@zigc/lib 0.17.0-dev.27 → 0.17.0-dev.296
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/c/fcntl.zig +6 -1
- package/c/inttypes.zig +0 -10
- package/c/math.zig +52 -112
- package/c/pthread.zig +57 -0
- package/c/search.zig +1 -27
- package/c/stdlib/drand48.zig +0 -57
- package/c/stdlib.zig +31 -121
- package/c/string.zig +20 -7
- package/c/strings.zig +0 -38
- package/c/unistd.zig +27 -26
- package/c/wchar.zig +10 -0
- package/c.zig +2 -2
- package/compiler/aro/aro/CodeGen.zig +7 -8
- package/compiler/aro/aro/Compilation.zig +17 -17
- package/compiler/aro/aro/Driver.zig +14 -13
- package/compiler/aro/aro/Parser.zig +20 -15
- package/compiler/aro/aro/Pragma.zig +3 -2
- package/compiler/aro/aro/Preprocessor.zig +9 -6
- package/compiler/aro/aro/pragmas/message.zig +3 -2
- package/compiler/aro/aro/text_literal.zig +3 -2
- package/compiler/aro/assembly_backend/x86_64.zig +4 -4
- package/compiler/build_runner.zig +2 -4
- package/compiler/reduce/Walk.zig +8 -13
- package/compiler/resinator/compile.zig +1 -1
- package/compiler/resinator/cvtres.zig +4 -4
- package/compiler/resinator/errors.zig +7 -7
- package/compiler/resinator/ico.zig +4 -4
- package/compiler/resinator/parse.zig +2 -2
- package/compiler/resinator/res.zig +1 -1
- package/compiler/std-docs.zig +1 -1
- package/compiler/test_runner.zig +1 -1
- package/compiler/translate-c/MacroTranslator.zig +1 -1
- package/compiler/translate-c/Translator.zig +6 -2
- package/compiler/translate-c/ast.zig +16 -23
- package/compiler/translate-c/main.zig +1 -1
- package/compiler_rt/addf3.zig +1 -1
- package/compiler_rt/arm.zig +3 -3
- package/compiler_rt/atomics.zig +1 -1
- package/compiler_rt/comparef.zig +3 -3
- package/compiler_rt/cos.zig +1 -3
- package/compiler_rt/count0bits.zig +1 -1
- package/compiler_rt/divdf3.zig +2 -2
- package/compiler_rt/divsf3.zig +1 -1
- package/compiler_rt/divtf3.zig +1 -1
- package/compiler_rt/divxf3.zig +1 -1
- package/compiler_rt/exp.zig +0 -2
- package/compiler_rt/exp2.zig +0 -2
- package/compiler_rt/extendf.zig +5 -5
- package/compiler_rt/fabs.zig +1 -3
- package/compiler_rt/float_from_int.zig +3 -4
- package/compiler_rt/fma.zig +0 -2
- package/compiler_rt/fmax.zig +1 -3
- package/compiler_rt/fmin.zig +1 -3
- package/compiler_rt/fmod.zig +2 -4
- package/compiler_rt/limb64.zig +3 -2
- package/compiler_rt/log.zig +0 -2
- package/compiler_rt/log10.zig +0 -2
- package/compiler_rt/log2.zig +0 -2
- package/compiler_rt/mulf3.zig +2 -2
- package/compiler_rt/rem_pio2l.zig +1 -1
- package/compiler_rt/round.zig +0 -2
- package/compiler_rt/sin.zig +1 -3
- package/compiler_rt/sincos.zig +1 -3
- package/compiler_rt/sqrt.zig +0 -2
- package/compiler_rt/ssp.zig +2 -2
- package/compiler_rt/tan.zig +1 -3
- package/compiler_rt/trunc.zig +0 -2
- package/compiler_rt/truncf.zig +3 -3
- package/compiler_rt/trunctfhf2.zig +3 -0
- package/compiler_rt.zig +5 -26
- package/docs/wasm/Walk.zig +1 -3
- package/docs/wasm/html_render.zig +1 -2
- package/docs/wasm/markdown/Parser.zig +16 -20
- package/fuzzer.zig +2 -0
- package/include/__clang_spirv_builtins.h +12 -12
- package/include/__float_float.h +176 -0
- package/include/__float_header_macro.h +12 -0
- package/include/__float_infinity_nan.h +20 -0
- package/include/amo.h +131 -0
- package/include/amxavx512intrin.h +215 -1
- package/include/amxintrin.h +0 -2
- package/include/arm_acle.h +37 -27
- package/include/arm_neon.h +218 -82
- package/include/arm_sme.h +8 -8
- package/include/arm_sve.h +4162 -3782
- package/include/avx10_2_512bf16intrin.h +19 -12
- package/include/avx10_2_512convertintrin.h +1 -1
- package/include/avx10_2_512niintrin.h +31 -31
- package/include/avx10_2_512satcvtdsintrin.h +1 -1
- package/include/avx10_2bf16intrin.h +54 -45
- package/include/avx10_2convertintrin.h +2 -2
- package/include/avx10_2copyintrin.h +1 -1
- package/include/avx10_2niintrin.h +14 -14
- package/include/avx10_2satcvtdsintrin.h +2 -2
- package/include/avx2intrin.h +275 -377
- package/include/avx512bf16intrin.h +25 -16
- package/include/avx512bitalgintrin.h +19 -30
- package/include/avx512bwintrin.h +386 -505
- package/include/avx512cdintrin.h +42 -55
- package/include/avx512dqintrin.h +132 -161
- package/include/avx512fintrin.h +1015 -1424
- package/include/avx512fp16intrin.h +112 -110
- package/include/avx512ifmaintrin.h +32 -34
- package/include/avx512ifmavlintrin.h +73 -46
- package/include/avx512vbmi2intrin.h +43 -32
- package/include/avx512vbmiintrin.h +19 -27
- package/include/avx512vbmivlintrin.h +35 -49
- package/include/avx512vlbf16intrin.h +32 -22
- package/include/avx512vlbitalgintrin.h +37 -53
- package/include/avx512vlbwintrin.h +470 -573
- package/include/avx512vlcdintrin.h +74 -102
- package/include/avx512vldqintrin.h +110 -127
- package/include/avx512vlfp16intrin.h +130 -111
- package/include/avx512vlintrin.h +945 -1299
- package/include/avx512vlvbmi2intrin.h +78 -63
- package/include/avx512vlvnniintrin.h +21 -18
- package/include/avx512vlvp2intersectintrin.h +2 -2
- package/include/avx512vnniintrin.h +10 -10
- package/include/avx512vp2intersectintrin.h +1 -2
- package/include/avx512vpopcntdqintrin.h +8 -10
- package/include/avx512vpopcntdqvlintrin.h +17 -15
- package/include/avxifmaintrin.h +16 -0
- package/include/avxintrin.h +165 -241
- package/include/avxvnniint16intrin.h +118 -99
- package/include/avxvnniint8intrin.h +56 -32
- package/include/avxvnniintrin.h +16 -8
- package/include/cpuid.h +101 -4
- package/include/emmintrin.h +168 -168
- package/include/f16cintrin.h +23 -9
- package/include/float.h +16 -155
- package/include/fma4intrin.h +98 -96
- package/include/fmaintrin.h +96 -66
- package/include/gfniintrin.h +21 -16
- package/include/hexagon_types.h +23 -20
- package/include/hvx_hexagon_protos.h +649 -860
- package/include/immintrin.h +0 -12
- package/include/intrin.h +4 -0
- package/include/lasxintrin.h +113 -0
- package/include/llvm_libc_wrappers/assert.h +3 -5
- package/include/llvm_libc_wrappers/ctype.h +3 -115
- package/include/llvm_libc_wrappers/inttypes.h +3 -5
- package/include/llvm_libc_wrappers/stdio.h +10 -38
- package/include/llvm_libc_wrappers/stdlib.h +3 -24
- package/include/llvm_libc_wrappers/string.h +2 -70
- package/include/llvm_libc_wrappers/time.h +4 -10
- package/include/mmintrin.h +188 -257
- package/include/module.modulemap +23 -4
- package/include/movrs_avx10_2_512intrin.h +2 -2
- package/include/movrs_avx10_2intrin.h +4 -4
- package/include/pmmintrin.h +12 -24
- package/include/ptrauth.h +16 -2
- package/include/riscv_mips.h +34 -0
- package/include/riscv_nds.h +89 -0
- package/include/sifive_vector.h +58 -2
- package/include/sm4evexintrin.h +2 -2
- package/include/smmintrin.h +77 -59
- package/include/spirvintrin.h +194 -0
- package/include/stddefer.h +19 -0
- package/include/tmmintrin.h +116 -147
- package/include/vaesintrin.h +1 -2
- package/include/xmmintrin.h +44 -70
- package/include/xopintrin.h +20 -10
- package/libc/include/aarch64-linux-any/asm/hwcap.h +1 -0
- package/libc/include/aarch64-linux-any/asm/unistd_64.h +1 -0
- package/libc/include/any-linux-any/asm-generic/errno.h +2 -0
- package/libc/include/any-linux-any/asm-generic/unistd.h +4 -1
- package/libc/include/any-linux-any/drm/amdgpu_drm.h +20 -6
- package/libc/include/any-linux-any/drm/amdxdna_accel.h +8 -0
- package/libc/include/any-linux-any/drm/drm_fourcc.h +6 -6
- package/libc/include/any-linux-any/drm/panfrost_drm.h +75 -1
- package/libc/include/any-linux-any/drm/panthor_drm.h +154 -3
- package/libc/include/any-linux-any/drm/rocket_accel.h +74 -24
- package/libc/include/any-linux-any/drm/xe_drm.h +89 -6
- package/libc/include/any-linux-any/linux/android/binder.h +1 -1
- package/libc/include/any-linux-any/linux/bpf.h +28 -0
- package/libc/include/any-linux-any/linux/btrfs.h +1 -0
- package/libc/include/any-linux-any/linux/btrfs_tree.h +32 -2
- package/libc/include/any-linux-any/linux/dma-buf.h +1 -0
- package/libc/include/any-linux-any/linux/dpll.h +1 -0
- package/libc/include/any-linux-any/linux/elf.h +2 -0
- package/libc/include/any-linux-any/linux/ethtool.h +21 -5
- package/libc/include/any-linux-any/linux/fs.h +1 -0
- package/libc/include/any-linux-any/linux/hyperv.h +1 -1
- package/libc/include/any-linux-any/linux/idxd.h +134 -134
- package/libc/include/any-linux-any/linux/if_alg.h +1 -1
- package/libc/include/any-linux-any/linux/if_link.h +1 -0
- package/libc/include/any-linux-any/linux/input-event-codes.h +4 -0
- package/libc/include/any-linux-any/linux/io_uring/bpf_filter.h +68 -0
- package/libc/include/any-linux-any/linux/io_uring/query.h +5 -1
- package/libc/include/any-linux-any/linux/io_uring.h +33 -2
- package/libc/include/any-linux-any/linux/iommufd.h +39 -0
- package/libc/include/any-linux-any/linux/kfd_ioctl.h +13 -3
- package/libc/include/any-linux-any/linux/kfd_sysfs.h +2 -1
- package/libc/include/any-linux-any/linux/kvm.h +30 -6
- package/libc/include/any-linux-any/linux/landlock.h +22 -8
- package/libc/include/any-linux-any/linux/magic.h +1 -0
- package/libc/include/any-linux-any/linux/mempolicy.h +3 -0
- package/libc/include/any-linux-any/linux/mount.h +11 -2
- package/libc/include/any-linux-any/linux/mptcp_pm.h +1 -1
- package/libc/include/any-linux-any/linux/mshv.h +2 -0
- package/libc/include/any-linux-any/linux/netfilter_bridge.h +5 -4
- package/libc/include/any-linux-any/linux/netfilter_ipv4.h +4 -5
- package/libc/include/any-linux-any/linux/netfilter_ipv6.h +3 -4
- package/libc/include/any-linux-any/linux/nfs.h +1 -1
- package/libc/include/any-linux-any/linux/nfsd_netlink.h +1 -0
- package/libc/include/any-linux-any/linux/nilfs2_api.h +2 -2
- package/libc/include/any-linux-any/linux/nilfs2_ondisk.h +97 -66
- package/libc/include/any-linux-any/linux/nl80211.h +104 -3
- package/libc/include/any-linux-any/linux/pci.h +7 -0
- package/libc/include/any-linux-any/linux/pci_regs.h +65 -6
- package/libc/include/any-linux-any/linux/pcitest.h +1 -0
- package/libc/include/any-linux-any/linux/perf_event.h +24 -3
- package/libc/include/any-linux-any/linux/pkt_sched.h +1 -0
- package/libc/include/any-linux-any/linux/prctl.h +30 -0
- package/libc/include/any-linux-any/linux/rseq.h +62 -5
- package/libc/include/any-linux-any/linux/shm.h +0 -1
- package/libc/include/any-linux-any/linux/stddef.h +4 -0
- package/libc/include/any-linux-any/linux/sysctl.h +1 -2
- package/libc/include/any-linux-any/linux/taskstats.h +12 -1
- package/libc/include/any-linux-any/linux/tcp.h +23 -3
- package/libc/include/any-linux-any/linux/typelimits.h +8 -0
- package/libc/include/any-linux-any/linux/ublk_cmd.h +120 -1
- package/libc/include/any-linux-any/linux/v4l2-controls.h +63 -0
- package/libc/include/any-linux-any/linux/vbox_vmmdev_types.h +2 -2
- package/libc/include/any-linux-any/linux/vduse.h +80 -5
- package/libc/include/any-linux-any/linux/version.h +3 -3
- package/libc/include/any-linux-any/linux/vfio.h +4 -0
- package/libc/include/any-linux-any/linux/videodev2.h +3 -0
- package/libc/include/any-linux-any/linux/virtio_ring.h +1 -2
- package/libc/include/any-linux-any/linux/vmclock-abi.h +20 -0
- package/libc/include/any-linux-any/rdma/bnxt_re-abi.h +16 -0
- package/libc/include/any-linux-any/rdma/ib_user_ioctl_cmds.h +16 -0
- package/libc/include/any-linux-any/rdma/mana-abi.h +3 -0
- package/libc/include/any-linux-any/scsi/scsi_bsg_ufs.h +8 -9
- package/libc/include/any-linux-any/sound/sof/tokens.h +6 -0
- package/libc/include/arc-linux-any/asm/swab.h +0 -63
- package/libc/include/arc-linux-any/asm/unistd_32.h +1 -0
- package/libc/include/arm-linux-any/asm/ptrace.h +0 -9
- package/libc/include/arm-linux-any/asm/unistd-eabi.h +1 -0
- package/libc/include/arm-linux-any/asm/unistd-oabi.h +1 -0
- package/libc/include/csky-linux-any/asm/unistd_32.h +1 -0
- package/libc/include/hexagon-linux-any/asm/unistd_32.h +1 -0
- package/libc/include/loongarch-linux-any/asm/hwcap.h +1 -0
- package/libc/include/loongarch-linux-any/asm/kvm.h +1 -0
- package/libc/include/loongarch-linux-any/asm/kvm_para.h +1 -0
- package/libc/include/loongarch-linux-any/asm/unistd_32.h +2 -0
- package/libc/include/loongarch-linux-any/asm/unistd_64.h +2 -0
- package/libc/include/m68k-linux-any/asm/unistd_32.h +1 -0
- package/libc/include/mips-linux-any/asm/errno.h +2 -0
- package/libc/include/mips-linux-any/asm/unistd_n32.h +1 -0
- package/libc/include/mips-linux-any/asm/unistd_n64.h +1 -0
- package/libc/include/mips-linux-any/asm/unistd_o32.h +1 -0
- package/libc/include/powerpc-linux-any/asm/unistd_32.h +1 -0
- package/libc/include/powerpc-linux-any/asm/unistd_64.h +1 -0
- package/libc/include/riscv-linux-any/asm/hwprobe.h +4 -0
- package/libc/include/riscv-linux-any/asm/kvm.h +3 -0
- package/libc/include/riscv-linux-any/asm/ptrace.h +37 -0
- package/libc/include/riscv-linux-any/asm/sigcontext.h +1 -0
- package/libc/include/riscv-linux-any/asm/unistd_32.h +1 -0
- package/libc/include/riscv-linux-any/asm/unistd_64.h +1 -0
- package/libc/include/s390x-linux-any/asm/unistd_64.h +1 -0
- package/libc/include/sparc-linux-any/asm/errno.h +2 -0
- package/libc/include/sparc-linux-any/asm/ioctls.h +4 -4
- package/libc/include/sparc-linux-any/asm/unistd_32.h +2 -0
- package/libc/include/sparc-linux-any/asm/unistd_64.h +2 -0
- package/libc/include/x86-linux-any/asm/auxvec.h +0 -4
- package/libc/include/x86-linux-any/asm/kvm.h +13 -8
- package/libc/include/x86-linux-any/asm/svm.h +16 -16
- package/libc/include/x86-linux-any/asm/unistd_32.h +1 -0
- package/libc/include/x86-linux-any/asm/unistd_64.h +1 -0
- package/libc/include/x86-linux-any/asm/unistd_x32.h +1 -0
- package/libc/include/xtensa-linux-any/asm/unistd_32.h +1 -0
- package/libc/musl/arch/mipsn32/syscall_arch.h +35 -32
- package/libcxx/include/__algorithm/all_of.h +11 -5
- package/libcxx/include/__algorithm/comp.h +4 -0
- package/libcxx/include/__algorithm/copy.h +28 -147
- package/libcxx/include/__algorithm/copy_backward.h +9 -24
- package/libcxx/include/__algorithm/copy_n.h +50 -16
- package/libcxx/include/__algorithm/count.h +2 -2
- package/libcxx/include/__algorithm/equal.h +43 -55
- package/libcxx/include/__algorithm/fill.h +26 -8
- package/libcxx/include/__algorithm/fill_n.h +32 -46
- package/libcxx/include/__algorithm/find.h +96 -39
- package/libcxx/include/__algorithm/find_end.h +105 -0
- package/libcxx/include/__algorithm/for_each.h +18 -24
- package/libcxx/include/__algorithm/for_each_n.h +20 -47
- package/libcxx/include/__algorithm/for_each_n_segment.h +1 -1
- package/libcxx/include/__algorithm/for_each_segment.h +26 -0
- package/libcxx/include/__algorithm/generate.h +4 -2
- package/libcxx/include/__algorithm/generate_n.h +19 -6
- package/libcxx/include/__algorithm/is_permutation.h +4 -4
- package/libcxx/include/__algorithm/iterator_operations.h +3 -0
- package/libcxx/include/__algorithm/lexicographical_compare.h +2 -2
- package/libcxx/include/__algorithm/lexicographical_compare_three_way.h +6 -6
- package/libcxx/include/__algorithm/make_heap.h +16 -4
- package/libcxx/include/__algorithm/mismatch.h +2 -2
- package/libcxx/include/__algorithm/move.h +8 -19
- package/libcxx/include/__algorithm/move_backward.h +9 -24
- package/libcxx/include/__algorithm/none_of.h +4 -4
- package/libcxx/include/__algorithm/partial_sort.h +1 -1
- package/libcxx/include/__algorithm/partial_sort_copy.h +1 -1
- package/libcxx/include/__algorithm/pstl.h +9 -9
- package/libcxx/include/__algorithm/radix_sort.h +27 -25
- package/libcxx/include/__algorithm/ranges_copy_n.h +3 -26
- package/libcxx/include/__algorithm/ranges_equal.h +17 -26
- package/libcxx/include/__algorithm/ranges_fill.h +7 -6
- package/libcxx/include/__algorithm/ranges_for_each.h +9 -1
- package/libcxx/include/__algorithm/ranges_generate_n.h +2 -6
- package/libcxx/include/__algorithm/ranges_search_n.h +2 -2
- package/libcxx/include/__algorithm/rotate.h +27 -44
- package/libcxx/include/__algorithm/search_n.h +49 -37
- package/libcxx/include/__algorithm/sift_down.h +19 -18
- package/libcxx/include/__algorithm/simd_utils.h +33 -4
- package/libcxx/include/__algorithm/specialized_algorithms.h +54 -0
- package/libcxx/include/__algorithm/stable_sort.h +1 -1
- package/libcxx/include/__assertion_handler +31 -4
- package/libcxx/include/__atomic/atomic.h +36 -60
- package/libcxx/include/__atomic/atomic_flag.h +19 -37
- package/libcxx/include/__atomic/atomic_ref.h +29 -12
- package/libcxx/include/__atomic/atomic_sync.h +127 -55
- package/libcxx/include/__atomic/atomic_sync_timed.h +144 -0
- package/libcxx/include/__atomic/atomic_waitable_traits.h +103 -0
- package/libcxx/include/__atomic/contention_t.h +27 -3
- package/libcxx/include/__atomic/floating_point_helper.h +55 -0
- package/libcxx/include/__bit/countl.h +1 -2
- package/libcxx/include/__bit/countr.h +1 -2
- package/libcxx/include/__bit/has_single_bit.h +1 -1
- package/libcxx/include/__bit/popcount.h +0 -1
- package/libcxx/include/__bit/rotate.h +15 -26
- package/libcxx/include/__bit_reference +207 -18
- package/libcxx/include/__charconv/from_chars_integral.h +1 -1
- package/libcxx/include/__charconv/from_chars_result.h +1 -1
- package/libcxx/include/__charconv/to_chars_integral.h +1 -0
- package/libcxx/include/__charconv/to_chars_result.h +1 -1
- package/libcxx/include/__charconv/traits.h +3 -24
- package/libcxx/include/__chrono/day.h +11 -0
- package/libcxx/include/__chrono/duration.h +58 -33
- package/libcxx/include/__chrono/file_clock.h +4 -2
- package/libcxx/include/__chrono/is_clock.h +72 -0
- package/libcxx/include/__chrono/leap_second.h +13 -0
- package/libcxx/include/__chrono/month.h +13 -0
- package/libcxx/include/__chrono/month_weekday.h +22 -0
- package/libcxx/include/__chrono/monthday.h +20 -0
- package/libcxx/include/__chrono/steady_clock.h +1 -1
- package/libcxx/include/__chrono/system_clock.h +3 -3
- package/libcxx/include/__chrono/time_point.h +37 -13
- package/libcxx/include/__chrono/weekday.h +25 -0
- package/libcxx/include/__chrono/year.h +11 -0
- package/libcxx/include/__chrono/year_month.h +13 -0
- package/libcxx/include/__chrono/year_month_day.h +23 -0
- package/libcxx/include/__chrono/year_month_weekday.h +26 -0
- package/libcxx/include/__chrono/zoned_time.h +16 -0
- package/libcxx/include/__compare/is_eq.h +6 -6
- package/libcxx/include/__compare/strong_order.h +12 -30
- package/libcxx/include/__compare/three_way_comparable.h +2 -2
- package/libcxx/include/__concepts/comparison_common_type.h +40 -0
- package/libcxx/include/__concepts/equality_comparable.h +2 -1
- package/libcxx/include/__condition_variable/condition_variable.h +1 -1
- package/libcxx/include/__config +63 -280
- package/libcxx/include/__configuration/abi.h +14 -24
- package/libcxx/include/__configuration/availability.h +65 -118
- package/libcxx/include/__configuration/compiler.h +6 -6
- package/libcxx/include/__configuration/experimental.h +38 -0
- package/libcxx/include/__configuration/hardening.h +215 -0
- package/libcxx/include/__configuration/language.h +3 -0
- package/libcxx/include/__configuration/platform.h +9 -16
- package/libcxx/include/__coroutine/coroutine_handle.h +9 -9
- package/libcxx/include/__coroutine/noop_coroutine_handle.h +11 -13
- package/libcxx/include/__debug_utils/strict_weak_ordering_check.h +1 -1
- package/libcxx/include/__exception/exception.h +6 -4
- package/libcxx/include/__exception/exception_ptr.h +27 -5
- package/libcxx/include/__exception/nested_exception.h +2 -2
- package/libcxx/include/__exception/operations.h +5 -5
- package/libcxx/include/__expected/bad_expected_access.h +8 -6
- package/libcxx/include/__expected/expected.h +62 -64
- package/libcxx/include/__expected/unexpected.h +4 -4
- package/libcxx/include/__filesystem/copy_options.h +4 -4
- package/libcxx/include/__filesystem/directory_entry.h +37 -33
- package/libcxx/include/__filesystem/directory_iterator.h +9 -11
- package/libcxx/include/__filesystem/directory_options.h +7 -4
- package/libcxx/include/__filesystem/file_status.h +3 -3
- package/libcxx/include/__filesystem/filesystem_error.h +9 -10
- package/libcxx/include/__filesystem/operations.h +97 -66
- package/libcxx/include/__filesystem/path.h +68 -64
- package/libcxx/include/__filesystem/path_iterator.h +1 -3
- package/libcxx/include/__filesystem/perm_options.h +4 -4
- package/libcxx/include/__filesystem/perms.h +4 -4
- package/libcxx/include/__filesystem/recursive_directory_iterator.h +9 -14
- package/libcxx/include/__filesystem/space_info.h +1 -1
- package/libcxx/include/__filesystem/u8path.h +12 -14
- package/libcxx/include/__flat_map/flat_map.h +88 -71
- package/libcxx/include/__flat_map/flat_multimap.h +251 -172
- package/libcxx/include/__flat_map/key_value_iterator.h +0 -1
- package/libcxx/include/__flat_map/utils.h +1 -0
- package/libcxx/include/__flat_set/flat_multiset.h +211 -143
- package/libcxx/include/__flat_set/flat_set.h +86 -68
- package/libcxx/include/__format/concepts.h +0 -14
- package/libcxx/include/__format/extended_grapheme_cluster_table.h +3 -2
- package/libcxx/include/__format/fmt_pair_like.h +42 -0
- package/libcxx/include/__format/format_arg.h +7 -10
- package/libcxx/include/__format/format_args.h +1 -1
- package/libcxx/include/__format/format_context.h +5 -5
- package/libcxx/include/__format/format_parse_context.h +2 -2
- package/libcxx/include/__format/formatter_output.h +30 -34
- package/libcxx/include/__format/indic_conjunct_break_table.h +3 -2
- package/libcxx/include/__format/range_default_formatter.h +2 -41
- package/libcxx/include/__format/range_format.h +71 -0
- package/libcxx/include/__format/range_formatter.h +1 -0
- package/libcxx/include/__format/width_estimation_table.h +4 -2
- package/libcxx/include/__functional/bind.h +10 -15
- package/libcxx/include/__functional/bind_back.h +1 -1
- package/libcxx/include/__functional/bind_front.h +1 -1
- package/libcxx/include/__functional/function.h +57 -75
- package/libcxx/include/__functional/hash.h +1 -10
- package/libcxx/include/__functional/identity.h +1 -1
- package/libcxx/include/__functional/is_transparent.h +8 -0
- package/libcxx/include/__functional/mem_fn.h +2 -1
- package/libcxx/include/__functional/operations.h +18 -0
- package/libcxx/include/__functional/ranges_operations.h +7 -0
- package/libcxx/include/__functional/reference_wrapper.h +7 -5
- package/libcxx/include/__functional/weak_result_type.h +14 -28
- package/libcxx/include/__fwd/ios.h +1 -1
- package/libcxx/include/__fwd/tuple.h +14 -0
- package/libcxx/include/__hash_table +371 -357
- package/libcxx/include/__ios/fpos.h +4 -4
- package/libcxx/include/__iterator/back_insert_iterator.h +1 -7
- package/libcxx/include/__iterator/bounded_iter.h +7 -8
- package/libcxx/include/__iterator/concepts.h +6 -9
- package/libcxx/include/__iterator/cpp17_iterator_concepts.h +13 -12
- package/libcxx/include/__iterator/distance.h +40 -18
- package/libcxx/include/__iterator/front_insert_iterator.h +1 -7
- package/libcxx/include/__iterator/insert_iterator.h +1 -7
- package/libcxx/include/__iterator/istream_iterator.h +6 -7
- package/libcxx/include/__iterator/istreambuf_iterator.h +6 -7
- package/libcxx/include/__iterator/iter_move.h +1 -1
- package/libcxx/include/__iterator/iterator.h +13 -0
- package/libcxx/include/__iterator/iterator_traits.h +13 -14
- package/libcxx/include/__iterator/ostream_iterator.h +1 -7
- package/libcxx/include/__iterator/ostreambuf_iterator.h +1 -7
- package/libcxx/include/__iterator/reverse_iterator.h +8 -13
- package/libcxx/include/__iterator/segmented_iterator.h +3 -8
- package/libcxx/include/__iterator/static_bounded_iter.h +3 -3
- package/libcxx/include/__iterator/wrap_iter.h +8 -6
- package/libcxx/include/__locale +3 -10
- package/libcxx/include/__locale_dir/locale_base_api/bsd_locale_fallbacks.h +0 -10
- package/libcxx/include/__locale_dir/locale_base_api.h +5 -28
- package/libcxx/include/__locale_dir/messages.h +1 -1
- package/libcxx/include/__locale_dir/money.h +2 -2
- package/libcxx/include/__locale_dir/num.h +190 -243
- package/libcxx/include/__locale_dir/pad_and_output.h +5 -6
- package/libcxx/include/__locale_dir/support/bsd_like.h +0 -20
- package/libcxx/include/__locale_dir/support/fuchsia.h +0 -7
- package/libcxx/include/__locale_dir/support/linux.h +0 -37
- package/libcxx/include/__locale_dir/support/netbsd.h +0 -2
- package/libcxx/include/__locale_dir/support/newlib.h +243 -0
- package/libcxx/include/__locale_dir/support/no_locale/characters.h +0 -4
- package/libcxx/include/__locale_dir/support/no_locale/strtonum.h +0 -9
- package/libcxx/include/__locale_dir/support/windows.h +0 -29
- package/libcxx/include/__locale_dir/time.h +3 -7
- package/libcxx/include/__math/hypot.h +1 -1
- package/libcxx/include/__math/logarithms.h +1 -1
- package/libcxx/include/__math/traits.h +80 -11
- package/libcxx/include/__mdspan/extents.h +7 -4
- package/libcxx/include/__mdspan/layout_stride.h +4 -5
- package/libcxx/include/__mdspan/mdspan.h +29 -23
- package/libcxx/include/__memory/addressof.h +7 -5
- package/libcxx/include/__memory/align.h +18 -1
- package/libcxx/include/__memory/allocate_at_least.h +15 -10
- package/libcxx/include/__memory/allocator.h +14 -26
- package/libcxx/include/__memory/allocator_traits.h +6 -4
- package/libcxx/include/__memory/compressed_pair.h +15 -9
- package/libcxx/include/__memory/construct_at.h +12 -23
- package/libcxx/include/__memory/inout_ptr.h +1 -1
- package/libcxx/include/__memory/is_sufficiently_aligned.h +1 -1
- package/libcxx/include/__memory/out_ptr.h +1 -1
- package/libcxx/include/__memory/pointer_traits.h +1 -1
- package/libcxx/include/__memory/raw_storage_iterator.h +3 -9
- package/libcxx/include/__memory/shared_count.h +9 -30
- package/libcxx/include/__memory/shared_ptr.h +100 -145
- package/libcxx/include/__memory/temp_value.h +1 -2
- package/libcxx/include/__memory/uninitialized_algorithms.h +44 -115
- package/libcxx/include/__memory/unique_ptr.h +14 -16
- package/libcxx/include/__memory/uses_allocator_construction.h +1 -0
- package/libcxx/include/__memory_resource/memory_resource.h +4 -2
- package/libcxx/include/__memory_resource/monotonic_buffer_resource.h +1 -1
- package/libcxx/include/__memory_resource/polymorphic_allocator.h +13 -8
- package/libcxx/include/__memory_resource/pool_options.h +1 -1
- package/libcxx/include/__memory_resource/synchronized_pool_resource.h +4 -2
- package/libcxx/include/__memory_resource/unsynchronized_pool_resource.h +1 -1
- package/libcxx/include/__mutex/mutex.h +2 -2
- package/libcxx/include/__mutex/once_flag.h +14 -11
- package/libcxx/include/__mutex/tag_types.h +3 -3
- package/libcxx/include/__mutex/unique_lock.h +8 -7
- package/libcxx/include/__new/align_val_t.h +6 -0
- package/libcxx/include/__new/allocate.h +1 -2
- package/libcxx/include/__new/exceptions.h +8 -2
- package/libcxx/include/__new/global_new_delete.h +4 -11
- package/libcxx/include/__new/interference_size.h +0 -4
- package/libcxx/include/__new/launder.h +3 -5
- package/libcxx/include/__new/nothrow_t.h +1 -1
- package/libcxx/include/__numeric/gcd_lcm.h +24 -34
- package/libcxx/include/__numeric/midpoint.h +9 -14
- package/libcxx/include/__numeric/pstl.h +2 -2
- package/libcxx/include/__numeric/saturation_arithmetic.h +13 -5
- package/libcxx/include/__ostream/basic_ostream.h +8 -8
- package/libcxx/include/__pstl/backends/default.h +14 -14
- package/libcxx/include/__pstl/backends/libdispatch.h +2 -2
- package/libcxx/include/__pstl/cpu_algos/find_if.h +1 -1
- package/libcxx/include/__pstl/cpu_algos/transform.h +5 -6
- package/libcxx/include/__pstl/cpu_algos/transform_reduce.h +5 -4
- package/libcxx/include/__random/binomial_distribution.h +10 -4
- package/libcxx/include/__random/mersenne_twister_engine.h +50 -154
- package/libcxx/include/__random/piecewise_constant_distribution.h +3 -2
- package/libcxx/include/__random/piecewise_linear_distribution.h +3 -2
- package/libcxx/include/__ranges/adjacent_transform_view.h +406 -0
- package/libcxx/include/__ranges/adjacent_view.h +419 -0
- package/libcxx/include/__ranges/as_rvalue_view.h +9 -9
- package/libcxx/include/__ranges/chunk_by_view.h +6 -6
- package/libcxx/include/__ranges/common_view.h +7 -7
- package/libcxx/include/__ranges/drop_view.h +8 -8
- package/libcxx/include/__ranges/drop_while_view.h +5 -5
- package/libcxx/include/__ranges/elements_of.h +49 -0
- package/libcxx/include/__ranges/empty_view.h +5 -5
- package/libcxx/include/__ranges/filter_view.h +10 -10
- package/libcxx/include/__ranges/iota_view.h +41 -22
- package/libcxx/include/__ranges/owning_view.h +15 -15
- package/libcxx/include/__ranges/ref_view.h +6 -6
- package/libcxx/include/__ranges/repeat_view.h +17 -10
- package/libcxx/include/__ranges/single_view.h +8 -8
- package/libcxx/include/__ranges/take_view.h +9 -9
- package/libcxx/include/__ranges/transform_view.h +1 -2
- package/libcxx/include/__ranges/view_interface.h +10 -10
- package/libcxx/include/__ranges/zip_transform_view.h +357 -0
- package/libcxx/include/__ranges/zip_view.h +20 -20
- package/libcxx/include/__split_buffer +612 -240
- package/libcxx/include/__stop_token/atomic_unique_lock.h +1 -1
- package/libcxx/include/__stop_token/stop_callback.h +2 -2
- package/libcxx/include/__stop_token/stop_source.h +1 -1
- package/libcxx/include/__stop_token/stop_state.h +4 -4
- package/libcxx/include/__stop_token/stop_token.h +1 -1
- package/libcxx/include/__string/char_traits.h +51 -31
- package/libcxx/include/__string/constexpr_c_functions.h +5 -5
- package/libcxx/include/__support/xlocale/__strtonum_fallback.h +0 -8
- package/libcxx/include/__system_error/error_category.h +8 -8
- package/libcxx/include/__system_error/error_code.h +5 -5
- package/libcxx/include/__system_error/error_condition.h +4 -4
- package/libcxx/include/__system_error/system_error.h +1 -1
- package/libcxx/include/__thread/id.h +1 -1
- package/libcxx/include/__thread/jthread.h +1 -1
- package/libcxx/include/__thread/poll_with_backoff.h +27 -8
- package/libcxx/include/__thread/support/c11.h +8 -8
- package/libcxx/include/__thread/support/pthread.h +8 -8
- package/libcxx/include/__thread/support/windows.h +8 -8
- package/libcxx/include/__thread/thread.h +13 -8
- package/libcxx/include/__thread/timed_backoff_policy.h +3 -2
- package/libcxx/include/__tree +849 -701
- package/libcxx/include/__tuple/sfinae_helpers.h +1 -44
- package/libcxx/include/__tuple/tuple_element.h +0 -12
- package/libcxx/include/__tuple/tuple_size.h +0 -4
- package/libcxx/include/__tuple/tuple_transform.h +45 -0
- package/libcxx/include/__type_traits/aligned_storage.h +13 -40
- package/libcxx/include/__type_traits/desugars_to.h +4 -0
- package/libcxx/include/__type_traits/invoke.h +8 -0
- package/libcxx/include/__type_traits/is_allocator.h +6 -7
- package/libcxx/include/__type_traits/is_array.h +26 -0
- package/libcxx/include/__type_traits/is_equality_comparable.h +16 -21
- package/libcxx/include/__type_traits/is_final.h +1 -1
- package/libcxx/include/__type_traits/is_floating_point.h +7 -6
- package/libcxx/include/__type_traits/is_generic_transparent_comparator.h +30 -0
- package/libcxx/include/__type_traits/is_specialization.h +2 -6
- package/libcxx/include/__type_traits/is_within_lifetime.h +29 -0
- package/libcxx/include/__type_traits/make_transparent.h +52 -0
- package/libcxx/include/__type_traits/reference_constructs_from_temporary.h +1 -7
- package/libcxx/include/__type_traits/reference_converts_from_temporary.h +1 -1
- package/libcxx/include/__utility/cmp.h +19 -7
- package/libcxx/include/__utility/default_three_way_comparator.h +70 -0
- package/libcxx/include/__utility/in_place.h +1 -1
- package/libcxx/include/__utility/integer_sequence.h +56 -41
- package/libcxx/include/__utility/lazy_synth_three_way_comparator.h +120 -0
- package/libcxx/include/__utility/pair.h +22 -25
- package/libcxx/include/__utility/scope_guard.h +2 -0
- package/libcxx/include/__utility/try_key_extraction.h +114 -0
- package/libcxx/include/__vector/vector.h +187 -160
- package/libcxx/include/__vector/vector_bool.h +76 -83
- package/libcxx/include/any +118 -155
- package/libcxx/include/array +88 -56
- package/libcxx/include/atomic +2 -0
- package/libcxx/include/barrier +20 -24
- package/libcxx/include/bitset +49 -30
- package/libcxx/include/ccomplex +3 -11
- package/libcxx/include/chrono +47 -0
- package/libcxx/include/ciso646 +3 -6
- package/libcxx/include/complex +77 -65
- package/libcxx/include/complex.h +10 -10
- package/libcxx/include/condition_variable +3 -3
- package/libcxx/include/cstdalign +3 -10
- package/libcxx/include/cstdbool +3 -10
- package/libcxx/include/ctgmath +2 -11
- package/libcxx/include/ctype.h +24 -24
- package/libcxx/include/cwchar +2 -2
- package/libcxx/include/deque +109 -225
- package/libcxx/include/errno.h +269 -269
- package/libcxx/include/exception +4 -1
- package/libcxx/include/ext/hash_map +7 -48
- package/libcxx/include/ext/hash_set +2 -8
- package/libcxx/include/fenv.h +43 -43
- package/libcxx/include/flat_map +663 -11
- package/libcxx/include/flat_set +543 -8
- package/libcxx/include/float.h +16 -16
- package/libcxx/include/forward_list +33 -53
- package/libcxx/include/fstream +57 -42
- package/libcxx/include/future +41 -51
- package/libcxx/include/initializer_list +9 -3
- package/libcxx/include/inttypes.h +16 -16
- package/libcxx/include/ios +28 -28
- package/libcxx/include/istream +19 -13
- package/libcxx/include/iterator +10 -0
- package/libcxx/include/latch +7 -5
- package/libcxx/include/limits +4 -4
- package/libcxx/include/list +61 -79
- package/libcxx/include/map +386 -274
- package/libcxx/include/math.h +19 -0
- package/libcxx/include/mdspan +1 -5
- package/libcxx/include/mutex +29 -19
- package/libcxx/include/optional +644 -149
- package/libcxx/include/print +9 -5
- package/libcxx/include/queue +39 -37
- package/libcxx/include/ranges +48 -0
- package/libcxx/include/regex +33 -31
- package/libcxx/include/scoped_allocator +16 -11
- package/libcxx/include/semaphore +15 -18
- package/libcxx/include/set +220 -192
- package/libcxx/include/shared_mutex +3 -8
- package/libcxx/include/span +53 -37
- package/libcxx/include/sstream +34 -28
- package/libcxx/include/stack +13 -15
- package/libcxx/include/stddef.h +10 -10
- package/libcxx/include/stdexcept +2 -2
- package/libcxx/include/stdio.h +20 -21
- package/libcxx/include/streambuf +59 -19
- package/libcxx/include/string +574 -573
- package/libcxx/include/string_view +113 -89
- package/libcxx/include/strstream +10 -10
- package/libcxx/include/syncstream +4 -4
- package/libcxx/include/tgmath.h +12 -12
- package/libcxx/include/tuple +242 -212
- package/libcxx/include/type_traits +8 -2
- package/libcxx/include/typeindex +5 -3
- package/libcxx/include/typeinfo +92 -89
- package/libcxx/include/unordered_map +179 -304
- package/libcxx/include/unordered_set +168 -195
- package/libcxx/include/utility +12 -0
- package/libcxx/include/valarray +106 -161
- package/libcxx/include/variant +25 -33
- package/libcxx/include/version +41 -24
- package/libcxx/include/wctype.h +29 -29
- package/libcxx/libc/hdr/errno_macros.h +3 -1
- package/libcxx/libc/hdr/stdint_proxy.h +18 -0
- package/libcxx/libc/hdr/types/wchar_t.h +23 -0
- package/libcxx/libc/hdr/wchar_overlay.h +69 -0
- package/libcxx/libc/include/llvm-libc-macros/cfloat128-macros.h +41 -0
- package/libcxx/libc/include/llvm-libc-macros/cfloat16-macros.h +20 -0
- package/libcxx/libc/include/llvm-libc-macros/wchar-macros.h +18 -0
- package/libcxx/libc/include/llvm-libc-types/cfloat128.h +11 -28
- package/libcxx/libc/include/llvm-libc-types/cfloat16.h +4 -7
- package/libcxx/libc/include/llvm-libc-types/wint_t.h +14 -0
- package/libcxx/libc/shared/libc_common.h +5 -0
- package/libcxx/libc/src/__support/CPP/bit.h +50 -12
- package/libcxx/libc/src/__support/CPP/type_traits/is_complex.h +7 -0
- package/libcxx/libc/src/__support/CPP/type_traits/is_destructible.h +2 -1
- package/libcxx/libc/src/__support/CPP/type_traits/is_unsigned.h +6 -0
- package/libcxx/libc/src/__support/CPP/utility/integer_sequence.h +10 -0
- package/libcxx/libc/src/__support/FPUtil/FPBits.h +6 -7
- package/libcxx/libc/src/__support/FPUtil/rounding_mode.h +50 -4
- package/libcxx/libc/src/__support/big_int.h +14 -14
- package/libcxx/libc/src/__support/common.h +3 -2
- package/libcxx/libc/src/__support/ctype_utils.h +19 -12
- package/libcxx/libc/src/__support/detailed_powers_of_ten.h +1 -2
- package/libcxx/libc/src/__support/high_precision_decimal.h +33 -14
- package/libcxx/libc/src/__support/libc_assert.h +2 -0
- package/libcxx/libc/src/__support/macros/attributes.h +44 -1
- package/libcxx/libc/src/__support/macros/config.h +23 -0
- package/libcxx/libc/src/__support/macros/null_check.h +1 -2
- package/libcxx/libc/src/__support/macros/optimization.h +3 -0
- package/libcxx/libc/src/__support/macros/properties/architectures.h +5 -1
- package/libcxx/libc/src/__support/macros/properties/compiler.h +8 -3
- package/libcxx/libc/src/__support/macros/properties/cpu_features.h +17 -1
- package/libcxx/libc/src/__support/macros/properties/types.h +2 -3
- package/libcxx/libc/src/__support/macros/sanitizer.h +0 -10
- package/libcxx/libc/src/__support/math_extras.h +20 -4
- package/libcxx/libc/src/__support/str_to_float.h +98 -125
- package/libcxx/libc/src/__support/str_to_integer.h +39 -27
- package/libcxx/libc/src/__support/wctype_utils.h +588 -0
- package/libcxx/src/any.cpp +4 -0
- package/libcxx/src/atomic.cpp +365 -80
- package/libcxx/src/barrier.cpp +4 -3
- package/libcxx/src/charconv.cpp +6 -3
- package/libcxx/src/condition_variable_destructor.cpp +1 -1
- package/libcxx/src/error_category.cpp +3 -1
- package/libcxx/src/exception.cpp +2 -10
- package/libcxx/src/experimental/time_zone.cpp +1 -1
- package/libcxx/src/experimental/tzdb.cpp +4 -1
- package/libcxx/src/filesystem/error.h +4 -22
- package/libcxx/src/filesystem/format_string.h +9 -18
- package/libcxx/src/filesystem/int128_builtins.cpp +2 -0
- package/libcxx/src/filesystem/operations.cpp +2 -9
- package/libcxx/src/filesystem/path.cpp +3 -1
- package/libcxx/src/include/aligned_alloc.h +65 -0
- package/libcxx/src/include/config_elast.h +1 -1
- package/libcxx/src/include/from_chars_floating_point.h +11 -7
- package/libcxx/src/include/overridable_function.h +8 -10
- package/libcxx/src/iostream.cpp +56 -37
- package/libcxx/src/locale.cpp +213 -196
- package/libcxx/src/memory.cpp +12 -14
- package/libcxx/src/mutex_destructor.cpp +1 -1
- package/libcxx/src/new.cpp +5 -5
- package/libcxx/src/optional.cpp +4 -0
- package/libcxx/src/print.cpp +9 -1
- package/libcxx/src/random.cpp +0 -26
- package/libcxx/src/string.cpp +10 -25
- package/libcxx/src/support/runtime/exception_fallback.ipp +2 -0
- package/libcxx/src/support/runtime/exception_glibcxx.ipp +3 -0
- package/libcxx/src/support/runtime/exception_libcxxabi.ipp +6 -2
- package/libcxx/src/support/runtime/exception_libcxxrt.ipp +2 -0
- package/libcxx/src/support/runtime/exception_msvc.ipp +2 -0
- package/libcxx/src/support/runtime/exception_pointer_cxxabi.ipp +9 -10
- package/libcxx/src/support/runtime/exception_pointer_glibcxx.ipp +2 -0
- package/libcxx/src/support/runtime/exception_pointer_msvc.ipp +1 -0
- package/libcxx/src/support/runtime/exception_pointer_unimplemented.ipp +1 -0
- package/libcxx/src/support/win32/locale_win32.cpp +1 -1
- package/libcxx/src/system_error.cpp +2 -0
- package/libcxx/src/thread.cpp +1 -3
- package/libcxx/src/valarray.cpp +1 -2
- package/libcxx/src/vector.cpp +2 -2
- package/libcxxabi/include/__cxxabi_config.h +42 -23
- package/libcxxabi/src/cxa_exception.cpp +4 -2
- package/libcxxabi/src/cxa_exception.h +16 -14
- package/libcxxabi/src/cxa_personality.cpp +126 -9
- package/libcxxabi/src/cxa_thread_atexit.cpp +2 -1
- package/libcxxabi/src/demangle/DemangleConfig.h +4 -0
- package/libcxxabi/src/demangle/ItaniumDemangle.h +9 -6
- package/libcxxabi/src/demangle/Utility.h +21 -7
- package/libcxxabi/src/fallback_malloc.cpp +1 -1
- package/libcxxabi/src/private_typeinfo.cpp +6 -0
- package/libcxxabi/src/stdlib_new_delete.cpp +5 -5
- package/libtsan/LICENSE.TXT +311 -0
- package/libtsan/builtins/assembly.h +41 -8
- package/libtsan/interception/interception_win.cpp +4 -0
- package/libtsan/sanitizer_common/sanitizer_allocator_primary32.h +1 -0
- package/libtsan/sanitizer_common/sanitizer_allocator_primary64.h +18 -0
- package/libtsan/sanitizer_common/sanitizer_common.h +15 -2
- package/libtsan/sanitizer_common/sanitizer_common_interceptors.inc +30 -4
- package/libtsan/sanitizer_common/sanitizer_common_interceptors_ioctl.inc +4 -0
- package/libtsan/sanitizer_common/sanitizer_common_interceptors_vfork_aarch64.inc.S +2 -1
- package/libtsan/sanitizer_common/sanitizer_common_interceptors_vfork_i386.inc.S +2 -0
- package/libtsan/sanitizer_common/sanitizer_common_interceptors_vfork_x86_64.inc.S +2 -0
- package/libtsan/sanitizer_common/sanitizer_common_syscalls.inc +18 -0
- package/libtsan/sanitizer_common/sanitizer_file.cpp +40 -12
- package/libtsan/sanitizer_common/sanitizer_file.h +3 -0
- package/libtsan/sanitizer_common/sanitizer_flags.inc +7 -0
- package/libtsan/sanitizer_common/sanitizer_fuchsia.cpp +30 -3
- package/libtsan/sanitizer_common/sanitizer_haiku.cpp +2 -2
- package/libtsan/sanitizer_common/sanitizer_internal_defs.h +1 -1
- package/libtsan/sanitizer_common/sanitizer_libc.cpp +8 -0
- package/libtsan/sanitizer_common/sanitizer_libc.h +1 -0
- package/libtsan/sanitizer_common/sanitizer_linux.cpp +15 -7
- package/libtsan/sanitizer_common/sanitizer_linux.h +3 -3
- package/libtsan/sanitizer_common/sanitizer_linux_libcdep.cpp +1 -0
- package/libtsan/sanitizer_common/sanitizer_mac.cpp +255 -104
- package/libtsan/sanitizer_common/sanitizer_mac.h +5 -0
- package/libtsan/sanitizer_common/sanitizer_netbsd.cpp +2 -2
- package/libtsan/sanitizer_common/sanitizer_platform.h +27 -1
- package/libtsan/sanitizer_common/sanitizer_platform_interceptors.h +5 -4
- package/libtsan/sanitizer_common/sanitizer_platform_limits_posix.cpp +15 -17
- package/libtsan/sanitizer_common/sanitizer_platform_limits_posix.h +32 -6
- package/libtsan/sanitizer_common/sanitizer_posix.cpp +3 -12
- package/libtsan/sanitizer_common/sanitizer_posix.h +2 -1
- package/libtsan/sanitizer_common/sanitizer_posix_libcdep.cpp +19 -0
- package/libtsan/sanitizer_common/sanitizer_procmaps_mac.cpp +102 -37
- package/libtsan/sanitizer_common/sanitizer_redefine_builtins.h +1 -1
- package/libtsan/sanitizer_common/sanitizer_signal_interceptors.inc +40 -2
- package/libtsan/sanitizer_common/sanitizer_stoptheworld.h +1 -1
- package/libtsan/sanitizer_common/sanitizer_stoptheworld_linux_libcdep.cpp +83 -12
- package/libtsan/sanitizer_common/sanitizer_stoptheworld_mac.cpp +3 -3
- package/libtsan/sanitizer_common/sanitizer_stoptheworld_netbsd_libcdep.cpp +7 -7
- package/libtsan/sanitizer_common/sanitizer_stoptheworld_win.cpp +2 -2
- package/libtsan/sanitizer_common/sanitizer_symbolizer_internal.h +5 -1
- package/libtsan/sanitizer_common/sanitizer_symbolizer_libcdep.cpp +11 -1
- package/libtsan/sanitizer_common/sanitizer_symbolizer_mac.cpp +86 -29
- package/libtsan/sanitizer_common/sanitizer_symbolizer_posix_libcdep.cpp +27 -16
- package/libtsan/sanitizer_common/sanitizer_thread_registry.cpp +5 -4
- package/libtsan/sanitizer_common/sanitizer_thread_registry.h +4 -4
- package/libtsan/sanitizer_common/sanitizer_win.cpp +1 -3
- package/libtsan/tsan_debugging.cpp +2 -2
- package/libtsan/tsan_flags.cpp +37 -0
- package/libtsan/tsan_flags.h +8 -0
- package/libtsan/tsan_flags.inc +12 -0
- package/libtsan/tsan_interceptors.h +9 -1
- package/libtsan/tsan_interceptors_mac.cpp +19 -0
- package/libtsan/tsan_interceptors_posix.cpp +78 -32
- package/libtsan/tsan_interface.h +3 -3
- package/libtsan/tsan_interface_ann.cpp +23 -9
- package/libtsan/tsan_mman.cpp +18 -4
- package/libtsan/tsan_platform.h +44 -7
- package/libtsan/tsan_platform_linux.cpp +42 -14
- package/libtsan/tsan_platform_mac.cpp +16 -3
- package/libtsan/tsan_report.h +14 -1
- package/libtsan/tsan_rtl.cpp +14 -0
- package/libtsan/tsan_rtl.h +7 -2
- package/libtsan/tsan_rtl_aarch64.S +3 -5
- package/libtsan/tsan_rtl_access.cpp +8 -3
- package/libtsan/tsan_rtl_amd64.S +2 -0
- package/libtsan/tsan_rtl_mutex.cpp +94 -49
- package/libtsan/tsan_rtl_report.cpp +132 -65
- package/libtsan/tsan_rtl_thread.cpp +31 -9
- package/libtsan/tsan_symbolize.cpp +1 -1
- package/libtsan/tsan_symbolize.h +1 -1
- package/libtsan/tsan_trace.h +1 -1
- package/libunwind/include/__libunwind_config.h +11 -2
- package/libunwind/include/libunwind.h +117 -11
- package/libunwind/include/unwind_arm_ehabi.h +4 -1
- package/libunwind/src/AddressSpace.hpp +42 -20
- package/libunwind/src/CompactUnwinder.hpp +16 -5
- package/libunwind/src/DwarfInstructions.hpp +24 -13
- package/libunwind/src/DwarfParser.hpp +62 -23
- package/libunwind/src/EHHeaderParser.hpp +7 -4
- package/libunwind/src/Registers.hpp +226 -22
- package/libunwind/src/Unwind-seh.cpp +6 -7
- package/libunwind/src/Unwind-wasm.c +7 -7
- package/libunwind/src/UnwindCursor.hpp +167 -49
- package/libunwind/src/UnwindLevel1.c +46 -17
- package/libunwind/src/UnwindRegistersRestore.S +46 -5
- package/libunwind/src/UnwindRegistersSave.S +86 -2
- package/libunwind/src/assembly.h +5 -1
- package/libunwind/src/config.h +9 -0
- package/libunwind/src/gcc_personality_v0.c +79 -6
- package/libunwind/src/libunwind.cpp +104 -4
- package/libunwind/src/libunwind_ext.h +7 -1
- package/libunwind/src/shadow_stack_unwind.h +2 -2
- package/lldb/pretty_printers.py +948 -0
- package/package.json +1 -1
- package/std/Build/Cache/Path.zig +1 -1
- package/std/Build/Cache.zig +6 -6
- package/std/Build/Step/Compile.zig +18 -20
- package/std/Build/Step/ConfigHeader.zig +1 -1
- package/std/Build/Step/Run.zig +15 -8
- package/std/Build/Step.zig +3 -7
- package/std/Build/Watch/FsEvents.zig +1 -1
- package/std/Build/WebServer.zig +2 -2
- package/std/Build/abi.zig +1 -1
- package/std/Build.zig +0 -3
- package/std/Io/Dir.zig +9 -4
- package/std/Io/Dispatch.zig +8 -18
- package/std/Io/File/Writer.zig +8 -6
- package/std/Io/File.zig +0 -9
- package/std/Io/Kqueue.zig +3 -3
- package/std/Io/Reader.zig +8 -14
- package/std/Io/Semaphore.zig +112 -17
- package/std/Io/Terminal.zig +1 -1
- package/std/Io/Threaded.zig +183 -46
- package/std/Io/Uring.zig +18 -20
- package/std/Io/Writer.zig +10 -7
- package/std/Io/net/HostName.zig +11 -6
- package/std/Io/net.zig +11 -11
- package/std/Io.zig +104 -38
- package/std/Random/ChaCha.zig +2 -2
- package/std/Random/benchmark.zig +2 -2
- package/std/Random.zig +6 -6
- package/std/SemanticVersion.zig +1 -1
- package/std/Target/Query.zig +2 -2
- package/std/Target/aarch64.zig +620 -77
- package/std/Target/amdgcn.zig +421 -21
- package/std/Target/arm.zig +40 -6
- package/std/Target/bpf.zig +6 -0
- package/std/Target/hexagon.zig +41 -6
- package/std/Target/loongarch.zig +18 -0
- package/std/Target/mips.zig +6 -0
- package/std/Target/nvptx.zig +58 -35
- package/std/Target/powerpc.zig +27 -19
- package/std/Target/riscv.zig +415 -177
- package/std/Target/sparc.zig +17 -0
- package/std/Target/wasm.zig +7 -0
- package/std/Target/x86.zig +200 -31
- package/std/Target/xtensa.zig +65 -0
- package/std/Target.zig +67 -14
- package/std/Thread.zig +5 -2
- package/std/array_hash_map.zig +103 -57
- package/std/array_list.zig +10 -31
- package/std/base64.zig +5 -5
- package/std/bit_set.zig +86 -69
- package/std/c/haiku.zig +3 -0
- package/std/c/serenity.zig +1 -6
- package/std/c.zig +123 -32
- package/std/compress/flate/Decompress.zig +3 -4
- package/std/compress/zstd/Decompress.zig +7 -9
- package/std/crypto/25519/curve25519.zig +2 -2
- package/std/crypto/25519/edwards25519.zig +3 -3
- package/std/crypto/25519/ristretto255.zig +2 -2
- package/std/crypto/25519/scalar.zig +6 -6
- package/std/crypto/25519/x25519.zig +1 -1
- package/std/crypto/Certificate.zig +14 -2
- package/std/crypto/Sha1.zig +1 -1
- package/std/crypto/aegis.zig +45 -39
- package/std/crypto/aes_ccm.zig +22 -22
- package/std/crypto/aes_gcm.zig +10 -12
- package/std/crypto/aes_ocb.zig +9 -9
- package/std/crypto/argon2.zig +15 -15
- package/std/crypto/ascon.zig +75 -33
- package/std/crypto/bcrypt.zig +20 -10
- package/std/crypto/benchmark.zig +7 -7
- package/std/crypto/blake2.zig +100 -64
- package/std/crypto/cbc_mac.zig +1 -1
- package/std/crypto/chacha20.zig +10 -10
- package/std/crypto/cmac.zig +3 -3
- package/std/crypto/codecs/asn1/Oid.zig +12 -1
- package/std/crypto/codecs/asn1/der/Encoder.zig +1 -1
- package/std/crypto/codecs/asn1.zig +1 -1
- package/std/crypto/codecs/base64_hex_ct.zig +3 -5
- package/std/crypto/ecdsa.zig +9 -9
- package/std/crypto/ff.zig +4 -5
- package/std/crypto/ghash_polyval.zig +4 -4
- package/std/crypto/hkdf.zig +1 -1
- package/std/crypto/isap.zig +3 -3
- package/std/crypto/kangarootwelve.zig +1 -1
- package/std/crypto/keccak_p.zig +8 -8
- package/std/crypto/md5.zig +1 -1
- package/std/crypto/ml_dsa.zig +39 -39
- package/std/crypto/ml_kem.zig +12 -34
- package/std/crypto/modes.zig +2 -2
- package/std/crypto/pbkdf2.zig +1 -1
- package/std/crypto/pcurves/common.zig +4 -5
- package/std/crypto/pcurves/p256/scalar.zig +3 -3
- package/std/crypto/pcurves/p384/scalar.zig +2 -2
- package/std/crypto/pcurves/secp256k1/scalar.zig +3 -3
- package/std/crypto/pcurves/tests/p256.zig +5 -5
- package/std/crypto/pcurves/tests/p384.zig +5 -5
- package/std/crypto/pcurves/tests/secp256k1.zig +3 -3
- package/std/crypto/salsa20.zig +8 -8
- package/std/crypto/sha2.zig +2 -2
- package/std/crypto/sha3.zig +2 -2
- package/std/crypto/siphash.zig +1 -1
- package/std/crypto/timing_safe.zig +10 -9
- package/std/crypto/tls/Client.zig +89 -13
- package/std/crypto/tls.zig +1 -1
- package/std/crypto.zig +4 -3
- package/std/debug/Dwarf.zig +15 -12
- package/std/debug/Pdb.zig +25 -17
- package/std/debug.zig +14 -8
- package/std/deque.zig +1 -1
- package/std/elf.zig +1 -1
- package/std/enums.zig +8 -8
- package/std/fmt/parse_float/decimal.zig +1 -1
- package/std/fmt/parse_float.zig +1 -1
- package/std/fmt.zig +14 -10
- package/std/fs/path.zig +7 -5
- package/std/hash/Adler32.zig +3 -3
- package/std/hash/benchmark.zig +2 -2
- package/std/hash/wyhash.zig +2 -2
- package/std/hash/xxhash.zig +0 -6
- package/std/hash_map.zig +20 -1
- package/std/heap/BufferFirstAllocator.zig +165 -0
- package/std/heap/debug_allocator.zig +2 -2
- package/std/heap/memory_pool.zig +34 -197
- package/std/heap.zig +2 -133
- package/std/http/Client.zig +21 -24
- package/std/http.zig +4 -5
- package/std/json/Scanner.zig +2 -2
- package/std/json/static.zig +2 -2
- package/std/{builtin.zig → lang.zig} +5 -1
- package/std/leb128.zig +2 -2
- package/std/math/big/int.zig +5 -5
- package/std/math/big.zig +4 -4
- package/std/math/copysign.zig +1 -1
- package/std/math/float.zig +62 -0
- package/std/math/frexp.zig +4 -4
- package/std/math/ilogb.zig +1 -1
- package/std/math/isfinite.zig +1 -1
- package/std/math/isinf.zig +1 -1
- package/std/math/isnan.zig +1 -2
- package/std/math/isnormal.zig +2 -2
- package/std/math/iszero.zig +2 -2
- package/std/math/ldexp.zig +2 -2
- package/std/math/log10.zig +0 -2
- package/std/math/modf.zig +1 -2
- package/std/math/nextafter.zig +3 -4
- package/std/math/pow.zig +1 -1
- package/std/math/powi.zig +2 -3
- package/std/math/signbit.zig +0 -1
- package/std/math.zig +21 -26
- package/std/mem/Allocator.zig +0 -6
- package/std/mem.zig +31 -31
- package/std/meta/trailer_flags.zig +1 -1
- package/std/meta.zig +4 -36
- package/std/multi_array_list.zig +24 -0
- package/std/os/emscripten.zig +7 -6
- package/std/os/linux/IoUring.zig +2 -0
- package/std/os/linux/aarch64.zig +41 -12
- package/std/os/linux/arc.zig +41 -12
- package/std/os/linux/arm.zig +41 -12
- package/std/os/linux/bpf.zig +6 -0
- package/std/os/linux/hexagon.zig +33 -11
- package/std/os/linux/ioctl.zig +2 -2
- package/std/os/linux/loongarch32.zig +41 -13
- package/std/os/linux/loongarch64.zig +41 -12
- package/std/os/linux/m68k.zig +41 -13
- package/std/os/linux/mips.zig +67 -36
- package/std/os/linux/mips64.zig +60 -29
- package/std/os/linux/mipsn32.zig +60 -29
- package/std/os/linux/or1k.zig +41 -12
- package/std/os/linux/powerpc.zig +41 -12
- package/std/os/linux/powerpc64.zig +41 -12
- package/std/os/linux/riscv32.zig +41 -12
- package/std/os/linux/riscv64.zig +41 -12
- package/std/os/linux/s390x.zig +44 -7
- package/std/os/linux/sparc64.zig +83 -52
- package/std/os/linux/syscalls.zig +26 -1
- package/std/os/linux/thumb.zig +52 -36
- package/std/os/linux/x32.zig +41 -12
- package/std/os/linux/x86.zig +44 -15
- package/std/os/linux/x86_64.zig +41 -12
- package/std/os/linux.zig +416 -450
- package/std/os/plan9.zig +1 -1
- package/std/os/uefi/hii.zig +1 -1
- package/std/os/uefi/tables/boot_services.zig +9 -8
- package/std/os/uefi/tables/runtime_services.zig +1 -1
- package/std/os/windows.zig +130 -1
- package/std/os.zig +41 -0
- package/std/posix.zig +1 -1
- package/std/process/Environ.zig +1 -1
- package/std/process.zig +5 -4
- package/std/simd.zig +4 -21
- package/std/sort.zig +3 -3
- package/std/start.zig +17 -4
- package/std/std.zig +8 -1
- package/std/tar/Writer.zig +39 -33
- package/std/tar.zig +8 -4
- package/std/testing.zig +2 -2
- package/std/unicode.zig +13 -8
- package/std/zig/Ast/Render.zig +6 -30
- package/std/zig/Ast.zig +12 -37
- package/std/zig/AstGen.zig +175 -332
- package/std/zig/AstRlAnnotate.zig +1 -22
- package/std/zig/AstSmith.zig +5 -11
- package/std/zig/BuiltinFn.zig +0 -32
- package/std/zig/ErrorBundle.zig +6 -2
- package/std/zig/LibCDirs.zig +15 -6
- package/std/zig/LibCInstallation.zig +34 -12
- package/std/zig/Parse.zig +11 -76
- package/std/zig/TokenSmith.zig +0 -6
- package/std/zig/WindowsSdk.zig +18 -18
- package/std/zig/Zir.zig +65 -132
- package/std/zig/ZonGen.zig +19 -27
- package/std/zig/llvm/Builder.zig +39 -33
- package/std/zig/llvm/bitcode_writer.zig +3 -3
- package/std/zig/llvm/ir.zig +8 -8
- package/std/zig/system/arm.zig +56 -2
- package/std/zig/system/windows.zig +34 -1
- package/std/zig/system/x86.zig +60 -16
- package/std/zig/system.zig +0 -10
- package/std/zig/target.zig +1 -4
- package/std/zig/tokenizer.zig +4 -43
- package/std/zig.zig +3 -14
- package/std/zip.zig +5 -5
- package/std/zon/parse.zig +1 -1
- package/zig.h +340 -1
- package/compiler_rt/long_double.zig +0 -37
- package/include/amxbf16transposeintrin.h +0 -94
- package/include/amxcomplextransposeintrin.h +0 -303
- package/include/amxfp16transposeintrin.h +0 -94
- package/include/amxmovrstransposeintrin.h +0 -200
- package/include/amxtf32transposeintrin.h +0 -105
- package/include/amxtransposeintrin.h +0 -248
- package/libc/include/hexagon-linux-any/asm/signal.h +0 -29
- package/libc/include/s390x-linux-any/asm/tape390.h +0 -103
- package/libc/mingw/math/lrintl.c +0 -18
- package/libc/mingw/math/rintl.c +0 -16
- package/libc/mingw/winpthreads/spinlock.c +0 -82
- package/libc/musl/src/linux/tee.c +0 -8
- package/libc/musl/src/math/i386/lrintl.c +0 -8
- package/libc/musl/src/math/i386/rintl.c +0 -7
- package/libc/musl/src/math/lrintl.c +0 -36
- package/libc/musl/src/math/rintl.c +0 -29
- package/libc/musl/src/math/s390x/rintl.c +0 -15
- package/libc/musl/src/math/x32/lrintl.s +0 -7
- package/libc/musl/src/math/x32/rintl.s +0 -6
- package/libc/musl/src/math/x86_64/lrintl.c +0 -8
- package/libc/musl/src/math/x86_64/rintl.c +0 -7
- package/libc/musl/src/string/strdup.c +0 -10
- package/libc/musl/src/string/strndup.c +0 -12
- package/libc/musl/src/string/wcsdup.c +0 -10
- package/libc/musl/src/thread/pthread_spin_destroy.c +0 -6
- package/libc/musl/src/thread/pthread_spin_init.c +0 -6
- package/libc/musl/src/thread/pthread_spin_lock.c +0 -8
- package/libc/musl/src/thread/pthread_spin_trylock.c +0 -7
- package/libc/musl/src/thread/pthread_spin_unlock.c +0 -7
- package/libc/musl/src/unistd/dup2.c +0 -20
- package/libc/musl/src/unistd/dup3.c +0 -26
- package/libc/wasi/thread-stub/pthread_spin_lock.c +0 -8
- package/libc/wasi/thread-stub/pthread_spin_trylock.c +0 -8
- package/libc/wasi/thread-stub/pthread_spin_unlock.c +0 -7
- package/libtsan/sanitizer_common/sanitizer_coverage_interface.inc +0 -43
- package/std/Build/Step/CheckObject.zig +0 -2764
- /package/std/{builtin → lang}/assembly.zig +0 -0
package/include/avx512vlintrin.h
CHANGED
|
@@ -15,14 +15,20 @@
|
|
|
15
15
|
#define __AVX512VLINTRIN_H
|
|
16
16
|
|
|
17
17
|
#define __DEFAULT_FN_ATTRS128 \
|
|
18
|
-
__attribute__((__always_inline__, __nodebug__,
|
|
19
|
-
__target__("avx512vl,no-evex512"), \
|
|
18
|
+
__attribute__((__always_inline__, __nodebug__, __target__("avx512vl"), \
|
|
20
19
|
__min_vector_width__(128)))
|
|
21
20
|
#define __DEFAULT_FN_ATTRS256 \
|
|
22
|
-
__attribute__((__always_inline__, __nodebug__,
|
|
23
|
-
__target__("avx512vl,no-evex512"), \
|
|
21
|
+
__attribute__((__always_inline__, __nodebug__, __target__("avx512vl"), \
|
|
24
22
|
__min_vector_width__(256)))
|
|
25
23
|
|
|
24
|
+
#if defined(__cplusplus) && (__cplusplus >= 201103L)
|
|
25
|
+
#define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256 constexpr
|
|
26
|
+
#define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128 constexpr
|
|
27
|
+
#else
|
|
28
|
+
#define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256
|
|
29
|
+
#define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128
|
|
30
|
+
#endif
|
|
31
|
+
|
|
26
32
|
typedef short __v2hi __attribute__((__vector_size__(4)));
|
|
27
33
|
typedef char __v4qi __attribute__((__vector_size__(4)));
|
|
28
34
|
typedef char __v2qi __attribute__((__vector_size__(2)));
|
|
@@ -229,209 +235,183 @@ typedef char __v2qi __attribute__((__vector_size__(2)));
|
|
|
229
235
|
#define _mm256_mask_cmpneq_epu64_mask(k, A, B) \
|
|
230
236
|
_mm256_mask_cmp_epu64_mask((k), (A), (B), _MM_CMPINT_NE)
|
|
231
237
|
|
|
232
|
-
static __inline__ __m256i
|
|
233
|
-
_mm256_mask_add_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
234
|
-
{
|
|
238
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
239
|
+
_mm256_mask_add_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
|
|
235
240
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
236
241
|
(__v8si)_mm256_add_epi32(__A, __B),
|
|
237
242
|
(__v8si)__W);
|
|
238
243
|
}
|
|
239
244
|
|
|
240
|
-
static __inline__ __m256i
|
|
241
|
-
_mm256_maskz_add_epi32(__mmask8 __U, __m256i __A, __m256i __B)
|
|
242
|
-
{
|
|
245
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
246
|
+
_mm256_maskz_add_epi32(__mmask8 __U, __m256i __A, __m256i __B) {
|
|
243
247
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
244
248
|
(__v8si)_mm256_add_epi32(__A, __B),
|
|
245
249
|
(__v8si)_mm256_setzero_si256());
|
|
246
250
|
}
|
|
247
251
|
|
|
248
|
-
static __inline__ __m256i
|
|
249
|
-
_mm256_mask_add_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
250
|
-
{
|
|
252
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
253
|
+
_mm256_mask_add_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
|
|
251
254
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
252
255
|
(__v4di)_mm256_add_epi64(__A, __B),
|
|
253
256
|
(__v4di)__W);
|
|
254
257
|
}
|
|
255
258
|
|
|
256
|
-
static __inline__ __m256i
|
|
257
|
-
_mm256_maskz_add_epi64(__mmask8 __U, __m256i __A, __m256i __B)
|
|
258
|
-
{
|
|
259
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
260
|
+
_mm256_maskz_add_epi64(__mmask8 __U, __m256i __A, __m256i __B) {
|
|
259
261
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
260
262
|
(__v4di)_mm256_add_epi64(__A, __B),
|
|
261
263
|
(__v4di)_mm256_setzero_si256());
|
|
262
264
|
}
|
|
263
265
|
|
|
264
|
-
static __inline__ __m256i
|
|
265
|
-
_mm256_mask_sub_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
266
|
-
{
|
|
266
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
267
|
+
_mm256_mask_sub_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
|
|
267
268
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
268
269
|
(__v8si)_mm256_sub_epi32(__A, __B),
|
|
269
270
|
(__v8si)__W);
|
|
270
271
|
}
|
|
271
272
|
|
|
272
|
-
static __inline__ __m256i
|
|
273
|
-
_mm256_maskz_sub_epi32(__mmask8 __U, __m256i __A, __m256i __B)
|
|
274
|
-
{
|
|
273
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
274
|
+
_mm256_maskz_sub_epi32(__mmask8 __U, __m256i __A, __m256i __B) {
|
|
275
275
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
276
276
|
(__v8si)_mm256_sub_epi32(__A, __B),
|
|
277
277
|
(__v8si)_mm256_setzero_si256());
|
|
278
278
|
}
|
|
279
279
|
|
|
280
|
-
static __inline__ __m256i
|
|
281
|
-
_mm256_mask_sub_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
282
|
-
{
|
|
280
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
281
|
+
_mm256_mask_sub_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) {
|
|
283
282
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
284
283
|
(__v4di)_mm256_sub_epi64(__A, __B),
|
|
285
284
|
(__v4di)__W);
|
|
286
285
|
}
|
|
287
286
|
|
|
288
|
-
static __inline__ __m256i
|
|
289
|
-
_mm256_maskz_sub_epi64(__mmask8 __U, __m256i __A, __m256i __B)
|
|
290
|
-
{
|
|
287
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
288
|
+
_mm256_maskz_sub_epi64(__mmask8 __U, __m256i __A, __m256i __B) {
|
|
291
289
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
292
290
|
(__v4di)_mm256_sub_epi64(__A, __B),
|
|
293
291
|
(__v4di)_mm256_setzero_si256());
|
|
294
292
|
}
|
|
295
293
|
|
|
296
|
-
static __inline__ __m128i
|
|
297
|
-
_mm_mask_add_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
298
|
-
{
|
|
294
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
295
|
+
_mm_mask_add_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
299
296
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
300
297
|
(__v4si)_mm_add_epi32(__A, __B),
|
|
301
298
|
(__v4si)__W);
|
|
302
299
|
}
|
|
303
300
|
|
|
304
|
-
static __inline__ __m128i
|
|
305
|
-
_mm_maskz_add_epi32(__mmask8 __U, __m128i __A, __m128i __B)
|
|
306
|
-
{
|
|
301
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
302
|
+
_mm_maskz_add_epi32(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
307
303
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
308
304
|
(__v4si)_mm_add_epi32(__A, __B),
|
|
309
305
|
(__v4si)_mm_setzero_si128());
|
|
310
306
|
}
|
|
311
307
|
|
|
312
|
-
static __inline__ __m128i
|
|
313
|
-
_mm_mask_add_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
314
|
-
{
|
|
308
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
309
|
+
_mm_mask_add_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
315
310
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
316
311
|
(__v2di)_mm_add_epi64(__A, __B),
|
|
317
312
|
(__v2di)__W);
|
|
318
313
|
}
|
|
319
314
|
|
|
320
|
-
static __inline__ __m128i
|
|
321
|
-
_mm_maskz_add_epi64(__mmask8 __U, __m128i __A, __m128i __B)
|
|
322
|
-
{
|
|
315
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
316
|
+
_mm_maskz_add_epi64(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
323
317
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
324
318
|
(__v2di)_mm_add_epi64(__A, __B),
|
|
325
319
|
(__v2di)_mm_setzero_si128());
|
|
326
320
|
}
|
|
327
321
|
|
|
328
|
-
static __inline__ __m128i
|
|
329
|
-
_mm_mask_sub_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
330
|
-
{
|
|
322
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
323
|
+
_mm_mask_sub_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
331
324
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
332
325
|
(__v4si)_mm_sub_epi32(__A, __B),
|
|
333
326
|
(__v4si)__W);
|
|
334
327
|
}
|
|
335
328
|
|
|
336
|
-
static __inline__ __m128i
|
|
337
|
-
_mm_maskz_sub_epi32(__mmask8 __U, __m128i __A, __m128i __B)
|
|
338
|
-
{
|
|
329
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
330
|
+
_mm_maskz_sub_epi32(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
339
331
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
340
332
|
(__v4si)_mm_sub_epi32(__A, __B),
|
|
341
333
|
(__v4si)_mm_setzero_si128());
|
|
342
334
|
}
|
|
343
335
|
|
|
344
|
-
static __inline__ __m128i
|
|
345
|
-
_mm_mask_sub_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
346
|
-
{
|
|
336
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
337
|
+
_mm_mask_sub_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
347
338
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
348
339
|
(__v2di)_mm_sub_epi64(__A, __B),
|
|
349
340
|
(__v2di)__W);
|
|
350
341
|
}
|
|
351
342
|
|
|
352
|
-
static __inline__ __m128i
|
|
353
|
-
_mm_maskz_sub_epi64(__mmask8 __U, __m128i __A, __m128i __B)
|
|
354
|
-
{
|
|
343
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
344
|
+
_mm_maskz_sub_epi64(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
355
345
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
356
346
|
(__v2di)_mm_sub_epi64(__A, __B),
|
|
357
347
|
(__v2di)_mm_setzero_si128());
|
|
358
348
|
}
|
|
359
349
|
|
|
360
|
-
static __inline__ __m256i
|
|
361
|
-
_mm256_mask_mul_epi32(__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y)
|
|
362
|
-
{
|
|
350
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
351
|
+
_mm256_mask_mul_epi32(__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y) {
|
|
363
352
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
364
353
|
(__v4di)_mm256_mul_epi32(__X, __Y),
|
|
365
354
|
(__v4di)__W);
|
|
366
355
|
}
|
|
367
356
|
|
|
368
|
-
static __inline__ __m256i
|
|
369
|
-
_mm256_maskz_mul_epi32(__mmask8 __M, __m256i __X, __m256i __Y)
|
|
370
|
-
{
|
|
357
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
358
|
+
_mm256_maskz_mul_epi32(__mmask8 __M, __m256i __X, __m256i __Y) {
|
|
371
359
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
372
360
|
(__v4di)_mm256_mul_epi32(__X, __Y),
|
|
373
361
|
(__v4di)_mm256_setzero_si256());
|
|
374
362
|
}
|
|
375
363
|
|
|
376
|
-
static __inline__ __m128i
|
|
377
|
-
_mm_mask_mul_epi32(__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y)
|
|
378
|
-
{
|
|
364
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
365
|
+
_mm_mask_mul_epi32(__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y) {
|
|
379
366
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
380
367
|
(__v2di)_mm_mul_epi32(__X, __Y),
|
|
381
368
|
(__v2di)__W);
|
|
382
369
|
}
|
|
383
370
|
|
|
384
|
-
static __inline__ __m128i
|
|
385
|
-
_mm_maskz_mul_epi32(__mmask8 __M, __m128i __X, __m128i __Y)
|
|
386
|
-
{
|
|
371
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
372
|
+
_mm_maskz_mul_epi32(__mmask8 __M, __m128i __X, __m128i __Y) {
|
|
387
373
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
388
374
|
(__v2di)_mm_mul_epi32(__X, __Y),
|
|
389
375
|
(__v2di)_mm_setzero_si128());
|
|
390
376
|
}
|
|
391
377
|
|
|
392
|
-
static __inline__ __m256i
|
|
393
|
-
_mm256_mask_mul_epu32(__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y)
|
|
394
|
-
{
|
|
378
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
379
|
+
_mm256_mask_mul_epu32(__m256i __W, __mmask8 __M, __m256i __X, __m256i __Y) {
|
|
395
380
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
396
381
|
(__v4di)_mm256_mul_epu32(__X, __Y),
|
|
397
382
|
(__v4di)__W);
|
|
398
383
|
}
|
|
399
384
|
|
|
400
|
-
static __inline__ __m256i
|
|
401
|
-
_mm256_maskz_mul_epu32(__mmask8 __M, __m256i __X, __m256i __Y)
|
|
402
|
-
{
|
|
385
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
386
|
+
_mm256_maskz_mul_epu32(__mmask8 __M, __m256i __X, __m256i __Y) {
|
|
403
387
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
404
388
|
(__v4di)_mm256_mul_epu32(__X, __Y),
|
|
405
389
|
(__v4di)_mm256_setzero_si256());
|
|
406
390
|
}
|
|
407
391
|
|
|
408
|
-
static __inline__ __m128i
|
|
409
|
-
_mm_mask_mul_epu32(__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y)
|
|
410
|
-
{
|
|
392
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
393
|
+
_mm_mask_mul_epu32(__m128i __W, __mmask8 __M, __m128i __X, __m128i __Y) {
|
|
411
394
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
412
395
|
(__v2di)_mm_mul_epu32(__X, __Y),
|
|
413
396
|
(__v2di)__W);
|
|
414
397
|
}
|
|
415
398
|
|
|
416
|
-
static __inline__ __m128i
|
|
417
|
-
_mm_maskz_mul_epu32(__mmask8 __M, __m128i __X, __m128i __Y)
|
|
418
|
-
{
|
|
399
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
400
|
+
_mm_maskz_mul_epu32(__mmask8 __M, __m128i __X, __m128i __Y) {
|
|
419
401
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
420
402
|
(__v2di)_mm_mul_epu32(__X, __Y),
|
|
421
403
|
(__v2di)_mm_setzero_si128());
|
|
422
404
|
}
|
|
423
405
|
|
|
424
|
-
static __inline__ __m256i
|
|
425
|
-
_mm256_maskz_mullo_epi32(__mmask8 __M, __m256i __A, __m256i __B)
|
|
426
|
-
{
|
|
406
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
407
|
+
_mm256_maskz_mullo_epi32(__mmask8 __M, __m256i __A, __m256i __B) {
|
|
427
408
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
428
409
|
(__v8si)_mm256_mullo_epi32(__A, __B),
|
|
429
410
|
(__v8si)_mm256_setzero_si256());
|
|
430
411
|
}
|
|
431
412
|
|
|
432
|
-
static __inline__ __m256i
|
|
433
|
-
_mm256_mask_mullo_epi32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B)
|
|
434
|
-
{
|
|
413
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
414
|
+
_mm256_mask_mullo_epi32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) {
|
|
435
415
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
436
416
|
(__v8si)_mm256_mullo_epi32(__A, __B),
|
|
437
417
|
(__v8si)__W);
|
|
@@ -453,9 +433,8 @@ _mm_mask_mullo_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B)
|
|
|
453
433
|
(__v4si)__W);
|
|
454
434
|
}
|
|
455
435
|
|
|
456
|
-
static __inline__ __m256i
|
|
457
|
-
_mm256_and_epi32(__m256i __a, __m256i __b)
|
|
458
|
-
{
|
|
436
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
437
|
+
_mm256_and_epi32(__m256i __a, __m256i __b) {
|
|
459
438
|
return (__m256i)((__v8su)__a & (__v8su)__b);
|
|
460
439
|
}
|
|
461
440
|
|
|
@@ -473,9 +452,8 @@ _mm256_maskz_and_epi32(__mmask8 __U, __m256i __A, __m256i __B)
|
|
|
473
452
|
return (__m256i)_mm256_mask_and_epi32(_mm256_setzero_si256(), __U, __A, __B);
|
|
474
453
|
}
|
|
475
454
|
|
|
476
|
-
static __inline__ __m128i
|
|
477
|
-
_mm_and_epi32(__m128i __a, __m128i __b)
|
|
478
|
-
{
|
|
455
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
456
|
+
_mm_and_epi32(__m128i __a, __m128i __b) {
|
|
479
457
|
return (__m128i)((__v4su)__a & (__v4su)__b);
|
|
480
458
|
}
|
|
481
459
|
|
|
@@ -896,329 +874,312 @@ _mm_maskz_xor_epi64(__mmask8 __U, __m128i __A, __m128i __B)
|
|
|
896
874
|
(__v2df)(__m128d)(b), (int)(p), \
|
|
897
875
|
(__mmask8)(m)))
|
|
898
876
|
|
|
899
|
-
static __inline__ __m128d
|
|
900
|
-
_mm_mask_fmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
|
|
901
|
-
|
|
902
|
-
|
|
903
|
-
__builtin_ia32_vfmaddpd ((__v2df) __A,
|
|
904
|
-
(__v2df) __B,
|
|
905
|
-
(__v2df) __C),
|
|
906
|
-
(__v2df) __A);
|
|
877
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
878
|
+
_mm_mask_fmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) {
|
|
879
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
880
|
+
(__mmask8)__U, (__v2df)_mm_fmadd_pd(__A, __B, __C), (__v2df)__A);
|
|
907
881
|
}
|
|
908
882
|
|
|
909
|
-
static __inline__ __m128d
|
|
910
|
-
_mm_mask3_fmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U)
|
|
911
|
-
|
|
912
|
-
|
|
913
|
-
__builtin_ia32_vfmaddpd ((__v2df) __A,
|
|
914
|
-
(__v2df) __B,
|
|
915
|
-
(__v2df) __C),
|
|
916
|
-
(__v2df) __C);
|
|
883
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
884
|
+
_mm_mask3_fmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) {
|
|
885
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
886
|
+
(__mmask8)__U, (__v2df)_mm_fmadd_pd(__A, __B, __C), (__v2df)__C);
|
|
917
887
|
}
|
|
918
888
|
|
|
919
|
-
static __inline__ __m128d
|
|
920
|
-
_mm_maskz_fmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C)
|
|
921
|
-
|
|
922
|
-
|
|
923
|
-
|
|
924
|
-
(__v2df) __B,
|
|
925
|
-
(__v2df) __C),
|
|
926
|
-
(__v2df)_mm_setzero_pd());
|
|
889
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
890
|
+
_mm_maskz_fmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) {
|
|
891
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
892
|
+
(__mmask8)__U, (__v2df)_mm_fmadd_pd(__A, __B, __C),
|
|
893
|
+
(__v2df)_mm_setzero_pd());
|
|
927
894
|
}
|
|
928
895
|
|
|
929
|
-
static __inline__ __m128d
|
|
930
|
-
_mm_mask_fmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
|
|
931
|
-
|
|
932
|
-
|
|
933
|
-
__builtin_ia32_vfmaddpd ((__v2df) __A,
|
|
934
|
-
(__v2df) __B,
|
|
935
|
-
-(__v2df) __C),
|
|
936
|
-
(__v2df) __A);
|
|
896
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
897
|
+
_mm_mask_fmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) {
|
|
898
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
899
|
+
(__mmask8)__U, (__v2df)_mm_fmsub_pd(__A, __B, __C), (__v2df)__A);
|
|
937
900
|
}
|
|
938
901
|
|
|
939
|
-
static __inline__ __m128d
|
|
940
|
-
|
|
941
|
-
|
|
942
|
-
|
|
943
|
-
__builtin_ia32_vfmaddpd ((__v2df) __A,
|
|
944
|
-
(__v2df) __B,
|
|
945
|
-
-(__v2df) __C),
|
|
946
|
-
(__v2df)_mm_setzero_pd());
|
|
902
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
903
|
+
_mm_mask3_fmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) {
|
|
904
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
905
|
+
(__mmask8)__U, (__v2df)_mm_fmsub_pd(__A, __B, __C), (__v2df)__C);
|
|
947
906
|
}
|
|
948
907
|
|
|
949
|
-
static __inline__ __m128d
|
|
950
|
-
|
|
951
|
-
|
|
952
|
-
|
|
953
|
-
|
|
954
|
-
(__v2df) __B,
|
|
955
|
-
(__v2df) __C),
|
|
956
|
-
(__v2df) __C);
|
|
908
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
909
|
+
_mm_maskz_fmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) {
|
|
910
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
911
|
+
(__mmask8)__U, (__v2df)_mm_fmsub_pd(__A, __B, __C),
|
|
912
|
+
(__v2df)_mm_setzero_pd());
|
|
957
913
|
}
|
|
958
914
|
|
|
959
|
-
static __inline__ __m128d
|
|
960
|
-
|
|
961
|
-
|
|
962
|
-
|
|
963
|
-
__builtin_ia32_vfmaddpd (-(__v2df) __A,
|
|
964
|
-
(__v2df) __B,
|
|
965
|
-
(__v2df) __C),
|
|
966
|
-
(__v2df)_mm_setzero_pd());
|
|
915
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
916
|
+
_mm_mask_fnmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) {
|
|
917
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
918
|
+
(__mmask8)__U, (__v2df)_mm_fnmadd_pd(__A, __B, __C), (__v2df)__A);
|
|
967
919
|
}
|
|
968
920
|
|
|
969
|
-
static __inline__ __m128d
|
|
970
|
-
|
|
971
|
-
|
|
972
|
-
|
|
973
|
-
__builtin_ia32_vfmaddpd (-(__v2df) __A,
|
|
974
|
-
(__v2df) __B,
|
|
975
|
-
-(__v2df) __C),
|
|
976
|
-
(__v2df)_mm_setzero_pd());
|
|
921
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
922
|
+
_mm_mask3_fnmadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) {
|
|
923
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
924
|
+
(__mmask8)__U, (__v2df)_mm_fnmadd_pd(__A, __B, __C), (__v2df)__C);
|
|
977
925
|
}
|
|
978
926
|
|
|
979
|
-
static __inline__
|
|
980
|
-
|
|
981
|
-
|
|
982
|
-
|
|
983
|
-
|
|
984
|
-
(__v4df) __B,
|
|
985
|
-
(__v4df) __C),
|
|
986
|
-
(__v4df) __A);
|
|
927
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
928
|
+
_mm_maskz_fnmadd_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) {
|
|
929
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
930
|
+
(__mmask8)__U, (__v2df)_mm_fnmadd_pd(__A, __B, __C),
|
|
931
|
+
(__v2df)_mm_setzero_pd());
|
|
987
932
|
}
|
|
988
933
|
|
|
989
|
-
static __inline__
|
|
990
|
-
|
|
991
|
-
|
|
992
|
-
|
|
993
|
-
__builtin_ia32_vfmaddpd256 ((__v4df) __A,
|
|
994
|
-
(__v4df) __B,
|
|
995
|
-
(__v4df) __C),
|
|
996
|
-
(__v4df) __C);
|
|
934
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
935
|
+
_mm_mask_fnmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) {
|
|
936
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
937
|
+
(__mmask8)__U, (__v2df)_mm_fnmsub_pd(__A, __B, __C), (__v2df)__A);
|
|
997
938
|
}
|
|
998
939
|
|
|
999
|
-
static __inline__
|
|
1000
|
-
|
|
1001
|
-
|
|
1002
|
-
|
|
1003
|
-
__builtin_ia32_vfmaddpd256 ((__v4df) __A,
|
|
1004
|
-
(__v4df) __B,
|
|
1005
|
-
(__v4df) __C),
|
|
1006
|
-
(__v4df)_mm256_setzero_pd());
|
|
940
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
941
|
+
_mm_mask3_fnmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U) {
|
|
942
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
943
|
+
(__mmask8)__U, (__v2df)_mm_fnmsub_pd(__A, __B, __C), (__v2df)__C);
|
|
1007
944
|
}
|
|
1008
945
|
|
|
1009
|
-
static __inline__
|
|
1010
|
-
|
|
1011
|
-
|
|
1012
|
-
|
|
1013
|
-
|
|
1014
|
-
(__v4df) __B,
|
|
1015
|
-
-(__v4df) __C),
|
|
1016
|
-
(__v4df) __A);
|
|
946
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
947
|
+
_mm_maskz_fnmsub_pd(__mmask8 __U, __m128d __A, __m128d __B, __m128d __C) {
|
|
948
|
+
return (__m128d)__builtin_ia32_selectpd_128(
|
|
949
|
+
(__mmask8)__U, (__v2df)_mm_fnmsub_pd(__A, __B, __C),
|
|
950
|
+
(__v2df)_mm_setzero_pd());
|
|
1017
951
|
}
|
|
1018
952
|
|
|
1019
|
-
static __inline__ __m256d
|
|
1020
|
-
|
|
1021
|
-
|
|
1022
|
-
|
|
1023
|
-
__builtin_ia32_vfmaddpd256 ((__v4df) __A,
|
|
1024
|
-
(__v4df) __B,
|
|
1025
|
-
-(__v4df) __C),
|
|
1026
|
-
(__v4df)_mm256_setzero_pd());
|
|
953
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
954
|
+
_mm256_mask_fmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) {
|
|
955
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
956
|
+
(__mmask8)__U, (__v4df)_mm256_fmadd_pd(__A, __B, __C), (__v4df)__A);
|
|
1027
957
|
}
|
|
1028
958
|
|
|
1029
|
-
static __inline__ __m256d
|
|
1030
|
-
|
|
1031
|
-
|
|
1032
|
-
|
|
1033
|
-
__builtin_ia32_vfmaddpd256 (-(__v4df) __A,
|
|
1034
|
-
(__v4df) __B,
|
|
1035
|
-
(__v4df) __C),
|
|
1036
|
-
(__v4df) __C);
|
|
959
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
960
|
+
_mm256_mask3_fmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) {
|
|
961
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
962
|
+
(__mmask8)__U, (__v4df)_mm256_fmadd_pd(__A, __B, __C), (__v4df)__C);
|
|
1037
963
|
}
|
|
1038
964
|
|
|
1039
|
-
static __inline__ __m256d
|
|
1040
|
-
|
|
1041
|
-
|
|
1042
|
-
|
|
1043
|
-
|
|
1044
|
-
(__v4df) __B,
|
|
1045
|
-
(__v4df) __C),
|
|
1046
|
-
(__v4df)_mm256_setzero_pd());
|
|
965
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
966
|
+
_mm256_maskz_fmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) {
|
|
967
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
968
|
+
(__mmask8)__U, (__v4df)_mm256_fmadd_pd(__A, __B, __C),
|
|
969
|
+
(__v4df)_mm256_setzero_pd());
|
|
1047
970
|
}
|
|
1048
971
|
|
|
1049
|
-
static __inline__ __m256d
|
|
1050
|
-
|
|
1051
|
-
|
|
1052
|
-
|
|
1053
|
-
__builtin_ia32_vfmaddpd256 (-(__v4df) __A,
|
|
1054
|
-
(__v4df) __B,
|
|
1055
|
-
-(__v4df) __C),
|
|
1056
|
-
(__v4df)_mm256_setzero_pd());
|
|
972
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
973
|
+
_mm256_mask_fmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) {
|
|
974
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
975
|
+
(__mmask8)__U, (__v4df)_mm256_fmsub_pd(__A, __B, __C), (__v4df)__A);
|
|
1057
976
|
}
|
|
1058
977
|
|
|
1059
|
-
static __inline__
|
|
1060
|
-
|
|
1061
|
-
|
|
1062
|
-
|
|
1063
|
-
__builtin_ia32_vfmaddps ((__v4sf) __A,
|
|
1064
|
-
(__v4sf) __B,
|
|
1065
|
-
(__v4sf) __C),
|
|
1066
|
-
(__v4sf) __A);
|
|
978
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
979
|
+
_mm256_mask3_fmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) {
|
|
980
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
981
|
+
(__mmask8)__U, (__v4df)_mm256_fmsub_pd(__A, __B, __C), (__v4df)__C);
|
|
1067
982
|
}
|
|
1068
983
|
|
|
1069
|
-
static __inline__
|
|
1070
|
-
|
|
1071
|
-
|
|
1072
|
-
|
|
1073
|
-
|
|
1074
|
-
(__v4sf) __B,
|
|
1075
|
-
(__v4sf) __C),
|
|
1076
|
-
(__v4sf) __C);
|
|
984
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
985
|
+
_mm256_maskz_fmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) {
|
|
986
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
987
|
+
(__mmask8)__U, (__v4df)_mm256_fmsub_pd(__A, __B, __C),
|
|
988
|
+
(__v4df)_mm256_setzero_pd());
|
|
1077
989
|
}
|
|
1078
990
|
|
|
1079
|
-
static __inline__
|
|
1080
|
-
|
|
1081
|
-
|
|
1082
|
-
|
|
1083
|
-
__builtin_ia32_vfmaddps ((__v4sf) __A,
|
|
1084
|
-
(__v4sf) __B,
|
|
1085
|
-
(__v4sf) __C),
|
|
1086
|
-
(__v4sf)_mm_setzero_ps());
|
|
991
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
992
|
+
_mm256_mask_fnmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) {
|
|
993
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
994
|
+
(__mmask8)__U, (__v4df)_mm256_fnmadd_pd(__A, __B, __C), (__v4df)__A);
|
|
1087
995
|
}
|
|
1088
996
|
|
|
1089
|
-
static __inline__
|
|
1090
|
-
|
|
1091
|
-
|
|
1092
|
-
|
|
1093
|
-
__builtin_ia32_vfmaddps ((__v4sf) __A,
|
|
1094
|
-
(__v4sf) __B,
|
|
1095
|
-
-(__v4sf) __C),
|
|
1096
|
-
(__v4sf) __A);
|
|
997
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
998
|
+
_mm256_mask3_fnmadd_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) {
|
|
999
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
1000
|
+
(__mmask8)__U, (__v4df)_mm256_fnmadd_pd(__A, __B, __C), (__v4df)__C);
|
|
1097
1001
|
}
|
|
1098
1002
|
|
|
1099
|
-
static __inline__
|
|
1100
|
-
|
|
1101
|
-
|
|
1102
|
-
|
|
1103
|
-
|
|
1104
|
-
(__v4sf) __B,
|
|
1105
|
-
-(__v4sf) __C),
|
|
1106
|
-
(__v4sf)_mm_setzero_ps());
|
|
1003
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1004
|
+
_mm256_maskz_fnmadd_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) {
|
|
1005
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
1006
|
+
(__mmask8)__U, (__v4df)_mm256_fnmadd_pd(__A, __B, __C),
|
|
1007
|
+
(__v4df)_mm256_setzero_pd());
|
|
1107
1008
|
}
|
|
1108
1009
|
|
|
1109
|
-
static __inline__
|
|
1110
|
-
|
|
1111
|
-
|
|
1112
|
-
|
|
1113
|
-
__builtin_ia32_vfmaddps (-(__v4sf) __A,
|
|
1114
|
-
(__v4sf) __B,
|
|
1115
|
-
(__v4sf) __C),
|
|
1116
|
-
(__v4sf) __C);
|
|
1010
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1011
|
+
_mm256_mask_fnmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C) {
|
|
1012
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
1013
|
+
(__mmask8)__U, (__v4df)_mm256_fnmsub_pd(__A, __B, __C), (__v4df)__A);
|
|
1117
1014
|
}
|
|
1118
1015
|
|
|
1119
|
-
static __inline__
|
|
1120
|
-
|
|
1121
|
-
|
|
1122
|
-
|
|
1123
|
-
__builtin_ia32_vfmaddps (-(__v4sf) __A,
|
|
1124
|
-
(__v4sf) __B,
|
|
1125
|
-
(__v4sf) __C),
|
|
1126
|
-
(__v4sf)_mm_setzero_ps());
|
|
1016
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1017
|
+
_mm256_mask3_fnmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U) {
|
|
1018
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
1019
|
+
(__mmask8)__U, (__v4df)_mm256_fnmsub_pd(__A, __B, __C), (__v4df)__C);
|
|
1127
1020
|
}
|
|
1128
1021
|
|
|
1129
|
-
static __inline__
|
|
1130
|
-
|
|
1131
|
-
|
|
1132
|
-
|
|
1133
|
-
|
|
1134
|
-
(__v4sf) __B,
|
|
1135
|
-
-(__v4sf) __C),
|
|
1136
|
-
(__v4sf)_mm_setzero_ps());
|
|
1022
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1023
|
+
_mm256_maskz_fnmsub_pd(__mmask8 __U, __m256d __A, __m256d __B, __m256d __C) {
|
|
1024
|
+
return (__m256d)__builtin_ia32_selectpd_256(
|
|
1025
|
+
(__mmask8)__U, (__v4df)_mm256_fnmsub_pd(__A, __B, __C),
|
|
1026
|
+
(__v4df)_mm256_setzero_pd());
|
|
1137
1027
|
}
|
|
1138
1028
|
|
|
1139
|
-
static __inline__
|
|
1140
|
-
|
|
1141
|
-
|
|
1142
|
-
|
|
1143
|
-
__builtin_ia32_vfmaddps256 ((__v8sf) __A,
|
|
1144
|
-
(__v8sf) __B,
|
|
1145
|
-
(__v8sf) __C),
|
|
1146
|
-
(__v8sf) __A);
|
|
1029
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1030
|
+
_mm_mask_fmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) {
|
|
1031
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1032
|
+
(__mmask8)__U, (__v4sf)_mm_fmadd_ps(__A, __B, __C), (__v4sf)__A);
|
|
1147
1033
|
}
|
|
1148
1034
|
|
|
1149
|
-
static __inline__
|
|
1150
|
-
|
|
1151
|
-
|
|
1152
|
-
|
|
1153
|
-
__builtin_ia32_vfmaddps256 ((__v8sf) __A,
|
|
1154
|
-
(__v8sf) __B,
|
|
1155
|
-
(__v8sf) __C),
|
|
1156
|
-
(__v8sf) __C);
|
|
1035
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1036
|
+
_mm_mask3_fmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) {
|
|
1037
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1038
|
+
(__mmask8)__U, (__v4sf)_mm_fmadd_ps(__A, __B, __C), (__v4sf)__C);
|
|
1157
1039
|
}
|
|
1158
1040
|
|
|
1159
|
-
static __inline__
|
|
1160
|
-
|
|
1161
|
-
|
|
1162
|
-
|
|
1163
|
-
|
|
1164
|
-
(__v8sf) __B,
|
|
1165
|
-
(__v8sf) __C),
|
|
1166
|
-
(__v8sf)_mm256_setzero_ps());
|
|
1041
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1042
|
+
_mm_maskz_fmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) {
|
|
1043
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1044
|
+
(__mmask8)__U, (__v4sf)_mm_fmadd_ps(__A, __B, __C),
|
|
1045
|
+
(__v4sf)_mm_setzero_ps());
|
|
1167
1046
|
}
|
|
1168
1047
|
|
|
1169
|
-
static __inline__
|
|
1170
|
-
|
|
1171
|
-
|
|
1172
|
-
|
|
1173
|
-
__builtin_ia32_vfmaddps256 ((__v8sf) __A,
|
|
1174
|
-
(__v8sf) __B,
|
|
1175
|
-
-(__v8sf) __C),
|
|
1176
|
-
(__v8sf) __A);
|
|
1048
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1049
|
+
_mm_mask_fmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) {
|
|
1050
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1051
|
+
(__mmask8)__U, (__v4sf)_mm_fmsub_ps(__A, __B, __C), (__v4sf)__A);
|
|
1177
1052
|
}
|
|
1178
1053
|
|
|
1179
|
-
static __inline__
|
|
1180
|
-
|
|
1181
|
-
|
|
1182
|
-
|
|
1183
|
-
__builtin_ia32_vfmaddps256 ((__v8sf) __A,
|
|
1184
|
-
(__v8sf) __B,
|
|
1185
|
-
-(__v8sf) __C),
|
|
1186
|
-
(__v8sf)_mm256_setzero_ps());
|
|
1054
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1055
|
+
_mm_mask3_fmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) {
|
|
1056
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1057
|
+
(__mmask8)__U, (__v4sf)_mm_fmsub_ps(__A, __B, __C), (__v4sf)__C);
|
|
1187
1058
|
}
|
|
1188
1059
|
|
|
1189
|
-
static __inline__
|
|
1190
|
-
|
|
1191
|
-
|
|
1192
|
-
|
|
1193
|
-
|
|
1194
|
-
(__v8sf) __B,
|
|
1195
|
-
(__v8sf) __C),
|
|
1196
|
-
(__v8sf) __C);
|
|
1060
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1061
|
+
_mm_maskz_fmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) {
|
|
1062
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1063
|
+
(__mmask8)__U, (__v4sf)_mm_fmsub_ps(__A, __B, __C),
|
|
1064
|
+
(__v4sf)_mm_setzero_ps());
|
|
1197
1065
|
}
|
|
1198
1066
|
|
|
1199
|
-
static __inline__
|
|
1200
|
-
|
|
1201
|
-
|
|
1202
|
-
|
|
1203
|
-
__builtin_ia32_vfmaddps256 (-(__v8sf) __A,
|
|
1204
|
-
(__v8sf) __B,
|
|
1205
|
-
(__v8sf) __C),
|
|
1206
|
-
(__v8sf)_mm256_setzero_ps());
|
|
1067
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1068
|
+
_mm_mask_fnmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) {
|
|
1069
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1070
|
+
(__mmask8)__U, (__v4sf)_mm_fnmadd_ps(__A, __B, __C), (__v4sf)__A);
|
|
1207
1071
|
}
|
|
1208
1072
|
|
|
1209
|
-
static __inline__
|
|
1210
|
-
|
|
1211
|
-
|
|
1212
|
-
|
|
1213
|
-
|
|
1214
|
-
|
|
1215
|
-
|
|
1216
|
-
|
|
1073
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1074
|
+
_mm_mask3_fnmadd_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) {
|
|
1075
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1076
|
+
(__mmask8)__U, (__v4sf)_mm_fnmadd_ps(__A, __B, __C), (__v4sf)__C);
|
|
1077
|
+
}
|
|
1078
|
+
|
|
1079
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1080
|
+
_mm_maskz_fnmadd_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) {
|
|
1081
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1082
|
+
(__mmask8)__U, (__v4sf)_mm_fnmadd_ps(__A, __B, __C),
|
|
1083
|
+
(__v4sf)_mm_setzero_ps());
|
|
1084
|
+
}
|
|
1085
|
+
|
|
1086
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1087
|
+
_mm_mask_fnmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C) {
|
|
1088
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1089
|
+
(__mmask8)__U, (__v4sf)_mm_fnmsub_ps(__A, __B, __C), (__v4sf)__A);
|
|
1090
|
+
}
|
|
1091
|
+
|
|
1092
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1093
|
+
_mm_mask3_fnmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U) {
|
|
1094
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1095
|
+
(__mmask8)__U, (__v4sf)_mm_fnmsub_ps(__A, __B, __C), (__v4sf)__C);
|
|
1096
|
+
}
|
|
1097
|
+
|
|
1098
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1099
|
+
_mm_maskz_fnmsub_ps(__mmask8 __U, __m128 __A, __m128 __B, __m128 __C) {
|
|
1100
|
+
return (__m128)__builtin_ia32_selectps_128(
|
|
1101
|
+
(__mmask8)__U, (__v4sf)_mm_fnmsub_ps(__A, __B, __C),
|
|
1102
|
+
(__v4sf)_mm_setzero_ps());
|
|
1103
|
+
}
|
|
1104
|
+
|
|
1105
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1106
|
+
_mm256_mask_fmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) {
|
|
1107
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1108
|
+
(__mmask8)__U, (__v8sf)_mm256_fmadd_ps(__A, __B, __C), (__v8sf)__A);
|
|
1109
|
+
}
|
|
1110
|
+
|
|
1111
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1112
|
+
_mm256_mask3_fmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) {
|
|
1113
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1114
|
+
(__mmask8)__U, (__v8sf)_mm256_fmadd_ps(__A, __B, __C), (__v8sf)__C);
|
|
1115
|
+
}
|
|
1116
|
+
|
|
1117
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1118
|
+
_mm256_maskz_fmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) {
|
|
1119
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1120
|
+
(__mmask8)__U, (__v8sf)_mm256_fmadd_ps(__A, __B, __C),
|
|
1121
|
+
(__v8sf)_mm256_setzero_ps());
|
|
1122
|
+
}
|
|
1123
|
+
|
|
1124
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1125
|
+
_mm256_mask_fmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) {
|
|
1126
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1127
|
+
(__mmask8)__U, (__v8sf)_mm256_fmsub_ps(__A, __B, __C), (__v8sf)__A);
|
|
1128
|
+
}
|
|
1129
|
+
|
|
1130
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1131
|
+
_mm256_mask3_fmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) {
|
|
1132
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1133
|
+
(__mmask8)__U, (__v8sf)_mm256_fmsub_ps(__A, __B, __C), (__v8sf)__C);
|
|
1134
|
+
}
|
|
1135
|
+
|
|
1136
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1137
|
+
_mm256_maskz_fmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) {
|
|
1138
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1139
|
+
(__mmask8)__U, (__v8sf)_mm256_fmsub_ps(__A, __B, __C),
|
|
1140
|
+
(__v8sf)_mm256_setzero_ps());
|
|
1141
|
+
}
|
|
1142
|
+
|
|
1143
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1144
|
+
_mm256_mask_fnmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) {
|
|
1145
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1146
|
+
(__mmask8)__U, (__v8sf)_mm256_fnmadd_ps(__A, __B, __C), (__v8sf)__A);
|
|
1147
|
+
}
|
|
1148
|
+
|
|
1149
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1150
|
+
_mm256_mask3_fnmadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) {
|
|
1151
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1152
|
+
(__mmask8)__U, (__v8sf)_mm256_fnmadd_ps(__A, __B, __C), (__v8sf)__C);
|
|
1153
|
+
}
|
|
1154
|
+
|
|
1155
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1156
|
+
_mm256_maskz_fnmadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) {
|
|
1157
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1158
|
+
(__mmask8)__U, (__v8sf)_mm256_fnmadd_ps(__A, __B, __C),
|
|
1159
|
+
(__v8sf)_mm256_setzero_ps());
|
|
1160
|
+
}
|
|
1161
|
+
|
|
1162
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1163
|
+
_mm256_mask_fnmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C) {
|
|
1164
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1165
|
+
(__mmask8)__U, (__v8sf)_mm256_fnmsub_ps(__A, __B, __C), (__v8sf)__A);
|
|
1166
|
+
}
|
|
1167
|
+
|
|
1168
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1169
|
+
_mm256_mask3_fnmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U) {
|
|
1170
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1171
|
+
(__mmask8)__U, (__v8sf)_mm256_fnmsub_ps(__A, __B, __C), (__v8sf)__C);
|
|
1172
|
+
}
|
|
1173
|
+
|
|
1174
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1175
|
+
_mm256_maskz_fnmsub_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C) {
|
|
1176
|
+
return (__m256)__builtin_ia32_selectps_256(
|
|
1177
|
+
(__mmask8)__U, (__v8sf)_mm256_fnmsub_ps(__A, __B, __C),
|
|
1178
|
+
(__v8sf)_mm256_setzero_ps());
|
|
1217
1179
|
}
|
|
1218
1180
|
|
|
1219
1181
|
static __inline__ __m128d __DEFAULT_FN_ATTRS128
|
|
1220
|
-
_mm_mask_fmaddsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
|
|
1221
|
-
{
|
|
1182
|
+
_mm_mask_fmaddsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C) {
|
|
1222
1183
|
return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U,
|
|
1223
1184
|
__builtin_ia32_vfmaddsubpd ((__v2df) __A,
|
|
1224
1185
|
(__v2df) __B,
|
|
@@ -1417,46 +1378,6 @@ _mm256_maskz_fmsubadd_ps(__mmask8 __U, __m256 __A, __m256 __B, __m256 __C)
|
|
|
1417
1378
|
(__v8sf)_mm256_setzero_ps());
|
|
1418
1379
|
}
|
|
1419
1380
|
|
|
1420
|
-
static __inline__ __m128d __DEFAULT_FN_ATTRS128
|
|
1421
|
-
_mm_mask3_fmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U)
|
|
1422
|
-
{
|
|
1423
|
-
return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U,
|
|
1424
|
-
__builtin_ia32_vfmaddpd ((__v2df) __A,
|
|
1425
|
-
(__v2df) __B,
|
|
1426
|
-
-(__v2df) __C),
|
|
1427
|
-
(__v2df) __C);
|
|
1428
|
-
}
|
|
1429
|
-
|
|
1430
|
-
static __inline__ __m256d __DEFAULT_FN_ATTRS256
|
|
1431
|
-
_mm256_mask3_fmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U)
|
|
1432
|
-
{
|
|
1433
|
-
return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U,
|
|
1434
|
-
__builtin_ia32_vfmaddpd256 ((__v4df) __A,
|
|
1435
|
-
(__v4df) __B,
|
|
1436
|
-
-(__v4df) __C),
|
|
1437
|
-
(__v4df) __C);
|
|
1438
|
-
}
|
|
1439
|
-
|
|
1440
|
-
static __inline__ __m128 __DEFAULT_FN_ATTRS128
|
|
1441
|
-
_mm_mask3_fmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
|
|
1442
|
-
{
|
|
1443
|
-
return (__m128) __builtin_ia32_selectps_128((__mmask8) __U,
|
|
1444
|
-
__builtin_ia32_vfmaddps ((__v4sf) __A,
|
|
1445
|
-
(__v4sf) __B,
|
|
1446
|
-
-(__v4sf) __C),
|
|
1447
|
-
(__v4sf) __C);
|
|
1448
|
-
}
|
|
1449
|
-
|
|
1450
|
-
static __inline__ __m256 __DEFAULT_FN_ATTRS256
|
|
1451
|
-
_mm256_mask3_fmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U)
|
|
1452
|
-
{
|
|
1453
|
-
return (__m256) __builtin_ia32_selectps_256((__mmask8) __U,
|
|
1454
|
-
__builtin_ia32_vfmaddps256 ((__v8sf) __A,
|
|
1455
|
-
(__v8sf) __B,
|
|
1456
|
-
-(__v8sf) __C),
|
|
1457
|
-
(__v8sf) __C);
|
|
1458
|
-
}
|
|
1459
|
-
|
|
1460
1381
|
static __inline__ __m128d __DEFAULT_FN_ATTRS128
|
|
1461
1382
|
_mm_mask3_fmsubadd_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U)
|
|
1462
1383
|
{
|
|
@@ -1497,233 +1418,113 @@ _mm256_mask3_fmsubadd_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U)
|
|
|
1497
1418
|
(__v8sf) __C);
|
|
1498
1419
|
}
|
|
1499
1420
|
|
|
1500
|
-
static __inline__ __m128d
|
|
1501
|
-
_mm_mask_fnmadd_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
|
|
1502
|
-
{
|
|
1503
|
-
return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U,
|
|
1504
|
-
__builtin_ia32_vfmaddpd ((__v2df) __A,
|
|
1505
|
-
-(__v2df) __B,
|
|
1506
|
-
(__v2df) __C),
|
|
1507
|
-
(__v2df) __A);
|
|
1508
|
-
}
|
|
1509
|
-
|
|
1510
|
-
static __inline__ __m256d __DEFAULT_FN_ATTRS256
|
|
1511
|
-
_mm256_mask_fnmadd_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C)
|
|
1512
|
-
{
|
|
1513
|
-
return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U,
|
|
1514
|
-
__builtin_ia32_vfmaddpd256 ((__v4df) __A,
|
|
1515
|
-
-(__v4df) __B,
|
|
1516
|
-
(__v4df) __C),
|
|
1517
|
-
(__v4df) __A);
|
|
1518
|
-
}
|
|
1519
|
-
|
|
1520
|
-
static __inline__ __m128 __DEFAULT_FN_ATTRS128
|
|
1521
|
-
_mm_mask_fnmadd_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
|
|
1522
|
-
{
|
|
1523
|
-
return (__m128) __builtin_ia32_selectps_128((__mmask8) __U,
|
|
1524
|
-
__builtin_ia32_vfmaddps ((__v4sf) __A,
|
|
1525
|
-
-(__v4sf) __B,
|
|
1526
|
-
(__v4sf) __C),
|
|
1527
|
-
(__v4sf) __A);
|
|
1528
|
-
}
|
|
1529
|
-
|
|
1530
|
-
static __inline__ __m256 __DEFAULT_FN_ATTRS256
|
|
1531
|
-
_mm256_mask_fnmadd_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C)
|
|
1532
|
-
{
|
|
1533
|
-
return (__m256) __builtin_ia32_selectps_256((__mmask8) __U,
|
|
1534
|
-
__builtin_ia32_vfmaddps256 ((__v8sf) __A,
|
|
1535
|
-
-(__v8sf) __B,
|
|
1536
|
-
(__v8sf) __C),
|
|
1537
|
-
(__v8sf) __A);
|
|
1538
|
-
}
|
|
1539
|
-
|
|
1540
|
-
static __inline__ __m128d __DEFAULT_FN_ATTRS128
|
|
1541
|
-
_mm_mask_fnmsub_pd(__m128d __A, __mmask8 __U, __m128d __B, __m128d __C)
|
|
1542
|
-
{
|
|
1543
|
-
return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U,
|
|
1544
|
-
__builtin_ia32_vfmaddpd ((__v2df) __A,
|
|
1545
|
-
-(__v2df) __B,
|
|
1546
|
-
-(__v2df) __C),
|
|
1547
|
-
(__v2df) __A);
|
|
1548
|
-
}
|
|
1549
|
-
|
|
1550
|
-
static __inline__ __m128d __DEFAULT_FN_ATTRS128
|
|
1551
|
-
_mm_mask3_fnmsub_pd(__m128d __A, __m128d __B, __m128d __C, __mmask8 __U)
|
|
1552
|
-
{
|
|
1553
|
-
return (__m128d) __builtin_ia32_selectpd_128((__mmask8) __U,
|
|
1554
|
-
__builtin_ia32_vfmaddpd ((__v2df) __A,
|
|
1555
|
-
-(__v2df) __B,
|
|
1556
|
-
-(__v2df) __C),
|
|
1557
|
-
(__v2df) __C);
|
|
1558
|
-
}
|
|
1559
|
-
|
|
1560
|
-
static __inline__ __m256d __DEFAULT_FN_ATTRS256
|
|
1561
|
-
_mm256_mask_fnmsub_pd(__m256d __A, __mmask8 __U, __m256d __B, __m256d __C)
|
|
1562
|
-
{
|
|
1563
|
-
return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U,
|
|
1564
|
-
__builtin_ia32_vfmaddpd256 ((__v4df) __A,
|
|
1565
|
-
-(__v4df) __B,
|
|
1566
|
-
-(__v4df) __C),
|
|
1567
|
-
(__v4df) __A);
|
|
1568
|
-
}
|
|
1569
|
-
|
|
1570
|
-
static __inline__ __m256d __DEFAULT_FN_ATTRS256
|
|
1571
|
-
_mm256_mask3_fnmsub_pd(__m256d __A, __m256d __B, __m256d __C, __mmask8 __U)
|
|
1572
|
-
{
|
|
1573
|
-
return (__m256d) __builtin_ia32_selectpd_256((__mmask8) __U,
|
|
1574
|
-
__builtin_ia32_vfmaddpd256 ((__v4df) __A,
|
|
1575
|
-
-(__v4df) __B,
|
|
1576
|
-
-(__v4df) __C),
|
|
1577
|
-
(__v4df) __C);
|
|
1578
|
-
}
|
|
1579
|
-
|
|
1580
|
-
static __inline__ __m128 __DEFAULT_FN_ATTRS128
|
|
1581
|
-
_mm_mask_fnmsub_ps(__m128 __A, __mmask8 __U, __m128 __B, __m128 __C)
|
|
1582
|
-
{
|
|
1583
|
-
return (__m128) __builtin_ia32_selectps_128((__mmask8) __U,
|
|
1584
|
-
__builtin_ia32_vfmaddps ((__v4sf) __A,
|
|
1585
|
-
-(__v4sf) __B,
|
|
1586
|
-
-(__v4sf) __C),
|
|
1587
|
-
(__v4sf) __A);
|
|
1588
|
-
}
|
|
1589
|
-
|
|
1590
|
-
static __inline__ __m128 __DEFAULT_FN_ATTRS128
|
|
1591
|
-
_mm_mask3_fnmsub_ps(__m128 __A, __m128 __B, __m128 __C, __mmask8 __U)
|
|
1592
|
-
{
|
|
1593
|
-
return (__m128) __builtin_ia32_selectps_128((__mmask8) __U,
|
|
1594
|
-
__builtin_ia32_vfmaddps ((__v4sf) __A,
|
|
1595
|
-
-(__v4sf) __B,
|
|
1596
|
-
-(__v4sf) __C),
|
|
1597
|
-
(__v4sf) __C);
|
|
1598
|
-
}
|
|
1599
|
-
|
|
1600
|
-
static __inline__ __m256 __DEFAULT_FN_ATTRS256
|
|
1601
|
-
_mm256_mask_fnmsub_ps(__m256 __A, __mmask8 __U, __m256 __B, __m256 __C)
|
|
1602
|
-
{
|
|
1603
|
-
return (__m256) __builtin_ia32_selectps_256((__mmask8) __U,
|
|
1604
|
-
__builtin_ia32_vfmaddps256 ((__v8sf) __A,
|
|
1605
|
-
-(__v8sf) __B,
|
|
1606
|
-
-(__v8sf) __C),
|
|
1607
|
-
(__v8sf) __A);
|
|
1608
|
-
}
|
|
1609
|
-
|
|
1610
|
-
static __inline__ __m256 __DEFAULT_FN_ATTRS256
|
|
1611
|
-
_mm256_mask3_fnmsub_ps(__m256 __A, __m256 __B, __m256 __C, __mmask8 __U)
|
|
1612
|
-
{
|
|
1613
|
-
return (__m256) __builtin_ia32_selectps_256((__mmask8) __U,
|
|
1614
|
-
__builtin_ia32_vfmaddps256 ((__v8sf) __A,
|
|
1615
|
-
-(__v8sf) __B,
|
|
1616
|
-
-(__v8sf) __C),
|
|
1617
|
-
(__v8sf) __C);
|
|
1618
|
-
}
|
|
1619
|
-
|
|
1620
|
-
static __inline__ __m128d __DEFAULT_FN_ATTRS128
|
|
1421
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1621
1422
|
_mm_mask_add_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
|
|
1622
1423
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
1623
1424
|
(__v2df)_mm_add_pd(__A, __B),
|
|
1624
1425
|
(__v2df)__W);
|
|
1625
1426
|
}
|
|
1626
1427
|
|
|
1627
|
-
static __inline__ __m128d
|
|
1428
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1628
1429
|
_mm_maskz_add_pd(__mmask8 __U, __m128d __A, __m128d __B) {
|
|
1629
1430
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
1630
1431
|
(__v2df)_mm_add_pd(__A, __B),
|
|
1631
1432
|
(__v2df)_mm_setzero_pd());
|
|
1632
1433
|
}
|
|
1633
1434
|
|
|
1634
|
-
static __inline__ __m256d
|
|
1435
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1635
1436
|
_mm256_mask_add_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
|
|
1636
1437
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
1637
1438
|
(__v4df)_mm256_add_pd(__A, __B),
|
|
1638
1439
|
(__v4df)__W);
|
|
1639
1440
|
}
|
|
1640
1441
|
|
|
1641
|
-
static __inline__ __m256d
|
|
1442
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1642
1443
|
_mm256_maskz_add_pd(__mmask8 __U, __m256d __A, __m256d __B) {
|
|
1643
1444
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
1644
1445
|
(__v4df)_mm256_add_pd(__A, __B),
|
|
1645
1446
|
(__v4df)_mm256_setzero_pd());
|
|
1646
1447
|
}
|
|
1647
1448
|
|
|
1648
|
-
static __inline__ __m128
|
|
1449
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1649
1450
|
_mm_mask_add_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
|
|
1650
1451
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
1651
1452
|
(__v4sf)_mm_add_ps(__A, __B),
|
|
1652
1453
|
(__v4sf)__W);
|
|
1653
1454
|
}
|
|
1654
1455
|
|
|
1655
|
-
static __inline__ __m128
|
|
1456
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1656
1457
|
_mm_maskz_add_ps(__mmask8 __U, __m128 __A, __m128 __B) {
|
|
1657
1458
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
1658
1459
|
(__v4sf)_mm_add_ps(__A, __B),
|
|
1659
1460
|
(__v4sf)_mm_setzero_ps());
|
|
1660
1461
|
}
|
|
1661
1462
|
|
|
1662
|
-
static __inline__ __m256
|
|
1463
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1663
1464
|
_mm256_mask_add_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
|
|
1664
1465
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
1665
1466
|
(__v8sf)_mm256_add_ps(__A, __B),
|
|
1666
1467
|
(__v8sf)__W);
|
|
1667
1468
|
}
|
|
1668
1469
|
|
|
1669
|
-
static __inline__ __m256
|
|
1470
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1670
1471
|
_mm256_maskz_add_ps(__mmask8 __U, __m256 __A, __m256 __B) {
|
|
1671
1472
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
1672
1473
|
(__v8sf)_mm256_add_ps(__A, __B),
|
|
1673
1474
|
(__v8sf)_mm256_setzero_ps());
|
|
1674
1475
|
}
|
|
1675
1476
|
|
|
1676
|
-
static __inline__ __m128i
|
|
1677
|
-
_mm_mask_blend_epi32
|
|
1477
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1478
|
+
_mm_mask_blend_epi32(__mmask8 __U, __m128i __A, __m128i __W) {
|
|
1678
1479
|
return (__m128i) __builtin_ia32_selectd_128 ((__mmask8) __U,
|
|
1679
1480
|
(__v4si) __W,
|
|
1680
1481
|
(__v4si) __A);
|
|
1681
1482
|
}
|
|
1682
1483
|
|
|
1683
|
-
static __inline__ __m256i
|
|
1684
|
-
_mm256_mask_blend_epi32
|
|
1484
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1485
|
+
_mm256_mask_blend_epi32(__mmask8 __U, __m256i __A, __m256i __W) {
|
|
1685
1486
|
return (__m256i) __builtin_ia32_selectd_256 ((__mmask8) __U,
|
|
1686
1487
|
(__v8si) __W,
|
|
1687
1488
|
(__v8si) __A);
|
|
1688
1489
|
}
|
|
1689
1490
|
|
|
1690
|
-
static __inline__ __m128d
|
|
1691
|
-
_mm_mask_blend_pd
|
|
1491
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1492
|
+
_mm_mask_blend_pd(__mmask8 __U, __m128d __A, __m128d __W) {
|
|
1692
1493
|
return (__m128d) __builtin_ia32_selectpd_128 ((__mmask8) __U,
|
|
1693
1494
|
(__v2df) __W,
|
|
1694
1495
|
(__v2df) __A);
|
|
1695
1496
|
}
|
|
1696
1497
|
|
|
1697
|
-
static __inline__ __m256d
|
|
1698
|
-
_mm256_mask_blend_pd
|
|
1498
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1499
|
+
_mm256_mask_blend_pd(__mmask8 __U, __m256d __A, __m256d __W) {
|
|
1699
1500
|
return (__m256d) __builtin_ia32_selectpd_256 ((__mmask8) __U,
|
|
1700
1501
|
(__v4df) __W,
|
|
1701
1502
|
(__v4df) __A);
|
|
1702
1503
|
}
|
|
1703
1504
|
|
|
1704
|
-
static __inline__ __m128
|
|
1705
|
-
_mm_mask_blend_ps
|
|
1505
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1506
|
+
_mm_mask_blend_ps(__mmask8 __U, __m128 __A, __m128 __W) {
|
|
1706
1507
|
return (__m128) __builtin_ia32_selectps_128 ((__mmask8) __U,
|
|
1707
1508
|
(__v4sf) __W,
|
|
1708
1509
|
(__v4sf) __A);
|
|
1709
1510
|
}
|
|
1710
1511
|
|
|
1711
|
-
static __inline__ __m256
|
|
1712
|
-
_mm256_mask_blend_ps
|
|
1512
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1513
|
+
_mm256_mask_blend_ps(__mmask8 __U, __m256 __A, __m256 __W) {
|
|
1713
1514
|
return (__m256) __builtin_ia32_selectps_256 ((__mmask8) __U,
|
|
1714
1515
|
(__v8sf) __W,
|
|
1715
1516
|
(__v8sf) __A);
|
|
1716
1517
|
}
|
|
1717
1518
|
|
|
1718
|
-
static __inline__ __m128i
|
|
1719
|
-
_mm_mask_blend_epi64
|
|
1519
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1520
|
+
_mm_mask_blend_epi64(__mmask8 __U, __m128i __A, __m128i __W) {
|
|
1720
1521
|
return (__m128i) __builtin_ia32_selectq_128 ((__mmask8) __U,
|
|
1721
1522
|
(__v2di) __W,
|
|
1722
1523
|
(__v2di) __A);
|
|
1723
1524
|
}
|
|
1724
1525
|
|
|
1725
|
-
static __inline__ __m256i
|
|
1726
|
-
_mm256_mask_blend_epi64
|
|
1526
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1527
|
+
_mm256_mask_blend_epi64(__mmask8 __U, __m256i __A, __m256i __W) {
|
|
1727
1528
|
return (__m256i) __builtin_ia32_selectq_256 ((__mmask8) __U,
|
|
1728
1529
|
(__v4di) __W,
|
|
1729
1530
|
(__v4di) __A);
|
|
@@ -1905,57 +1706,57 @@ _mm256_mask_compressstoreu_epi32 (void *__P, __mmask8 __U, __m256i __A) {
|
|
|
1905
1706
|
(__mmask8) __U);
|
|
1906
1707
|
}
|
|
1907
1708
|
|
|
1908
|
-
static __inline__ __m128d
|
|
1909
|
-
_mm_mask_cvtepi32_pd
|
|
1709
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1710
|
+
_mm_mask_cvtepi32_pd(__m128d __W, __mmask8 __U, __m128i __A) {
|
|
1910
1711
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8) __U,
|
|
1911
1712
|
(__v2df)_mm_cvtepi32_pd(__A),
|
|
1912
1713
|
(__v2df)__W);
|
|
1913
1714
|
}
|
|
1914
1715
|
|
|
1915
|
-
static __inline__ __m128d
|
|
1916
|
-
_mm_maskz_cvtepi32_pd
|
|
1716
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1717
|
+
_mm_maskz_cvtepi32_pd(__mmask8 __U, __m128i __A) {
|
|
1917
1718
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8) __U,
|
|
1918
1719
|
(__v2df)_mm_cvtepi32_pd(__A),
|
|
1919
1720
|
(__v2df)_mm_setzero_pd());
|
|
1920
1721
|
}
|
|
1921
1722
|
|
|
1922
|
-
static __inline__ __m256d
|
|
1923
|
-
_mm256_mask_cvtepi32_pd
|
|
1723
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1724
|
+
_mm256_mask_cvtepi32_pd(__m256d __W, __mmask8 __U, __m128i __A) {
|
|
1924
1725
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8) __U,
|
|
1925
1726
|
(__v4df)_mm256_cvtepi32_pd(__A),
|
|
1926
1727
|
(__v4df)__W);
|
|
1927
1728
|
}
|
|
1928
1729
|
|
|
1929
|
-
static __inline__ __m256d
|
|
1930
|
-
_mm256_maskz_cvtepi32_pd
|
|
1730
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1731
|
+
_mm256_maskz_cvtepi32_pd(__mmask8 __U, __m128i __A) {
|
|
1931
1732
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8) __U,
|
|
1932
1733
|
(__v4df)_mm256_cvtepi32_pd(__A),
|
|
1933
1734
|
(__v4df)_mm256_setzero_pd());
|
|
1934
1735
|
}
|
|
1935
1736
|
|
|
1936
|
-
static __inline__ __m128
|
|
1937
|
-
_mm_mask_cvtepi32_ps
|
|
1737
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1738
|
+
_mm_mask_cvtepi32_ps(__m128 __W, __mmask8 __U, __m128i __A) {
|
|
1938
1739
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
1939
1740
|
(__v4sf)_mm_cvtepi32_ps(__A),
|
|
1940
1741
|
(__v4sf)__W);
|
|
1941
1742
|
}
|
|
1942
1743
|
|
|
1943
|
-
static __inline__ __m128
|
|
1944
|
-
_mm_maskz_cvtepi32_ps
|
|
1744
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1745
|
+
_mm_maskz_cvtepi32_ps(__mmask8 __U, __m128i __A) {
|
|
1945
1746
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
1946
1747
|
(__v4sf)_mm_cvtepi32_ps(__A),
|
|
1947
1748
|
(__v4sf)_mm_setzero_ps());
|
|
1948
1749
|
}
|
|
1949
1750
|
|
|
1950
|
-
static __inline__ __m256
|
|
1951
|
-
_mm256_mask_cvtepi32_ps
|
|
1751
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1752
|
+
_mm256_mask_cvtepi32_ps(__m256 __W, __mmask8 __U, __m256i __A) {
|
|
1952
1753
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
1953
1754
|
(__v8sf)_mm256_cvtepi32_ps(__A),
|
|
1954
1755
|
(__v8sf)__W);
|
|
1955
1756
|
}
|
|
1956
1757
|
|
|
1957
|
-
static __inline__ __m256
|
|
1958
|
-
_mm256_maskz_cvtepi32_ps
|
|
1758
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1759
|
+
_mm256_maskz_cvtepi32_ps(__mmask8 __U, __m256i __A) {
|
|
1959
1760
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
1960
1761
|
(__v8sf)_mm256_cvtepi32_ps(__A),
|
|
1961
1762
|
(__v8sf)_mm256_setzero_ps());
|
|
@@ -1990,30 +1791,30 @@ _mm256_maskz_cvtpd_epi32 (__mmask8 __U, __m256d __A) {
|
|
|
1990
1791
|
(__v4si)_mm_setzero_si128());
|
|
1991
1792
|
}
|
|
1992
1793
|
|
|
1993
|
-
static __inline__ __m128
|
|
1994
|
-
_mm_mask_cvtpd_ps
|
|
1794
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1795
|
+
_mm_mask_cvtpd_ps(__m128 __W, __mmask8 __U, __m128d __A) {
|
|
1995
1796
|
return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A,
|
|
1996
1797
|
(__v4sf) __W,
|
|
1997
1798
|
(__mmask8) __U);
|
|
1998
1799
|
}
|
|
1999
1800
|
|
|
2000
|
-
static __inline__ __m128
|
|
2001
|
-
_mm_maskz_cvtpd_ps
|
|
1801
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
1802
|
+
_mm_maskz_cvtpd_ps(__mmask8 __U, __m128d __A) {
|
|
2002
1803
|
return (__m128) __builtin_ia32_cvtpd2ps_mask ((__v2df) __A,
|
|
2003
1804
|
(__v4sf)
|
|
2004
1805
|
_mm_setzero_ps (),
|
|
2005
1806
|
(__mmask8) __U);
|
|
2006
1807
|
}
|
|
2007
1808
|
|
|
2008
|
-
static __inline__ __m128
|
|
2009
|
-
_mm256_mask_cvtpd_ps
|
|
1809
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1810
|
+
_mm256_mask_cvtpd_ps(__m128 __W, __mmask8 __U, __m256d __A) {
|
|
2010
1811
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2011
1812
|
(__v4sf)_mm256_cvtpd_ps(__A),
|
|
2012
1813
|
(__v4sf)__W);
|
|
2013
1814
|
}
|
|
2014
1815
|
|
|
2015
|
-
static __inline__ __m128
|
|
2016
|
-
_mm256_maskz_cvtpd_ps
|
|
1816
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
1817
|
+
_mm256_maskz_cvtpd_ps(__mmask8 __U, __m256d __A) {
|
|
2017
1818
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2018
1819
|
(__v4sf)_mm256_cvtpd_ps(__A),
|
|
2019
1820
|
(__v4sf)_mm_setzero_ps());
|
|
@@ -2316,133 +2117,133 @@ _mm256_maskz_cvttps_epu32 (__mmask8 __U, __m256 __A) {
|
|
|
2316
2117
|
(__mmask8) __U);
|
|
2317
2118
|
}
|
|
2318
2119
|
|
|
2319
|
-
static __inline__ __m128d
|
|
2320
|
-
_mm_cvtepu32_pd
|
|
2120
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2121
|
+
_mm_cvtepu32_pd(__m128i __A) {
|
|
2321
2122
|
return (__m128d) __builtin_convertvector(
|
|
2322
2123
|
__builtin_shufflevector((__v4su)__A, (__v4su)__A, 0, 1), __v2df);
|
|
2323
2124
|
}
|
|
2324
2125
|
|
|
2325
|
-
static __inline__ __m128d
|
|
2326
|
-
_mm_mask_cvtepu32_pd
|
|
2126
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2127
|
+
_mm_mask_cvtepu32_pd(__m128d __W, __mmask8 __U, __m128i __A) {
|
|
2327
2128
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8) __U,
|
|
2328
2129
|
(__v2df)_mm_cvtepu32_pd(__A),
|
|
2329
2130
|
(__v2df)__W);
|
|
2330
2131
|
}
|
|
2331
2132
|
|
|
2332
|
-
static __inline__ __m128d
|
|
2333
|
-
_mm_maskz_cvtepu32_pd
|
|
2133
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2134
|
+
_mm_maskz_cvtepu32_pd(__mmask8 __U, __m128i __A) {
|
|
2334
2135
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8) __U,
|
|
2335
2136
|
(__v2df)_mm_cvtepu32_pd(__A),
|
|
2336
2137
|
(__v2df)_mm_setzero_pd());
|
|
2337
2138
|
}
|
|
2338
2139
|
|
|
2339
|
-
static __inline__ __m256d
|
|
2340
|
-
_mm256_cvtepu32_pd
|
|
2140
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2141
|
+
_mm256_cvtepu32_pd(__m128i __A) {
|
|
2341
2142
|
return (__m256d)__builtin_convertvector((__v4su)__A, __v4df);
|
|
2342
2143
|
}
|
|
2343
2144
|
|
|
2344
|
-
static __inline__ __m256d
|
|
2345
|
-
_mm256_mask_cvtepu32_pd
|
|
2145
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2146
|
+
_mm256_mask_cvtepu32_pd(__m256d __W, __mmask8 __U, __m128i __A) {
|
|
2346
2147
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8) __U,
|
|
2347
2148
|
(__v4df)_mm256_cvtepu32_pd(__A),
|
|
2348
2149
|
(__v4df)__W);
|
|
2349
2150
|
}
|
|
2350
2151
|
|
|
2351
|
-
static __inline__ __m256d
|
|
2352
|
-
_mm256_maskz_cvtepu32_pd
|
|
2152
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2153
|
+
_mm256_maskz_cvtepu32_pd(__mmask8 __U, __m128i __A) {
|
|
2353
2154
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8) __U,
|
|
2354
2155
|
(__v4df)_mm256_cvtepu32_pd(__A),
|
|
2355
2156
|
(__v4df)_mm256_setzero_pd());
|
|
2356
2157
|
}
|
|
2357
2158
|
|
|
2358
|
-
static __inline__ __m128
|
|
2359
|
-
_mm_cvtepu32_ps
|
|
2159
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2160
|
+
_mm_cvtepu32_ps(__m128i __A) {
|
|
2360
2161
|
return (__m128)__builtin_convertvector((__v4su)__A, __v4sf);
|
|
2361
2162
|
}
|
|
2362
2163
|
|
|
2363
|
-
static __inline__ __m128
|
|
2364
|
-
_mm_mask_cvtepu32_ps
|
|
2164
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2165
|
+
_mm_mask_cvtepu32_ps(__m128 __W, __mmask8 __U, __m128i __A) {
|
|
2365
2166
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2366
2167
|
(__v4sf)_mm_cvtepu32_ps(__A),
|
|
2367
2168
|
(__v4sf)__W);
|
|
2368
2169
|
}
|
|
2369
2170
|
|
|
2370
|
-
static __inline__ __m128
|
|
2371
|
-
_mm_maskz_cvtepu32_ps
|
|
2171
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2172
|
+
_mm_maskz_cvtepu32_ps(__mmask8 __U, __m128i __A) {
|
|
2372
2173
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2373
2174
|
(__v4sf)_mm_cvtepu32_ps(__A),
|
|
2374
2175
|
(__v4sf)_mm_setzero_ps());
|
|
2375
2176
|
}
|
|
2376
2177
|
|
|
2377
|
-
static __inline__ __m256
|
|
2378
|
-
_mm256_cvtepu32_ps
|
|
2178
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2179
|
+
_mm256_cvtepu32_ps(__m256i __A) {
|
|
2379
2180
|
return (__m256)__builtin_convertvector((__v8su)__A, __v8sf);
|
|
2380
2181
|
}
|
|
2381
2182
|
|
|
2382
|
-
static __inline__ __m256
|
|
2383
|
-
_mm256_mask_cvtepu32_ps
|
|
2183
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2184
|
+
_mm256_mask_cvtepu32_ps(__m256 __W, __mmask8 __U, __m256i __A) {
|
|
2384
2185
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
2385
2186
|
(__v8sf)_mm256_cvtepu32_ps(__A),
|
|
2386
2187
|
(__v8sf)__W);
|
|
2387
2188
|
}
|
|
2388
2189
|
|
|
2389
|
-
static __inline__ __m256
|
|
2390
|
-
_mm256_maskz_cvtepu32_ps
|
|
2190
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2191
|
+
_mm256_maskz_cvtepu32_ps(__mmask8 __U, __m256i __A) {
|
|
2391
2192
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
2392
2193
|
(__v8sf)_mm256_cvtepu32_ps(__A),
|
|
2393
2194
|
(__v8sf)_mm256_setzero_ps());
|
|
2394
2195
|
}
|
|
2395
2196
|
|
|
2396
|
-
static __inline__ __m128d
|
|
2197
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2397
2198
|
_mm_mask_div_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
|
|
2398
2199
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
2399
2200
|
(__v2df)_mm_div_pd(__A, __B),
|
|
2400
2201
|
(__v2df)__W);
|
|
2401
2202
|
}
|
|
2402
2203
|
|
|
2403
|
-
static __inline__ __m128d
|
|
2204
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2404
2205
|
_mm_maskz_div_pd(__mmask8 __U, __m128d __A, __m128d __B) {
|
|
2405
2206
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
2406
2207
|
(__v2df)_mm_div_pd(__A, __B),
|
|
2407
2208
|
(__v2df)_mm_setzero_pd());
|
|
2408
2209
|
}
|
|
2409
2210
|
|
|
2410
|
-
static __inline__ __m256d
|
|
2211
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2411
2212
|
_mm256_mask_div_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
|
|
2412
2213
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
2413
2214
|
(__v4df)_mm256_div_pd(__A, __B),
|
|
2414
2215
|
(__v4df)__W);
|
|
2415
2216
|
}
|
|
2416
2217
|
|
|
2417
|
-
static __inline__ __m256d
|
|
2218
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2418
2219
|
_mm256_maskz_div_pd(__mmask8 __U, __m256d __A, __m256d __B) {
|
|
2419
2220
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
2420
2221
|
(__v4df)_mm256_div_pd(__A, __B),
|
|
2421
2222
|
(__v4df)_mm256_setzero_pd());
|
|
2422
2223
|
}
|
|
2423
2224
|
|
|
2424
|
-
static __inline__ __m128
|
|
2225
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2425
2226
|
_mm_mask_div_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
|
|
2426
2227
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2427
2228
|
(__v4sf)_mm_div_ps(__A, __B),
|
|
2428
2229
|
(__v4sf)__W);
|
|
2429
2230
|
}
|
|
2430
2231
|
|
|
2431
|
-
static __inline__ __m128
|
|
2232
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2432
2233
|
_mm_maskz_div_ps(__mmask8 __U, __m128 __A, __m128 __B) {
|
|
2433
2234
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2434
2235
|
(__v4sf)_mm_div_ps(__A, __B),
|
|
2435
2236
|
(__v4sf)_mm_setzero_ps());
|
|
2436
2237
|
}
|
|
2437
2238
|
|
|
2438
|
-
static __inline__ __m256
|
|
2239
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2439
2240
|
_mm256_mask_div_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
|
|
2440
2241
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
2441
2242
|
(__v8sf)_mm256_div_ps(__A, __B),
|
|
2442
2243
|
(__v8sf)__W);
|
|
2443
2244
|
}
|
|
2444
2245
|
|
|
2445
|
-
static __inline__ __m256
|
|
2246
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2446
2247
|
_mm256_maskz_div_ps(__mmask8 __U, __m256 __A, __m256 __B) {
|
|
2447
2248
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
2448
2249
|
(__v8sf)_mm256_div_ps(__A, __B),
|
|
@@ -2796,499 +2597,499 @@ _mm256_maskz_getexp_ps (__mmask8 __U, __m256 __A) {
|
|
|
2796
2597
|
(__mmask8) __U);
|
|
2797
2598
|
}
|
|
2798
2599
|
|
|
2799
|
-
static __inline__ __m128d
|
|
2600
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2800
2601
|
_mm_mask_max_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
|
|
2801
2602
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
2802
2603
|
(__v2df)_mm_max_pd(__A, __B),
|
|
2803
2604
|
(__v2df)__W);
|
|
2804
2605
|
}
|
|
2805
2606
|
|
|
2806
|
-
static __inline__ __m128d
|
|
2607
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2807
2608
|
_mm_maskz_max_pd(__mmask8 __U, __m128d __A, __m128d __B) {
|
|
2808
2609
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
2809
2610
|
(__v2df)_mm_max_pd(__A, __B),
|
|
2810
2611
|
(__v2df)_mm_setzero_pd());
|
|
2811
2612
|
}
|
|
2812
2613
|
|
|
2813
|
-
static __inline__ __m256d
|
|
2614
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2814
2615
|
_mm256_mask_max_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
|
|
2815
2616
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
2816
2617
|
(__v4df)_mm256_max_pd(__A, __B),
|
|
2817
2618
|
(__v4df)__W);
|
|
2818
2619
|
}
|
|
2819
2620
|
|
|
2820
|
-
static __inline__ __m256d
|
|
2621
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2821
2622
|
_mm256_maskz_max_pd(__mmask8 __U, __m256d __A, __m256d __B) {
|
|
2822
2623
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
2823
2624
|
(__v4df)_mm256_max_pd(__A, __B),
|
|
2824
2625
|
(__v4df)_mm256_setzero_pd());
|
|
2825
2626
|
}
|
|
2826
2627
|
|
|
2827
|
-
static __inline__ __m128
|
|
2628
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2828
2629
|
_mm_mask_max_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
|
|
2829
2630
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2830
2631
|
(__v4sf)_mm_max_ps(__A, __B),
|
|
2831
2632
|
(__v4sf)__W);
|
|
2832
2633
|
}
|
|
2833
2634
|
|
|
2834
|
-
static __inline__ __m128
|
|
2635
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2835
2636
|
_mm_maskz_max_ps(__mmask8 __U, __m128 __A, __m128 __B) {
|
|
2836
2637
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2837
2638
|
(__v4sf)_mm_max_ps(__A, __B),
|
|
2838
2639
|
(__v4sf)_mm_setzero_ps());
|
|
2839
2640
|
}
|
|
2840
2641
|
|
|
2841
|
-
static __inline__ __m256
|
|
2642
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2842
2643
|
_mm256_mask_max_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
|
|
2843
2644
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
2844
2645
|
(__v8sf)_mm256_max_ps(__A, __B),
|
|
2845
2646
|
(__v8sf)__W);
|
|
2846
2647
|
}
|
|
2847
2648
|
|
|
2848
|
-
static __inline__ __m256
|
|
2649
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2849
2650
|
_mm256_maskz_max_ps(__mmask8 __U, __m256 __A, __m256 __B) {
|
|
2850
2651
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
2851
2652
|
(__v8sf)_mm256_max_ps(__A, __B),
|
|
2852
2653
|
(__v8sf)_mm256_setzero_ps());
|
|
2853
2654
|
}
|
|
2854
2655
|
|
|
2855
|
-
static __inline__ __m128d
|
|
2656
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2856
2657
|
_mm_mask_min_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
|
|
2857
2658
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
2858
2659
|
(__v2df)_mm_min_pd(__A, __B),
|
|
2859
2660
|
(__v2df)__W);
|
|
2860
2661
|
}
|
|
2861
2662
|
|
|
2862
|
-
static __inline__ __m128d
|
|
2663
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2863
2664
|
_mm_maskz_min_pd(__mmask8 __U, __m128d __A, __m128d __B) {
|
|
2864
2665
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
2865
2666
|
(__v2df)_mm_min_pd(__A, __B),
|
|
2866
2667
|
(__v2df)_mm_setzero_pd());
|
|
2867
2668
|
}
|
|
2868
2669
|
|
|
2869
|
-
static __inline__ __m256d
|
|
2670
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2870
2671
|
_mm256_mask_min_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
|
|
2871
2672
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
2872
2673
|
(__v4df)_mm256_min_pd(__A, __B),
|
|
2873
2674
|
(__v4df)__W);
|
|
2874
2675
|
}
|
|
2875
2676
|
|
|
2876
|
-
static __inline__ __m256d
|
|
2677
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2877
2678
|
_mm256_maskz_min_pd(__mmask8 __U, __m256d __A, __m256d __B) {
|
|
2878
2679
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
2879
2680
|
(__v4df)_mm256_min_pd(__A, __B),
|
|
2880
2681
|
(__v4df)_mm256_setzero_pd());
|
|
2881
2682
|
}
|
|
2882
2683
|
|
|
2883
|
-
static __inline__ __m128
|
|
2684
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2884
2685
|
_mm_mask_min_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
|
|
2885
2686
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2886
2687
|
(__v4sf)_mm_min_ps(__A, __B),
|
|
2887
2688
|
(__v4sf)__W);
|
|
2888
2689
|
}
|
|
2889
2690
|
|
|
2890
|
-
static __inline__ __m128
|
|
2691
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2891
2692
|
_mm_maskz_min_ps(__mmask8 __U, __m128 __A, __m128 __B) {
|
|
2892
2693
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2893
2694
|
(__v4sf)_mm_min_ps(__A, __B),
|
|
2894
2695
|
(__v4sf)_mm_setzero_ps());
|
|
2895
2696
|
}
|
|
2896
2697
|
|
|
2897
|
-
static __inline__ __m256
|
|
2698
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2898
2699
|
_mm256_mask_min_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
|
|
2899
2700
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
2900
2701
|
(__v8sf)_mm256_min_ps(__A, __B),
|
|
2901
2702
|
(__v8sf)__W);
|
|
2902
2703
|
}
|
|
2903
2704
|
|
|
2904
|
-
static __inline__ __m256
|
|
2705
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2905
2706
|
_mm256_maskz_min_ps(__mmask8 __U, __m256 __A, __m256 __B) {
|
|
2906
2707
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
2907
2708
|
(__v8sf)_mm256_min_ps(__A, __B),
|
|
2908
2709
|
(__v8sf)_mm256_setzero_ps());
|
|
2909
2710
|
}
|
|
2910
2711
|
|
|
2911
|
-
static __inline__ __m128d
|
|
2712
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2912
2713
|
_mm_mask_mul_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
|
|
2913
2714
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
2914
2715
|
(__v2df)_mm_mul_pd(__A, __B),
|
|
2915
2716
|
(__v2df)__W);
|
|
2916
2717
|
}
|
|
2917
2718
|
|
|
2918
|
-
static __inline__ __m128d
|
|
2719
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2919
2720
|
_mm_maskz_mul_pd(__mmask8 __U, __m128d __A, __m128d __B) {
|
|
2920
2721
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
2921
2722
|
(__v2df)_mm_mul_pd(__A, __B),
|
|
2922
2723
|
(__v2df)_mm_setzero_pd());
|
|
2923
2724
|
}
|
|
2924
2725
|
|
|
2925
|
-
static __inline__ __m256d
|
|
2726
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2926
2727
|
_mm256_mask_mul_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
|
|
2927
2728
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
2928
2729
|
(__v4df)_mm256_mul_pd(__A, __B),
|
|
2929
2730
|
(__v4df)__W);
|
|
2930
2731
|
}
|
|
2931
2732
|
|
|
2932
|
-
static __inline__ __m256d
|
|
2733
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2933
2734
|
_mm256_maskz_mul_pd(__mmask8 __U, __m256d __A, __m256d __B) {
|
|
2934
2735
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
2935
2736
|
(__v4df)_mm256_mul_pd(__A, __B),
|
|
2936
2737
|
(__v4df)_mm256_setzero_pd());
|
|
2937
2738
|
}
|
|
2938
2739
|
|
|
2939
|
-
static __inline__ __m128
|
|
2740
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2940
2741
|
_mm_mask_mul_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
|
|
2941
2742
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2942
2743
|
(__v4sf)_mm_mul_ps(__A, __B),
|
|
2943
2744
|
(__v4sf)__W);
|
|
2944
2745
|
}
|
|
2945
2746
|
|
|
2946
|
-
static __inline__ __m128
|
|
2747
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2947
2748
|
_mm_maskz_mul_ps(__mmask8 __U, __m128 __A, __m128 __B) {
|
|
2948
2749
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
2949
2750
|
(__v4sf)_mm_mul_ps(__A, __B),
|
|
2950
2751
|
(__v4sf)_mm_setzero_ps());
|
|
2951
2752
|
}
|
|
2952
2753
|
|
|
2953
|
-
static __inline__ __m256
|
|
2754
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2954
2755
|
_mm256_mask_mul_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
|
|
2955
2756
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
2956
2757
|
(__v8sf)_mm256_mul_ps(__A, __B),
|
|
2957
2758
|
(__v8sf)__W);
|
|
2958
2759
|
}
|
|
2959
2760
|
|
|
2960
|
-
static __inline__ __m256
|
|
2761
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2961
2762
|
_mm256_maskz_mul_ps(__mmask8 __U, __m256 __A, __m256 __B) {
|
|
2962
2763
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
2963
2764
|
(__v8sf)_mm256_mul_ps(__A, __B),
|
|
2964
2765
|
(__v8sf)_mm256_setzero_ps());
|
|
2965
2766
|
}
|
|
2966
2767
|
|
|
2967
|
-
static __inline__ __m128i
|
|
2768
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2968
2769
|
_mm_mask_abs_epi32(__m128i __W, __mmask8 __U, __m128i __A) {
|
|
2969
2770
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
2970
2771
|
(__v4si)_mm_abs_epi32(__A),
|
|
2971
2772
|
(__v4si)__W);
|
|
2972
2773
|
}
|
|
2973
2774
|
|
|
2974
|
-
static __inline__ __m128i
|
|
2775
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2975
2776
|
_mm_maskz_abs_epi32(__mmask8 __U, __m128i __A) {
|
|
2976
2777
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
2977
2778
|
(__v4si)_mm_abs_epi32(__A),
|
|
2978
2779
|
(__v4si)_mm_setzero_si128());
|
|
2979
2780
|
}
|
|
2980
2781
|
|
|
2981
|
-
static __inline__ __m256i
|
|
2782
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2982
2783
|
_mm256_mask_abs_epi32(__m256i __W, __mmask8 __U, __m256i __A) {
|
|
2983
2784
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
2984
2785
|
(__v8si)_mm256_abs_epi32(__A),
|
|
2985
2786
|
(__v8si)__W);
|
|
2986
2787
|
}
|
|
2987
2788
|
|
|
2988
|
-
static __inline__ __m256i
|
|
2789
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2989
2790
|
_mm256_maskz_abs_epi32(__mmask8 __U, __m256i __A) {
|
|
2990
2791
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
2991
2792
|
(__v8si)_mm256_abs_epi32(__A),
|
|
2992
2793
|
(__v8si)_mm256_setzero_si256());
|
|
2993
2794
|
}
|
|
2994
2795
|
|
|
2995
|
-
static __inline__ __m128i
|
|
2996
|
-
_mm_abs_epi64
|
|
2796
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2797
|
+
_mm_abs_epi64(__m128i __A) {
|
|
2997
2798
|
return (__m128i)__builtin_elementwise_abs((__v2di)__A);
|
|
2998
2799
|
}
|
|
2999
2800
|
|
|
3000
|
-
static __inline__ __m128i
|
|
3001
|
-
_mm_mask_abs_epi64
|
|
2801
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2802
|
+
_mm_mask_abs_epi64(__m128i __W, __mmask8 __U, __m128i __A) {
|
|
3002
2803
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
3003
2804
|
(__v2di)_mm_abs_epi64(__A),
|
|
3004
2805
|
(__v2di)__W);
|
|
3005
2806
|
}
|
|
3006
2807
|
|
|
3007
|
-
static __inline__ __m128i
|
|
3008
|
-
_mm_maskz_abs_epi64
|
|
2808
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2809
|
+
_mm_maskz_abs_epi64(__mmask8 __U, __m128i __A) {
|
|
3009
2810
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
3010
2811
|
(__v2di)_mm_abs_epi64(__A),
|
|
3011
2812
|
(__v2di)_mm_setzero_si128());
|
|
3012
2813
|
}
|
|
3013
2814
|
|
|
3014
|
-
static __inline__ __m256i
|
|
3015
|
-
_mm256_abs_epi64
|
|
2815
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2816
|
+
_mm256_abs_epi64(__m256i __A) {
|
|
3016
2817
|
return (__m256i)__builtin_elementwise_abs((__v4di)__A);
|
|
3017
2818
|
}
|
|
3018
2819
|
|
|
3019
|
-
static __inline__ __m256i
|
|
3020
|
-
_mm256_mask_abs_epi64
|
|
2820
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2821
|
+
_mm256_mask_abs_epi64(__m256i __W, __mmask8 __U, __m256i __A) {
|
|
3021
2822
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
3022
2823
|
(__v4di)_mm256_abs_epi64(__A),
|
|
3023
2824
|
(__v4di)__W);
|
|
3024
2825
|
}
|
|
3025
2826
|
|
|
3026
|
-
static __inline__ __m256i
|
|
3027
|
-
_mm256_maskz_abs_epi64
|
|
2827
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2828
|
+
_mm256_maskz_abs_epi64(__mmask8 __U, __m256i __A) {
|
|
3028
2829
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
3029
2830
|
(__v4di)_mm256_abs_epi64(__A),
|
|
3030
2831
|
(__v4di)_mm256_setzero_si256());
|
|
3031
2832
|
}
|
|
3032
2833
|
|
|
3033
|
-
static __inline__ __m128i
|
|
2834
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3034
2835
|
_mm_maskz_max_epi32(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
3035
2836
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
|
|
3036
2837
|
(__v4si)_mm_max_epi32(__A, __B),
|
|
3037
2838
|
(__v4si)_mm_setzero_si128());
|
|
3038
2839
|
}
|
|
3039
2840
|
|
|
3040
|
-
static __inline__ __m128i
|
|
2841
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3041
2842
|
_mm_mask_max_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
3042
2843
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
|
|
3043
2844
|
(__v4si)_mm_max_epi32(__A, __B),
|
|
3044
2845
|
(__v4si)__W);
|
|
3045
2846
|
}
|
|
3046
2847
|
|
|
3047
|
-
static __inline__ __m256i
|
|
2848
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3048
2849
|
_mm256_maskz_max_epi32(__mmask8 __M, __m256i __A, __m256i __B) {
|
|
3049
2850
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
3050
2851
|
(__v8si)_mm256_max_epi32(__A, __B),
|
|
3051
2852
|
(__v8si)_mm256_setzero_si256());
|
|
3052
2853
|
}
|
|
3053
2854
|
|
|
3054
|
-
static __inline__ __m256i
|
|
2855
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3055
2856
|
_mm256_mask_max_epi32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) {
|
|
3056
2857
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
3057
2858
|
(__v8si)_mm256_max_epi32(__A, __B),
|
|
3058
2859
|
(__v8si)__W);
|
|
3059
2860
|
}
|
|
3060
2861
|
|
|
3061
|
-
static __inline__ __m128i
|
|
3062
|
-
_mm_max_epi64
|
|
2862
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2863
|
+
_mm_max_epi64(__m128i __A, __m128i __B) {
|
|
3063
2864
|
return (__m128i)__builtin_elementwise_max((__v2di)__A, (__v2di)__B);
|
|
3064
2865
|
}
|
|
3065
2866
|
|
|
3066
|
-
static __inline__ __m128i
|
|
3067
|
-
_mm_maskz_max_epi64
|
|
2867
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2868
|
+
_mm_maskz_max_epi64(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
3068
2869
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
3069
2870
|
(__v2di)_mm_max_epi64(__A, __B),
|
|
3070
2871
|
(__v2di)_mm_setzero_si128());
|
|
3071
2872
|
}
|
|
3072
2873
|
|
|
3073
|
-
static __inline__ __m128i
|
|
3074
|
-
_mm_mask_max_epi64
|
|
2874
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2875
|
+
_mm_mask_max_epi64(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
3075
2876
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
3076
2877
|
(__v2di)_mm_max_epi64(__A, __B),
|
|
3077
2878
|
(__v2di)__W);
|
|
3078
2879
|
}
|
|
3079
2880
|
|
|
3080
|
-
static __inline__ __m256i
|
|
3081
|
-
_mm256_max_epi64
|
|
2881
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2882
|
+
_mm256_max_epi64(__m256i __A, __m256i __B) {
|
|
3082
2883
|
return (__m256i)__builtin_elementwise_max((__v4di)__A, (__v4di)__B);
|
|
3083
2884
|
}
|
|
3084
2885
|
|
|
3085
|
-
static __inline__ __m256i
|
|
3086
|
-
_mm256_maskz_max_epi64
|
|
2886
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2887
|
+
_mm256_maskz_max_epi64(__mmask8 __M, __m256i __A, __m256i __B) {
|
|
3087
2888
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
3088
2889
|
(__v4di)_mm256_max_epi64(__A, __B),
|
|
3089
2890
|
(__v4di)_mm256_setzero_si256());
|
|
3090
2891
|
}
|
|
3091
2892
|
|
|
3092
|
-
static __inline__ __m256i
|
|
3093
|
-
_mm256_mask_max_epi64
|
|
2893
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2894
|
+
_mm256_mask_max_epi64(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) {
|
|
3094
2895
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
3095
2896
|
(__v4di)_mm256_max_epi64(__A, __B),
|
|
3096
2897
|
(__v4di)__W);
|
|
3097
2898
|
}
|
|
3098
2899
|
|
|
3099
|
-
static __inline__ __m128i
|
|
2900
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3100
2901
|
_mm_maskz_max_epu32(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
3101
2902
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
|
|
3102
2903
|
(__v4si)_mm_max_epu32(__A, __B),
|
|
3103
2904
|
(__v4si)_mm_setzero_si128());
|
|
3104
2905
|
}
|
|
3105
2906
|
|
|
3106
|
-
static __inline__ __m128i
|
|
2907
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3107
2908
|
_mm_mask_max_epu32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
3108
2909
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
|
|
3109
2910
|
(__v4si)_mm_max_epu32(__A, __B),
|
|
3110
2911
|
(__v4si)__W);
|
|
3111
2912
|
}
|
|
3112
2913
|
|
|
3113
|
-
static __inline__ __m256i
|
|
2914
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3114
2915
|
_mm256_maskz_max_epu32(__mmask8 __M, __m256i __A, __m256i __B) {
|
|
3115
2916
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
3116
2917
|
(__v8si)_mm256_max_epu32(__A, __B),
|
|
3117
2918
|
(__v8si)_mm256_setzero_si256());
|
|
3118
2919
|
}
|
|
3119
2920
|
|
|
3120
|
-
static __inline__ __m256i
|
|
2921
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3121
2922
|
_mm256_mask_max_epu32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) {
|
|
3122
2923
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
3123
2924
|
(__v8si)_mm256_max_epu32(__A, __B),
|
|
3124
2925
|
(__v8si)__W);
|
|
3125
2926
|
}
|
|
3126
2927
|
|
|
3127
|
-
static __inline__ __m128i
|
|
3128
|
-
_mm_max_epu64
|
|
2928
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2929
|
+
_mm_max_epu64(__m128i __A, __m128i __B) {
|
|
3129
2930
|
return (__m128i)__builtin_elementwise_max((__v2du)__A, (__v2du)__B);
|
|
3130
2931
|
}
|
|
3131
2932
|
|
|
3132
|
-
static __inline__ __m128i
|
|
3133
|
-
_mm_maskz_max_epu64
|
|
2933
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2934
|
+
_mm_maskz_max_epu64(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
3134
2935
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
3135
2936
|
(__v2di)_mm_max_epu64(__A, __B),
|
|
3136
2937
|
(__v2di)_mm_setzero_si128());
|
|
3137
2938
|
}
|
|
3138
2939
|
|
|
3139
|
-
static __inline__ __m128i
|
|
3140
|
-
_mm_mask_max_epu64
|
|
2940
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2941
|
+
_mm_mask_max_epu64(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
3141
2942
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
3142
2943
|
(__v2di)_mm_max_epu64(__A, __B),
|
|
3143
2944
|
(__v2di)__W);
|
|
3144
2945
|
}
|
|
3145
2946
|
|
|
3146
|
-
static __inline__ __m256i
|
|
3147
|
-
_mm256_max_epu64
|
|
2947
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2948
|
+
_mm256_max_epu64(__m256i __A, __m256i __B) {
|
|
3148
2949
|
return (__m256i)__builtin_elementwise_max((__v4du)__A, (__v4du)__B);
|
|
3149
2950
|
}
|
|
3150
2951
|
|
|
3151
|
-
static __inline__ __m256i
|
|
3152
|
-
_mm256_maskz_max_epu64
|
|
2952
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2953
|
+
_mm256_maskz_max_epu64(__mmask8 __M, __m256i __A, __m256i __B) {
|
|
3153
2954
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
3154
2955
|
(__v4di)_mm256_max_epu64(__A, __B),
|
|
3155
2956
|
(__v4di)_mm256_setzero_si256());
|
|
3156
2957
|
}
|
|
3157
2958
|
|
|
3158
|
-
static __inline__ __m256i
|
|
3159
|
-
_mm256_mask_max_epu64
|
|
2959
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
2960
|
+
_mm256_mask_max_epu64(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) {
|
|
3160
2961
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
3161
2962
|
(__v4di)_mm256_max_epu64(__A, __B),
|
|
3162
2963
|
(__v4di)__W);
|
|
3163
2964
|
}
|
|
3164
2965
|
|
|
3165
|
-
static __inline__ __m128i
|
|
2966
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3166
2967
|
_mm_maskz_min_epi32(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
3167
2968
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
|
|
3168
2969
|
(__v4si)_mm_min_epi32(__A, __B),
|
|
3169
2970
|
(__v4si)_mm_setzero_si128());
|
|
3170
2971
|
}
|
|
3171
2972
|
|
|
3172
|
-
static __inline__ __m128i
|
|
2973
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3173
2974
|
_mm_mask_min_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
3174
2975
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
|
|
3175
2976
|
(__v4si)_mm_min_epi32(__A, __B),
|
|
3176
2977
|
(__v4si)__W);
|
|
3177
2978
|
}
|
|
3178
2979
|
|
|
3179
|
-
static __inline__ __m256i
|
|
2980
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3180
2981
|
_mm256_maskz_min_epi32(__mmask8 __M, __m256i __A, __m256i __B) {
|
|
3181
2982
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
3182
2983
|
(__v8si)_mm256_min_epi32(__A, __B),
|
|
3183
2984
|
(__v8si)_mm256_setzero_si256());
|
|
3184
2985
|
}
|
|
3185
2986
|
|
|
3186
|
-
static __inline__ __m256i
|
|
2987
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3187
2988
|
_mm256_mask_min_epi32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) {
|
|
3188
2989
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
3189
2990
|
(__v8si)_mm256_min_epi32(__A, __B),
|
|
3190
2991
|
(__v8si)__W);
|
|
3191
2992
|
}
|
|
3192
2993
|
|
|
3193
|
-
static __inline__ __m128i
|
|
3194
|
-
_mm_min_epi64
|
|
2994
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
2995
|
+
_mm_min_epi64(__m128i __A, __m128i __B) {
|
|
3195
2996
|
return (__m128i)__builtin_elementwise_min((__v2di)__A, (__v2di)__B);
|
|
3196
2997
|
}
|
|
3197
2998
|
|
|
3198
|
-
static __inline__ __m128i
|
|
3199
|
-
_mm_mask_min_epi64
|
|
2999
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3000
|
+
_mm_mask_min_epi64(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
3200
3001
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
3201
3002
|
(__v2di)_mm_min_epi64(__A, __B),
|
|
3202
3003
|
(__v2di)__W);
|
|
3203
3004
|
}
|
|
3204
3005
|
|
|
3205
|
-
static __inline__ __m128i
|
|
3206
|
-
_mm_maskz_min_epi64
|
|
3006
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3007
|
+
_mm_maskz_min_epi64(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
3207
3008
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
3208
3009
|
(__v2di)_mm_min_epi64(__A, __B),
|
|
3209
3010
|
(__v2di)_mm_setzero_si128());
|
|
3210
3011
|
}
|
|
3211
3012
|
|
|
3212
|
-
static __inline__ __m256i
|
|
3213
|
-
_mm256_min_epi64
|
|
3013
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3014
|
+
_mm256_min_epi64(__m256i __A, __m256i __B) {
|
|
3214
3015
|
return (__m256i)__builtin_elementwise_min((__v4di)__A, (__v4di)__B);
|
|
3215
3016
|
}
|
|
3216
3017
|
|
|
3217
|
-
static __inline__ __m256i
|
|
3218
|
-
_mm256_mask_min_epi64
|
|
3018
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3019
|
+
_mm256_mask_min_epi64(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) {
|
|
3219
3020
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
3220
3021
|
(__v4di)_mm256_min_epi64(__A, __B),
|
|
3221
3022
|
(__v4di)__W);
|
|
3222
3023
|
}
|
|
3223
3024
|
|
|
3224
|
-
static __inline__ __m256i
|
|
3225
|
-
_mm256_maskz_min_epi64
|
|
3025
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3026
|
+
_mm256_maskz_min_epi64(__mmask8 __M, __m256i __A, __m256i __B) {
|
|
3226
3027
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
3227
3028
|
(__v4di)_mm256_min_epi64(__A, __B),
|
|
3228
3029
|
(__v4di)_mm256_setzero_si256());
|
|
3229
3030
|
}
|
|
3230
3031
|
|
|
3231
|
-
static __inline__ __m128i
|
|
3032
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3232
3033
|
_mm_maskz_min_epu32(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
3233
3034
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
|
|
3234
3035
|
(__v4si)_mm_min_epu32(__A, __B),
|
|
3235
3036
|
(__v4si)_mm_setzero_si128());
|
|
3236
3037
|
}
|
|
3237
3038
|
|
|
3238
|
-
static __inline__ __m128i
|
|
3039
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3239
3040
|
_mm_mask_min_epu32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
3240
3041
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
|
|
3241
3042
|
(__v4si)_mm_min_epu32(__A, __B),
|
|
3242
3043
|
(__v4si)__W);
|
|
3243
3044
|
}
|
|
3244
3045
|
|
|
3245
|
-
static __inline__ __m256i
|
|
3046
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3246
3047
|
_mm256_maskz_min_epu32(__mmask8 __M, __m256i __A, __m256i __B) {
|
|
3247
3048
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
3248
3049
|
(__v8si)_mm256_min_epu32(__A, __B),
|
|
3249
3050
|
(__v8si)_mm256_setzero_si256());
|
|
3250
3051
|
}
|
|
3251
3052
|
|
|
3252
|
-
static __inline__ __m256i
|
|
3053
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3253
3054
|
_mm256_mask_min_epu32(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) {
|
|
3254
3055
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
3255
3056
|
(__v8si)_mm256_min_epu32(__A, __B),
|
|
3256
3057
|
(__v8si)__W);
|
|
3257
3058
|
}
|
|
3258
3059
|
|
|
3259
|
-
static __inline__ __m128i
|
|
3260
|
-
_mm_min_epu64
|
|
3060
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3061
|
+
_mm_min_epu64(__m128i __A, __m128i __B) {
|
|
3261
3062
|
return (__m128i)__builtin_elementwise_min((__v2du)__A, (__v2du)__B);
|
|
3262
3063
|
}
|
|
3263
3064
|
|
|
3264
|
-
static __inline__ __m128i
|
|
3265
|
-
_mm_mask_min_epu64
|
|
3065
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3066
|
+
_mm_mask_min_epu64(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) {
|
|
3266
3067
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
3267
3068
|
(__v2di)_mm_min_epu64(__A, __B),
|
|
3268
3069
|
(__v2di)__W);
|
|
3269
3070
|
}
|
|
3270
3071
|
|
|
3271
|
-
static __inline__ __m128i
|
|
3272
|
-
_mm_maskz_min_epu64
|
|
3072
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3073
|
+
_mm_maskz_min_epu64(__mmask8 __M, __m128i __A, __m128i __B) {
|
|
3273
3074
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__M,
|
|
3274
3075
|
(__v2di)_mm_min_epu64(__A, __B),
|
|
3275
3076
|
(__v2di)_mm_setzero_si128());
|
|
3276
3077
|
}
|
|
3277
3078
|
|
|
3278
|
-
static __inline__ __m256i
|
|
3279
|
-
_mm256_min_epu64
|
|
3079
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3080
|
+
_mm256_min_epu64(__m256i __A, __m256i __B) {
|
|
3280
3081
|
return (__m256i)__builtin_elementwise_min((__v4du)__A, (__v4du)__B);
|
|
3281
3082
|
}
|
|
3282
3083
|
|
|
3283
|
-
static __inline__ __m256i
|
|
3284
|
-
_mm256_mask_min_epu64
|
|
3084
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3085
|
+
_mm256_mask_min_epu64(__m256i __W, __mmask8 __M, __m256i __A, __m256i __B) {
|
|
3285
3086
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
3286
3087
|
(__v4di)_mm256_min_epu64(__A, __B),
|
|
3287
3088
|
(__v4di)__W);
|
|
3288
3089
|
}
|
|
3289
3090
|
|
|
3290
|
-
static __inline__ __m256i
|
|
3291
|
-
_mm256_maskz_min_epu64
|
|
3091
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3092
|
+
_mm256_maskz_min_epu64(__mmask8 __M, __m256i __A, __m256i __B) {
|
|
3292
3093
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
3293
3094
|
(__v4di)_mm256_min_epu64(__A, __B),
|
|
3294
3095
|
(__v4di)_mm256_setzero_si256());
|
|
@@ -3691,69 +3492,69 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3691
3492
|
(__v8sf)_mm256_setzero_ps());
|
|
3692
3493
|
}
|
|
3693
3494
|
|
|
3694
|
-
static __inline__ __m128d
|
|
3495
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3695
3496
|
_mm_mask_sub_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
|
|
3696
3497
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
3697
3498
|
(__v2df)_mm_sub_pd(__A, __B),
|
|
3698
3499
|
(__v2df)__W);
|
|
3699
3500
|
}
|
|
3700
3501
|
|
|
3701
|
-
static __inline__ __m128d
|
|
3502
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3702
3503
|
_mm_maskz_sub_pd(__mmask8 __U, __m128d __A, __m128d __B) {
|
|
3703
3504
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
3704
3505
|
(__v2df)_mm_sub_pd(__A, __B),
|
|
3705
3506
|
(__v2df)_mm_setzero_pd());
|
|
3706
3507
|
}
|
|
3707
3508
|
|
|
3708
|
-
static __inline__ __m256d
|
|
3509
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3709
3510
|
_mm256_mask_sub_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
|
|
3710
3511
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
3711
3512
|
(__v4df)_mm256_sub_pd(__A, __B),
|
|
3712
3513
|
(__v4df)__W);
|
|
3713
3514
|
}
|
|
3714
3515
|
|
|
3715
|
-
static __inline__ __m256d
|
|
3516
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3716
3517
|
_mm256_maskz_sub_pd(__mmask8 __U, __m256d __A, __m256d __B) {
|
|
3717
3518
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
3718
3519
|
(__v4df)_mm256_sub_pd(__A, __B),
|
|
3719
3520
|
(__v4df)_mm256_setzero_pd());
|
|
3720
3521
|
}
|
|
3721
3522
|
|
|
3722
|
-
static __inline__ __m128
|
|
3523
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3723
3524
|
_mm_mask_sub_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
|
|
3724
3525
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
3725
3526
|
(__v4sf)_mm_sub_ps(__A, __B),
|
|
3726
3527
|
(__v4sf)__W);
|
|
3727
3528
|
}
|
|
3728
3529
|
|
|
3729
|
-
static __inline__ __m128
|
|
3530
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3730
3531
|
_mm_maskz_sub_ps(__mmask8 __U, __m128 __A, __m128 __B) {
|
|
3731
3532
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
3732
3533
|
(__v4sf)_mm_sub_ps(__A, __B),
|
|
3733
3534
|
(__v4sf)_mm_setzero_ps());
|
|
3734
3535
|
}
|
|
3735
3536
|
|
|
3736
|
-
static __inline__ __m256
|
|
3537
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3737
3538
|
_mm256_mask_sub_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
|
|
3738
3539
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
3739
3540
|
(__v8sf)_mm256_sub_ps(__A, __B),
|
|
3740
3541
|
(__v8sf)__W);
|
|
3741
3542
|
}
|
|
3742
3543
|
|
|
3743
|
-
static __inline__ __m256
|
|
3544
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3744
3545
|
_mm256_maskz_sub_ps(__mmask8 __U, __m256 __A, __m256 __B) {
|
|
3745
3546
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
3746
3547
|
(__v8sf)_mm256_sub_ps(__A, __B),
|
|
3747
3548
|
(__v8sf)_mm256_setzero_ps());
|
|
3748
3549
|
}
|
|
3749
3550
|
|
|
3750
|
-
static __inline__ __m128i
|
|
3551
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3751
3552
|
_mm_permutex2var_epi32(__m128i __A, __m128i __I, __m128i __B) {
|
|
3752
3553
|
return (__m128i)__builtin_ia32_vpermi2vard128((__v4si) __A, (__v4si)__I,
|
|
3753
3554
|
(__v4si)__B);
|
|
3754
3555
|
}
|
|
3755
3556
|
|
|
3756
|
-
static __inline__ __m128i
|
|
3557
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3757
3558
|
_mm_mask_permutex2var_epi32(__m128i __A, __mmask8 __U, __m128i __I,
|
|
3758
3559
|
__m128i __B) {
|
|
3759
3560
|
return (__m128i)__builtin_ia32_selectd_128(__U,
|
|
@@ -3761,7 +3562,7 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3761
3562
|
(__v4si)__A);
|
|
3762
3563
|
}
|
|
3763
3564
|
|
|
3764
|
-
static __inline__ __m128i
|
|
3565
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3765
3566
|
_mm_mask2_permutex2var_epi32(__m128i __A, __m128i __I, __mmask8 __U,
|
|
3766
3567
|
__m128i __B) {
|
|
3767
3568
|
return (__m128i)__builtin_ia32_selectd_128(__U,
|
|
@@ -3769,7 +3570,7 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3769
3570
|
(__v4si)__I);
|
|
3770
3571
|
}
|
|
3771
3572
|
|
|
3772
|
-
static __inline__ __m128i
|
|
3573
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3773
3574
|
_mm_maskz_permutex2var_epi32(__mmask8 __U, __m128i __A, __m128i __I,
|
|
3774
3575
|
__m128i __B) {
|
|
3775
3576
|
return (__m128i)__builtin_ia32_selectd_128(__U,
|
|
@@ -3777,13 +3578,13 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3777
3578
|
(__v4si)_mm_setzero_si128());
|
|
3778
3579
|
}
|
|
3779
3580
|
|
|
3780
|
-
static __inline__ __m256i
|
|
3581
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3781
3582
|
_mm256_permutex2var_epi32(__m256i __A, __m256i __I, __m256i __B) {
|
|
3782
3583
|
return (__m256i)__builtin_ia32_vpermi2vard256((__v8si)__A, (__v8si) __I,
|
|
3783
3584
|
(__v8si) __B);
|
|
3784
3585
|
}
|
|
3785
3586
|
|
|
3786
|
-
static __inline__ __m256i
|
|
3587
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3787
3588
|
_mm256_mask_permutex2var_epi32(__m256i __A, __mmask8 __U, __m256i __I,
|
|
3788
3589
|
__m256i __B) {
|
|
3789
3590
|
return (__m256i)__builtin_ia32_selectd_256(__U,
|
|
@@ -3791,7 +3592,7 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3791
3592
|
(__v8si)__A);
|
|
3792
3593
|
}
|
|
3793
3594
|
|
|
3794
|
-
static __inline__ __m256i
|
|
3595
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3795
3596
|
_mm256_mask2_permutex2var_epi32(__m256i __A, __m256i __I, __mmask8 __U,
|
|
3796
3597
|
__m256i __B) {
|
|
3797
3598
|
return (__m256i)__builtin_ia32_selectd_256(__U,
|
|
@@ -3799,7 +3600,7 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3799
3600
|
(__v8si)__I);
|
|
3800
3601
|
}
|
|
3801
3602
|
|
|
3802
|
-
static __inline__ __m256i
|
|
3603
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3803
3604
|
_mm256_maskz_permutex2var_epi32(__mmask8 __U, __m256i __A, __m256i __I,
|
|
3804
3605
|
__m256i __B) {
|
|
3805
3606
|
return (__m256i)__builtin_ia32_selectd_256(__U,
|
|
@@ -3807,40 +3608,43 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3807
3608
|
(__v8si)_mm256_setzero_si256());
|
|
3808
3609
|
}
|
|
3809
3610
|
|
|
3810
|
-
static __inline__ __m128d
|
|
3611
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3811
3612
|
_mm_permutex2var_pd(__m128d __A, __m128i __I, __m128d __B) {
|
|
3812
3613
|
return (__m128d)__builtin_ia32_vpermi2varpd128((__v2df)__A, (__v2di)__I,
|
|
3813
3614
|
(__v2df)__B);
|
|
3814
3615
|
}
|
|
3815
3616
|
|
|
3816
|
-
static __inline__ __m128d
|
|
3817
|
-
_mm_mask_permutex2var_pd(__m128d __A, __mmask8 __U, __m128i __I,
|
|
3617
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3618
|
+
_mm_mask_permutex2var_pd(__m128d __A, __mmask8 __U, __m128i __I,
|
|
3619
|
+
__m128d __B) {
|
|
3818
3620
|
return (__m128d)__builtin_ia32_selectpd_128(__U,
|
|
3819
3621
|
(__v2df)_mm_permutex2var_pd(__A, __I, __B),
|
|
3820
3622
|
(__v2df)__A);
|
|
3821
3623
|
}
|
|
3822
3624
|
|
|
3823
|
-
static __inline__ __m128d
|
|
3824
|
-
_mm_mask2_permutex2var_pd(__m128d __A, __m128i __I, __mmask8 __U,
|
|
3625
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3626
|
+
_mm_mask2_permutex2var_pd(__m128d __A, __m128i __I, __mmask8 __U,
|
|
3627
|
+
__m128d __B) {
|
|
3825
3628
|
return (__m128d)__builtin_ia32_selectpd_128(__U,
|
|
3826
3629
|
(__v2df)_mm_permutex2var_pd(__A, __I, __B),
|
|
3827
3630
|
(__v2df)(__m128d)__I);
|
|
3828
3631
|
}
|
|
3829
3632
|
|
|
3830
|
-
static __inline__ __m128d
|
|
3831
|
-
_mm_maskz_permutex2var_pd(__mmask8 __U, __m128d __A, __m128i __I,
|
|
3633
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3634
|
+
_mm_maskz_permutex2var_pd(__mmask8 __U, __m128d __A, __m128i __I,
|
|
3635
|
+
__m128d __B) {
|
|
3832
3636
|
return (__m128d)__builtin_ia32_selectpd_128(__U,
|
|
3833
3637
|
(__v2df)_mm_permutex2var_pd(__A, __I, __B),
|
|
3834
3638
|
(__v2df)_mm_setzero_pd());
|
|
3835
3639
|
}
|
|
3836
3640
|
|
|
3837
|
-
static __inline__ __m256d
|
|
3641
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3838
3642
|
_mm256_permutex2var_pd(__m256d __A, __m256i __I, __m256d __B) {
|
|
3839
3643
|
return (__m256d)__builtin_ia32_vpermi2varpd256((__v4df)__A, (__v4di)__I,
|
|
3840
3644
|
(__v4df)__B);
|
|
3841
3645
|
}
|
|
3842
3646
|
|
|
3843
|
-
static __inline__ __m256d
|
|
3647
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3844
3648
|
_mm256_mask_permutex2var_pd(__m256d __A, __mmask8 __U, __m256i __I,
|
|
3845
3649
|
__m256d __B) {
|
|
3846
3650
|
return (__m256d)__builtin_ia32_selectpd_256(__U,
|
|
@@ -3848,7 +3652,7 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3848
3652
|
(__v4df)__A);
|
|
3849
3653
|
}
|
|
3850
3654
|
|
|
3851
|
-
static __inline__ __m256d
|
|
3655
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3852
3656
|
_mm256_mask2_permutex2var_pd(__m256d __A, __m256i __I, __mmask8 __U,
|
|
3853
3657
|
__m256d __B) {
|
|
3854
3658
|
return (__m256d)__builtin_ia32_selectpd_256(__U,
|
|
@@ -3856,7 +3660,7 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3856
3660
|
(__v4df)(__m256d)__I);
|
|
3857
3661
|
}
|
|
3858
3662
|
|
|
3859
|
-
static __inline__ __m256d
|
|
3663
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3860
3664
|
_mm256_maskz_permutex2var_pd(__mmask8 __U, __m256d __A, __m256i __I,
|
|
3861
3665
|
__m256d __B) {
|
|
3862
3666
|
return (__m256d)__builtin_ia32_selectpd_256(__U,
|
|
@@ -3864,47 +3668,48 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3864
3668
|
(__v4df)_mm256_setzero_pd());
|
|
3865
3669
|
}
|
|
3866
3670
|
|
|
3867
|
-
static __inline__ __m128
|
|
3671
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3868
3672
|
_mm_permutex2var_ps(__m128 __A, __m128i __I, __m128 __B) {
|
|
3869
3673
|
return (__m128)__builtin_ia32_vpermi2varps128((__v4sf)__A, (__v4si)__I,
|
|
3870
3674
|
(__v4sf)__B);
|
|
3871
3675
|
}
|
|
3872
3676
|
|
|
3873
|
-
static __inline__ __m128
|
|
3677
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3874
3678
|
_mm_mask_permutex2var_ps(__m128 __A, __mmask8 __U, __m128i __I, __m128 __B) {
|
|
3875
3679
|
return (__m128)__builtin_ia32_selectps_128(__U,
|
|
3876
3680
|
(__v4sf)_mm_permutex2var_ps(__A, __I, __B),
|
|
3877
3681
|
(__v4sf)__A);
|
|
3878
3682
|
}
|
|
3879
3683
|
|
|
3880
|
-
static __inline__ __m128
|
|
3684
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3881
3685
|
_mm_mask2_permutex2var_ps(__m128 __A, __m128i __I, __mmask8 __U, __m128 __B) {
|
|
3882
3686
|
return (__m128)__builtin_ia32_selectps_128(__U,
|
|
3883
3687
|
(__v4sf)_mm_permutex2var_ps(__A, __I, __B),
|
|
3884
3688
|
(__v4sf)(__m128)__I);
|
|
3885
3689
|
}
|
|
3886
3690
|
|
|
3887
|
-
static __inline__ __m128
|
|
3691
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3888
3692
|
_mm_maskz_permutex2var_ps(__mmask8 __U, __m128 __A, __m128i __I, __m128 __B) {
|
|
3889
3693
|
return (__m128)__builtin_ia32_selectps_128(__U,
|
|
3890
3694
|
(__v4sf)_mm_permutex2var_ps(__A, __I, __B),
|
|
3891
3695
|
(__v4sf)_mm_setzero_ps());
|
|
3892
3696
|
}
|
|
3893
3697
|
|
|
3894
|
-
static __inline__ __m256
|
|
3698
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3895
3699
|
_mm256_permutex2var_ps(__m256 __A, __m256i __I, __m256 __B) {
|
|
3896
3700
|
return (__m256)__builtin_ia32_vpermi2varps256((__v8sf)__A, (__v8si)__I,
|
|
3897
3701
|
(__v8sf) __B);
|
|
3898
3702
|
}
|
|
3899
3703
|
|
|
3900
|
-
static __inline__ __m256
|
|
3901
|
-
_mm256_mask_permutex2var_ps(__m256 __A, __mmask8 __U, __m256i __I,
|
|
3704
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3705
|
+
_mm256_mask_permutex2var_ps(__m256 __A, __mmask8 __U, __m256i __I,
|
|
3706
|
+
__m256 __B) {
|
|
3902
3707
|
return (__m256)__builtin_ia32_selectps_256(__U,
|
|
3903
3708
|
(__v8sf)_mm256_permutex2var_ps(__A, __I, __B),
|
|
3904
3709
|
(__v8sf)__A);
|
|
3905
3710
|
}
|
|
3906
3711
|
|
|
3907
|
-
static __inline__ __m256
|
|
3712
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3908
3713
|
_mm256_mask2_permutex2var_ps(__m256 __A, __m256i __I, __mmask8 __U,
|
|
3909
3714
|
__m256 __B) {
|
|
3910
3715
|
return (__m256)__builtin_ia32_selectps_256(__U,
|
|
@@ -3912,7 +3717,7 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3912
3717
|
(__v8sf)(__m256)__I);
|
|
3913
3718
|
}
|
|
3914
3719
|
|
|
3915
|
-
static __inline__ __m256
|
|
3720
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3916
3721
|
_mm256_maskz_permutex2var_ps(__mmask8 __U, __m256 __A, __m256i __I,
|
|
3917
3722
|
__m256 __B) {
|
|
3918
3723
|
return (__m256)__builtin_ia32_selectps_256(__U,
|
|
@@ -3920,13 +3725,13 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3920
3725
|
(__v8sf)_mm256_setzero_ps());
|
|
3921
3726
|
}
|
|
3922
3727
|
|
|
3923
|
-
static __inline__ __m128i
|
|
3728
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3924
3729
|
_mm_permutex2var_epi64(__m128i __A, __m128i __I, __m128i __B) {
|
|
3925
3730
|
return (__m128i)__builtin_ia32_vpermi2varq128((__v2di)__A, (__v2di)__I,
|
|
3926
3731
|
(__v2di)__B);
|
|
3927
3732
|
}
|
|
3928
3733
|
|
|
3929
|
-
static __inline__ __m128i
|
|
3734
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3930
3735
|
_mm_mask_permutex2var_epi64(__m128i __A, __mmask8 __U, __m128i __I,
|
|
3931
3736
|
__m128i __B) {
|
|
3932
3737
|
return (__m128i)__builtin_ia32_selectq_128(__U,
|
|
@@ -3934,7 +3739,7 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3934
3739
|
(__v2di)__A);
|
|
3935
3740
|
}
|
|
3936
3741
|
|
|
3937
|
-
static __inline__ __m128i
|
|
3742
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3938
3743
|
_mm_mask2_permutex2var_epi64(__m128i __A, __m128i __I, __mmask8 __U,
|
|
3939
3744
|
__m128i __B) {
|
|
3940
3745
|
return (__m128i)__builtin_ia32_selectq_128(__U,
|
|
@@ -3942,7 +3747,7 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3942
3747
|
(__v2di)__I);
|
|
3943
3748
|
}
|
|
3944
3749
|
|
|
3945
|
-
static __inline__ __m128i
|
|
3750
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
3946
3751
|
_mm_maskz_permutex2var_epi64(__mmask8 __U, __m128i __A, __m128i __I,
|
|
3947
3752
|
__m128i __B) {
|
|
3948
3753
|
return (__m128i)__builtin_ia32_selectq_128(__U,
|
|
@@ -3950,14 +3755,13 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3950
3755
|
(__v2di)_mm_setzero_si128());
|
|
3951
3756
|
}
|
|
3952
3757
|
|
|
3953
|
-
|
|
3954
|
-
static __inline__ __m256i __DEFAULT_FN_ATTRS256
|
|
3758
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3955
3759
|
_mm256_permutex2var_epi64(__m256i __A, __m256i __I, __m256i __B) {
|
|
3956
3760
|
return (__m256i)__builtin_ia32_vpermi2varq256((__v4di)__A, (__v4di) __I,
|
|
3957
3761
|
(__v4di) __B);
|
|
3958
3762
|
}
|
|
3959
3763
|
|
|
3960
|
-
static __inline__ __m256i
|
|
3764
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3961
3765
|
_mm256_mask_permutex2var_epi64(__m256i __A, __mmask8 __U, __m256i __I,
|
|
3962
3766
|
__m256i __B) {
|
|
3963
3767
|
return (__m256i)__builtin_ia32_selectq_256(__U,
|
|
@@ -3965,7 +3769,7 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3965
3769
|
(__v4di)__A);
|
|
3966
3770
|
}
|
|
3967
3771
|
|
|
3968
|
-
static __inline__ __m256i
|
|
3772
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3969
3773
|
_mm256_mask2_permutex2var_epi64(__m256i __A, __m256i __I, __mmask8 __U,
|
|
3970
3774
|
__m256i __B) {
|
|
3971
3775
|
return (__m256i)__builtin_ia32_selectq_256(__U,
|
|
@@ -3973,7 +3777,7 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
3973
3777
|
(__v4di)__I);
|
|
3974
3778
|
}
|
|
3975
3779
|
|
|
3976
|
-
static __inline__ __m256i
|
|
3780
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
3977
3781
|
_mm256_maskz_permutex2var_epi64(__mmask8 __U, __m256i __A, __m256i __I,
|
|
3978
3782
|
__m256i __B) {
|
|
3979
3783
|
return (__m256i)__builtin_ia32_selectq_256(__U,
|
|
@@ -4355,13 +4159,13 @@ _mm256_maskz_scalef_ps (__mmask8 __U, __m256 __A, __m256 __B) {
|
|
|
4355
4159
|
(__v4di)_mm256_rol_epi64((a), (b)), \
|
|
4356
4160
|
(__v4di)_mm256_setzero_si256()))
|
|
4357
4161
|
|
|
4358
|
-
static __inline__ __m128i
|
|
4162
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4359
4163
|
_mm_rolv_epi32 (__m128i __A, __m128i __B)
|
|
4360
4164
|
{
|
|
4361
|
-
return (__m128i)
|
|
4165
|
+
return (__m128i)__builtin_elementwise_fshl((__v4su)__A, (__v4su)__A, (__v4su)__B);
|
|
4362
4166
|
}
|
|
4363
4167
|
|
|
4364
|
-
static __inline__ __m128i
|
|
4168
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4365
4169
|
_mm_mask_rolv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
4366
4170
|
{
|
|
4367
4171
|
return (__m128i)__builtin_ia32_selectd_128(__U,
|
|
@@ -4369,7 +4173,7 @@ _mm_mask_rolv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
|
4369
4173
|
(__v4si)__W);
|
|
4370
4174
|
}
|
|
4371
4175
|
|
|
4372
|
-
static __inline__ __m128i
|
|
4176
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4373
4177
|
_mm_maskz_rolv_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
|
|
4374
4178
|
{
|
|
4375
4179
|
return (__m128i)__builtin_ia32_selectd_128(__U,
|
|
@@ -4377,13 +4181,13 @@ _mm_maskz_rolv_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
|
|
|
4377
4181
|
(__v4si)_mm_setzero_si128());
|
|
4378
4182
|
}
|
|
4379
4183
|
|
|
4380
|
-
static __inline__ __m256i
|
|
4184
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4381
4185
|
_mm256_rolv_epi32 (__m256i __A, __m256i __B)
|
|
4382
4186
|
{
|
|
4383
|
-
return (__m256i)
|
|
4187
|
+
return (__m256i)__builtin_elementwise_fshl((__v8su)__A, (__v8su)__A, (__v8su)__B);
|
|
4384
4188
|
}
|
|
4385
4189
|
|
|
4386
|
-
static __inline__ __m256i
|
|
4190
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4387
4191
|
_mm256_mask_rolv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
4388
4192
|
{
|
|
4389
4193
|
return (__m256i)__builtin_ia32_selectd_256(__U,
|
|
@@ -4391,7 +4195,7 @@ _mm256_mask_rolv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
|
4391
4195
|
(__v8si)__W);
|
|
4392
4196
|
}
|
|
4393
4197
|
|
|
4394
|
-
static __inline__ __m256i
|
|
4198
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4395
4199
|
_mm256_maskz_rolv_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
|
|
4396
4200
|
{
|
|
4397
4201
|
return (__m256i)__builtin_ia32_selectd_256(__U,
|
|
@@ -4399,13 +4203,13 @@ _mm256_maskz_rolv_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
|
|
|
4399
4203
|
(__v8si)_mm256_setzero_si256());
|
|
4400
4204
|
}
|
|
4401
4205
|
|
|
4402
|
-
static __inline__ __m128i
|
|
4206
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4403
4207
|
_mm_rolv_epi64 (__m128i __A, __m128i __B)
|
|
4404
4208
|
{
|
|
4405
|
-
return (__m128i)
|
|
4209
|
+
return (__m128i)__builtin_elementwise_fshl((__v2du)__A, (__v2du)__A, (__v2du)__B);
|
|
4406
4210
|
}
|
|
4407
4211
|
|
|
4408
|
-
static __inline__ __m128i
|
|
4212
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4409
4213
|
_mm_mask_rolv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
4410
4214
|
{
|
|
4411
4215
|
return (__m128i)__builtin_ia32_selectq_128(__U,
|
|
@@ -4413,7 +4217,7 @@ _mm_mask_rolv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
|
4413
4217
|
(__v2di)__W);
|
|
4414
4218
|
}
|
|
4415
4219
|
|
|
4416
|
-
static __inline__ __m128i
|
|
4220
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4417
4221
|
_mm_maskz_rolv_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
|
|
4418
4222
|
{
|
|
4419
4223
|
return (__m128i)__builtin_ia32_selectq_128(__U,
|
|
@@ -4421,13 +4225,13 @@ _mm_maskz_rolv_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
|
|
|
4421
4225
|
(__v2di)_mm_setzero_si128());
|
|
4422
4226
|
}
|
|
4423
4227
|
|
|
4424
|
-
static __inline__ __m256i
|
|
4228
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4425
4229
|
_mm256_rolv_epi64 (__m256i __A, __m256i __B)
|
|
4426
4230
|
{
|
|
4427
|
-
return (__m256i)
|
|
4231
|
+
return (__m256i)__builtin_elementwise_fshl((__v4du)__A, (__v4du)__A, (__v4du)__B);
|
|
4428
4232
|
}
|
|
4429
4233
|
|
|
4430
|
-
static __inline__ __m256i
|
|
4234
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4431
4235
|
_mm256_mask_rolv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
4432
4236
|
{
|
|
4433
4237
|
return (__m256i)__builtin_ia32_selectq_256(__U,
|
|
@@ -4435,7 +4239,7 @@ _mm256_mask_rolv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
|
4435
4239
|
(__v4di)__W);
|
|
4436
4240
|
}
|
|
4437
4241
|
|
|
4438
|
-
static __inline__ __m256i
|
|
4242
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4439
4243
|
_mm256_maskz_rolv_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
|
|
4440
4244
|
{
|
|
4441
4245
|
return (__m256i)__builtin_ia32_selectq_256(__U,
|
|
@@ -4495,33 +4299,29 @@ _mm256_maskz_rolv_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
|
|
|
4495
4299
|
(__v4di)_mm256_ror_epi64((a), (b)), \
|
|
4496
4300
|
(__v4di)_mm256_setzero_si256()))
|
|
4497
4301
|
|
|
4498
|
-
static __inline__ __m128i
|
|
4499
|
-
_mm_mask_sll_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
4500
|
-
{
|
|
4302
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4303
|
+
_mm_mask_sll_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
4501
4304
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
4502
4305
|
(__v4si)_mm_sll_epi32(__A, __B),
|
|
4503
4306
|
(__v4si)__W);
|
|
4504
4307
|
}
|
|
4505
4308
|
|
|
4506
|
-
static __inline__ __m128i
|
|
4507
|
-
_mm_maskz_sll_epi32(__mmask8 __U, __m128i __A, __m128i __B)
|
|
4508
|
-
{
|
|
4309
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4310
|
+
_mm_maskz_sll_epi32(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
4509
4311
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
4510
4312
|
(__v4si)_mm_sll_epi32(__A, __B),
|
|
4511
4313
|
(__v4si)_mm_setzero_si128());
|
|
4512
4314
|
}
|
|
4513
4315
|
|
|
4514
|
-
static __inline__ __m256i
|
|
4515
|
-
_mm256_mask_sll_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B)
|
|
4516
|
-
{
|
|
4316
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4317
|
+
_mm256_mask_sll_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) {
|
|
4517
4318
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
4518
4319
|
(__v8si)_mm256_sll_epi32(__A, __B),
|
|
4519
4320
|
(__v8si)__W);
|
|
4520
4321
|
}
|
|
4521
4322
|
|
|
4522
|
-
static __inline__ __m256i
|
|
4523
|
-
_mm256_maskz_sll_epi32(__mmask8 __U, __m256i __A, __m128i __B)
|
|
4524
|
-
{
|
|
4323
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4324
|
+
_mm256_maskz_sll_epi32(__mmask8 __U, __m256i __A, __m128i __B) {
|
|
4525
4325
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
4526
4326
|
(__v8si)_mm256_sll_epi32(__A, __B),
|
|
4527
4327
|
(__v8si)_mm256_setzero_si256());
|
|
@@ -4543,49 +4343,44 @@ _mm_maskz_slli_epi32(__mmask8 __U, __m128i __A, unsigned int __B)
|
|
|
4543
4343
|
(__v4si)_mm_setzero_si128());
|
|
4544
4344
|
}
|
|
4545
4345
|
|
|
4546
|
-
static __inline__ __m256i
|
|
4547
|
-
_mm256_mask_slli_epi32(__m256i __W, __mmask8 __U, __m256i __A,
|
|
4548
|
-
{
|
|
4346
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4347
|
+
_mm256_mask_slli_epi32(__m256i __W, __mmask8 __U, __m256i __A,
|
|
4348
|
+
unsigned int __B) {
|
|
4549
4349
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
4550
4350
|
(__v8si)_mm256_slli_epi32(__A, (int)__B),
|
|
4551
4351
|
(__v8si)__W);
|
|
4552
4352
|
}
|
|
4553
4353
|
|
|
4554
|
-
static __inline__ __m256i
|
|
4555
|
-
_mm256_maskz_slli_epi32(__mmask8 __U, __m256i __A, unsigned int __B)
|
|
4556
|
-
{
|
|
4354
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4355
|
+
_mm256_maskz_slli_epi32(__mmask8 __U, __m256i __A, unsigned int __B) {
|
|
4557
4356
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
4558
4357
|
(__v8si)_mm256_slli_epi32(__A, (int)__B),
|
|
4559
4358
|
(__v8si)_mm256_setzero_si256());
|
|
4560
4359
|
}
|
|
4561
4360
|
|
|
4562
|
-
static __inline__ __m128i
|
|
4563
|
-
_mm_mask_sll_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
4564
|
-
{
|
|
4361
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4362
|
+
_mm_mask_sll_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
4565
4363
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
4566
4364
|
(__v2di)_mm_sll_epi64(__A, __B),
|
|
4567
4365
|
(__v2di)__W);
|
|
4568
4366
|
}
|
|
4569
4367
|
|
|
4570
|
-
static __inline__ __m128i
|
|
4571
|
-
_mm_maskz_sll_epi64(__mmask8 __U, __m128i __A, __m128i __B)
|
|
4572
|
-
{
|
|
4368
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4369
|
+
_mm_maskz_sll_epi64(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
4573
4370
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
4574
4371
|
(__v2di)_mm_sll_epi64(__A, __B),
|
|
4575
4372
|
(__v2di)_mm_setzero_si128());
|
|
4576
4373
|
}
|
|
4577
4374
|
|
|
4578
|
-
static __inline__ __m256i
|
|
4579
|
-
_mm256_mask_sll_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B)
|
|
4580
|
-
{
|
|
4375
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4376
|
+
_mm256_mask_sll_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) {
|
|
4581
4377
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
4582
4378
|
(__v4di)_mm256_sll_epi64(__A, __B),
|
|
4583
4379
|
(__v4di)__W);
|
|
4584
4380
|
}
|
|
4585
4381
|
|
|
4586
|
-
static __inline__ __m256i
|
|
4587
|
-
_mm256_maskz_sll_epi64(__mmask8 __U, __m256i __A, __m128i __B)
|
|
4588
|
-
{
|
|
4382
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4383
|
+
_mm256_maskz_sll_epi64(__mmask8 __U, __m256i __A, __m128i __B) {
|
|
4589
4384
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
4590
4385
|
(__v4di)_mm256_sll_epi64(__A, __B),
|
|
4591
4386
|
(__v4di)_mm256_setzero_si256());
|
|
@@ -4607,29 +4402,28 @@ _mm_maskz_slli_epi64(__mmask8 __U, __m128i __A, unsigned int __B)
|
|
|
4607
4402
|
(__v2di)_mm_setzero_si128());
|
|
4608
4403
|
}
|
|
4609
4404
|
|
|
4610
|
-
static __inline__ __m256i
|
|
4611
|
-
_mm256_mask_slli_epi64(__m256i __W, __mmask8 __U, __m256i __A,
|
|
4612
|
-
{
|
|
4405
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4406
|
+
_mm256_mask_slli_epi64(__m256i __W, __mmask8 __U, __m256i __A,
|
|
4407
|
+
unsigned int __B) {
|
|
4613
4408
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
4614
4409
|
(__v4di)_mm256_slli_epi64(__A, (int)__B),
|
|
4615
4410
|
(__v4di)__W);
|
|
4616
4411
|
}
|
|
4617
4412
|
|
|
4618
|
-
static __inline__ __m256i
|
|
4619
|
-
_mm256_maskz_slli_epi64(__mmask8 __U, __m256i __A, unsigned int __B)
|
|
4620
|
-
{
|
|
4413
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4414
|
+
_mm256_maskz_slli_epi64(__mmask8 __U, __m256i __A, unsigned int __B) {
|
|
4621
4415
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
4622
4416
|
(__v4di)_mm256_slli_epi64(__A, (int)__B),
|
|
4623
4417
|
(__v4di)_mm256_setzero_si256());
|
|
4624
4418
|
}
|
|
4625
4419
|
|
|
4626
|
-
static __inline__ __m128i
|
|
4420
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4627
4421
|
_mm_rorv_epi32 (__m128i __A, __m128i __B)
|
|
4628
4422
|
{
|
|
4629
|
-
return (__m128i)
|
|
4423
|
+
return (__m128i)__builtin_elementwise_fshr((__v4su)__A, (__v4su)__A, (__v4su)__B);
|
|
4630
4424
|
}
|
|
4631
4425
|
|
|
4632
|
-
static __inline__ __m128i
|
|
4426
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4633
4427
|
_mm_mask_rorv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
4634
4428
|
{
|
|
4635
4429
|
return (__m128i)__builtin_ia32_selectd_128(__U,
|
|
@@ -4637,7 +4431,7 @@ _mm_mask_rorv_epi32 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
|
4637
4431
|
(__v4si)__W);
|
|
4638
4432
|
}
|
|
4639
4433
|
|
|
4640
|
-
static __inline__ __m128i
|
|
4434
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4641
4435
|
_mm_maskz_rorv_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
|
|
4642
4436
|
{
|
|
4643
4437
|
return (__m128i)__builtin_ia32_selectd_128(__U,
|
|
@@ -4645,13 +4439,13 @@ _mm_maskz_rorv_epi32 (__mmask8 __U, __m128i __A, __m128i __B)
|
|
|
4645
4439
|
(__v4si)_mm_setzero_si128());
|
|
4646
4440
|
}
|
|
4647
4441
|
|
|
4648
|
-
static __inline__ __m256i
|
|
4442
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4649
4443
|
_mm256_rorv_epi32 (__m256i __A, __m256i __B)
|
|
4650
4444
|
{
|
|
4651
|
-
return (__m256i)
|
|
4445
|
+
return (__m256i)__builtin_elementwise_fshr((__v8su)__A, (__v8su)__A, (__v8su)__B);
|
|
4652
4446
|
}
|
|
4653
4447
|
|
|
4654
|
-
static __inline__ __m256i
|
|
4448
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4655
4449
|
_mm256_mask_rorv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
4656
4450
|
{
|
|
4657
4451
|
return (__m256i)__builtin_ia32_selectd_256(__U,
|
|
@@ -4659,7 +4453,7 @@ _mm256_mask_rorv_epi32 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
|
4659
4453
|
(__v8si)__W);
|
|
4660
4454
|
}
|
|
4661
4455
|
|
|
4662
|
-
static __inline__ __m256i
|
|
4456
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4663
4457
|
_mm256_maskz_rorv_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
|
|
4664
4458
|
{
|
|
4665
4459
|
return (__m256i)__builtin_ia32_selectd_256(__U,
|
|
@@ -4667,13 +4461,13 @@ _mm256_maskz_rorv_epi32 (__mmask8 __U, __m256i __A, __m256i __B)
|
|
|
4667
4461
|
(__v8si)_mm256_setzero_si256());
|
|
4668
4462
|
}
|
|
4669
4463
|
|
|
4670
|
-
static __inline__ __m128i
|
|
4464
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4671
4465
|
_mm_rorv_epi64 (__m128i __A, __m128i __B)
|
|
4672
4466
|
{
|
|
4673
|
-
return (__m128i)
|
|
4467
|
+
return (__m128i)__builtin_elementwise_fshr((__v2du)__A, (__v2du)__A, (__v2du)__B);
|
|
4674
4468
|
}
|
|
4675
4469
|
|
|
4676
|
-
static __inline__ __m128i
|
|
4470
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4677
4471
|
_mm_mask_rorv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
4678
4472
|
{
|
|
4679
4473
|
return (__m128i)__builtin_ia32_selectq_128(__U,
|
|
@@ -4681,7 +4475,7 @@ _mm_mask_rorv_epi64 (__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
|
4681
4475
|
(__v2di)__W);
|
|
4682
4476
|
}
|
|
4683
4477
|
|
|
4684
|
-
static __inline__ __m128i
|
|
4478
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4685
4479
|
_mm_maskz_rorv_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
|
|
4686
4480
|
{
|
|
4687
4481
|
return (__m128i)__builtin_ia32_selectq_128(__U,
|
|
@@ -4689,13 +4483,13 @@ _mm_maskz_rorv_epi64 (__mmask8 __U, __m128i __A, __m128i __B)
|
|
|
4689
4483
|
(__v2di)_mm_setzero_si128());
|
|
4690
4484
|
}
|
|
4691
4485
|
|
|
4692
|
-
static __inline__ __m256i
|
|
4486
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4693
4487
|
_mm256_rorv_epi64 (__m256i __A, __m256i __B)
|
|
4694
4488
|
{
|
|
4695
|
-
return (__m256i)
|
|
4489
|
+
return (__m256i)__builtin_elementwise_fshr((__v4du)__A, (__v4du)__A, (__v4du)__B);
|
|
4696
4490
|
}
|
|
4697
4491
|
|
|
4698
|
-
static __inline__ __m256i
|
|
4492
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4699
4493
|
_mm256_mask_rorv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
4700
4494
|
{
|
|
4701
4495
|
return (__m256i)__builtin_ia32_selectq_256(__U,
|
|
@@ -4703,7 +4497,7 @@ _mm256_mask_rorv_epi64 (__m256i __W, __mmask8 __U, __m256i __A, __m256i __B)
|
|
|
4703
4497
|
(__v4di)__W);
|
|
4704
4498
|
}
|
|
4705
4499
|
|
|
4706
|
-
static __inline__ __m256i
|
|
4500
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4707
4501
|
_mm256_maskz_rorv_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
|
|
4708
4502
|
{
|
|
4709
4503
|
return (__m256i)__builtin_ia32_selectq_256(__U,
|
|
@@ -4711,7 +4505,7 @@ _mm256_maskz_rorv_epi64 (__mmask8 __U, __m256i __A, __m256i __B)
|
|
|
4711
4505
|
(__v4di)_mm256_setzero_si256());
|
|
4712
4506
|
}
|
|
4713
4507
|
|
|
4714
|
-
static __inline__ __m128i
|
|
4508
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4715
4509
|
_mm_mask_sllv_epi64(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
4716
4510
|
{
|
|
4717
4511
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
@@ -4719,7 +4513,7 @@ _mm_mask_sllv_epi64(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
4719
4513
|
(__v2di)__W);
|
|
4720
4514
|
}
|
|
4721
4515
|
|
|
4722
|
-
static __inline__ __m128i
|
|
4516
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4723
4517
|
_mm_maskz_sllv_epi64(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
4724
4518
|
{
|
|
4725
4519
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
@@ -4727,7 +4521,7 @@ _mm_maskz_sllv_epi64(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
4727
4521
|
(__v2di)_mm_setzero_si128());
|
|
4728
4522
|
}
|
|
4729
4523
|
|
|
4730
|
-
static __inline__ __m256i
|
|
4524
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4731
4525
|
_mm256_mask_sllv_epi64(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
4732
4526
|
{
|
|
4733
4527
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
@@ -4735,7 +4529,7 @@ _mm256_mask_sllv_epi64(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
|
4735
4529
|
(__v4di)__W);
|
|
4736
4530
|
}
|
|
4737
4531
|
|
|
4738
|
-
static __inline__ __m256i
|
|
4532
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4739
4533
|
_mm256_maskz_sllv_epi64(__mmask8 __U, __m256i __X, __m256i __Y)
|
|
4740
4534
|
{
|
|
4741
4535
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
@@ -4743,7 +4537,7 @@ _mm256_maskz_sllv_epi64(__mmask8 __U, __m256i __X, __m256i __Y)
|
|
|
4743
4537
|
(__v4di)_mm256_setzero_si256());
|
|
4744
4538
|
}
|
|
4745
4539
|
|
|
4746
|
-
static __inline__ __m128i
|
|
4540
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4747
4541
|
_mm_mask_sllv_epi32(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
4748
4542
|
{
|
|
4749
4543
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
@@ -4751,7 +4545,7 @@ _mm_mask_sllv_epi32(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
4751
4545
|
(__v4si)__W);
|
|
4752
4546
|
}
|
|
4753
4547
|
|
|
4754
|
-
static __inline__ __m128i
|
|
4548
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4755
4549
|
_mm_maskz_sllv_epi32(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
4756
4550
|
{
|
|
4757
4551
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
@@ -4759,7 +4553,7 @@ _mm_maskz_sllv_epi32(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
4759
4553
|
(__v4si)_mm_setzero_si128());
|
|
4760
4554
|
}
|
|
4761
4555
|
|
|
4762
|
-
static __inline__ __m256i
|
|
4556
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4763
4557
|
_mm256_mask_sllv_epi32(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
4764
4558
|
{
|
|
4765
4559
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
@@ -4767,7 +4561,7 @@ _mm256_mask_sllv_epi32(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
|
4767
4561
|
(__v8si)__W);
|
|
4768
4562
|
}
|
|
4769
4563
|
|
|
4770
|
-
static __inline__ __m256i
|
|
4564
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4771
4565
|
_mm256_maskz_sllv_epi32(__mmask8 __U, __m256i __X, __m256i __Y)
|
|
4772
4566
|
{
|
|
4773
4567
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
@@ -4775,7 +4569,7 @@ _mm256_maskz_sllv_epi32(__mmask8 __U, __m256i __X, __m256i __Y)
|
|
|
4775
4569
|
(__v8si)_mm256_setzero_si256());
|
|
4776
4570
|
}
|
|
4777
4571
|
|
|
4778
|
-
static __inline__ __m128i
|
|
4572
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4779
4573
|
_mm_mask_srlv_epi64(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
4780
4574
|
{
|
|
4781
4575
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
@@ -4783,7 +4577,7 @@ _mm_mask_srlv_epi64(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
4783
4577
|
(__v2di)__W);
|
|
4784
4578
|
}
|
|
4785
4579
|
|
|
4786
|
-
static __inline__ __m128i
|
|
4580
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4787
4581
|
_mm_maskz_srlv_epi64(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
4788
4582
|
{
|
|
4789
4583
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
@@ -4791,7 +4585,7 @@ _mm_maskz_srlv_epi64(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
4791
4585
|
(__v2di)_mm_setzero_si128());
|
|
4792
4586
|
}
|
|
4793
4587
|
|
|
4794
|
-
static __inline__ __m256i
|
|
4588
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4795
4589
|
_mm256_mask_srlv_epi64(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
4796
4590
|
{
|
|
4797
4591
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
@@ -4799,7 +4593,7 @@ _mm256_mask_srlv_epi64(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
|
4799
4593
|
(__v4di)__W);
|
|
4800
4594
|
}
|
|
4801
4595
|
|
|
4802
|
-
static __inline__ __m256i
|
|
4596
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4803
4597
|
_mm256_maskz_srlv_epi64(__mmask8 __U, __m256i __X, __m256i __Y)
|
|
4804
4598
|
{
|
|
4805
4599
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
@@ -4807,7 +4601,7 @@ _mm256_maskz_srlv_epi64(__mmask8 __U, __m256i __X, __m256i __Y)
|
|
|
4807
4601
|
(__v4di)_mm256_setzero_si256());
|
|
4808
4602
|
}
|
|
4809
4603
|
|
|
4810
|
-
static __inline__ __m128i
|
|
4604
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4811
4605
|
_mm_mask_srlv_epi32(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
4812
4606
|
{
|
|
4813
4607
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
@@ -4815,7 +4609,7 @@ _mm_mask_srlv_epi32(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
4815
4609
|
(__v4si)__W);
|
|
4816
4610
|
}
|
|
4817
4611
|
|
|
4818
|
-
static __inline__ __m128i
|
|
4612
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4819
4613
|
_mm_maskz_srlv_epi32(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
4820
4614
|
{
|
|
4821
4615
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
@@ -4823,7 +4617,7 @@ _mm_maskz_srlv_epi32(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
4823
4617
|
(__v4si)_mm_setzero_si128());
|
|
4824
4618
|
}
|
|
4825
4619
|
|
|
4826
|
-
static __inline__ __m256i
|
|
4620
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4827
4621
|
_mm256_mask_srlv_epi32(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
4828
4622
|
{
|
|
4829
4623
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
@@ -4831,7 +4625,7 @@ _mm256_mask_srlv_epi32(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
|
4831
4625
|
(__v8si)__W);
|
|
4832
4626
|
}
|
|
4833
4627
|
|
|
4834
|
-
static __inline__ __m256i
|
|
4628
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4835
4629
|
_mm256_maskz_srlv_epi32(__mmask8 __U, __m256i __X, __m256i __Y)
|
|
4836
4630
|
{
|
|
4837
4631
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
@@ -4839,33 +4633,29 @@ _mm256_maskz_srlv_epi32(__mmask8 __U, __m256i __X, __m256i __Y)
|
|
|
4839
4633
|
(__v8si)_mm256_setzero_si256());
|
|
4840
4634
|
}
|
|
4841
4635
|
|
|
4842
|
-
static __inline__ __m128i
|
|
4843
|
-
_mm_mask_srl_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
4844
|
-
{
|
|
4636
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4637
|
+
_mm_mask_srl_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
4845
4638
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
4846
4639
|
(__v4si)_mm_srl_epi32(__A, __B),
|
|
4847
4640
|
(__v4si)__W);
|
|
4848
4641
|
}
|
|
4849
4642
|
|
|
4850
|
-
static __inline__ __m128i
|
|
4851
|
-
_mm_maskz_srl_epi32(__mmask8 __U, __m128i __A, __m128i __B)
|
|
4852
|
-
{
|
|
4643
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4644
|
+
_mm_maskz_srl_epi32(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
4853
4645
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
4854
4646
|
(__v4si)_mm_srl_epi32(__A, __B),
|
|
4855
4647
|
(__v4si)_mm_setzero_si128());
|
|
4856
4648
|
}
|
|
4857
4649
|
|
|
4858
|
-
static __inline__ __m256i
|
|
4859
|
-
_mm256_mask_srl_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B)
|
|
4860
|
-
{
|
|
4650
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4651
|
+
_mm256_mask_srl_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) {
|
|
4861
4652
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
4862
4653
|
(__v8si)_mm256_srl_epi32(__A, __B),
|
|
4863
4654
|
(__v8si)__W);
|
|
4864
4655
|
}
|
|
4865
4656
|
|
|
4866
|
-
static __inline__ __m256i
|
|
4867
|
-
_mm256_maskz_srl_epi32(__mmask8 __U, __m256i __A, __m128i __B)
|
|
4868
|
-
{
|
|
4657
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4658
|
+
_mm256_maskz_srl_epi32(__mmask8 __U, __m256i __A, __m128i __B) {
|
|
4869
4659
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
4870
4660
|
(__v8si)_mm256_srl_epi32(__A, __B),
|
|
4871
4661
|
(__v8si)_mm256_setzero_si256());
|
|
@@ -4887,49 +4677,44 @@ _mm_maskz_srli_epi32(__mmask8 __U, __m128i __A, unsigned int __B)
|
|
|
4887
4677
|
(__v4si)_mm_setzero_si128());
|
|
4888
4678
|
}
|
|
4889
4679
|
|
|
4890
|
-
static __inline__ __m256i
|
|
4891
|
-
_mm256_mask_srli_epi32(__m256i __W, __mmask8 __U, __m256i __A,
|
|
4892
|
-
{
|
|
4680
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4681
|
+
_mm256_mask_srli_epi32(__m256i __W, __mmask8 __U, __m256i __A,
|
|
4682
|
+
unsigned int __B) {
|
|
4893
4683
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
4894
4684
|
(__v8si)_mm256_srli_epi32(__A, (int)__B),
|
|
4895
4685
|
(__v8si)__W);
|
|
4896
4686
|
}
|
|
4897
4687
|
|
|
4898
|
-
static __inline__ __m256i
|
|
4899
|
-
_mm256_maskz_srli_epi32(__mmask8 __U, __m256i __A, unsigned int __B)
|
|
4900
|
-
{
|
|
4688
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4689
|
+
_mm256_maskz_srli_epi32(__mmask8 __U, __m256i __A, unsigned int __B) {
|
|
4901
4690
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
4902
4691
|
(__v8si)_mm256_srli_epi32(__A, (int)__B),
|
|
4903
4692
|
(__v8si)_mm256_setzero_si256());
|
|
4904
4693
|
}
|
|
4905
4694
|
|
|
4906
|
-
static __inline__ __m128i
|
|
4907
|
-
_mm_mask_srl_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
4908
|
-
{
|
|
4695
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4696
|
+
_mm_mask_srl_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
4909
4697
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
4910
4698
|
(__v2di)_mm_srl_epi64(__A, __B),
|
|
4911
4699
|
(__v2di)__W);
|
|
4912
4700
|
}
|
|
4913
4701
|
|
|
4914
|
-
static __inline__ __m128i
|
|
4915
|
-
_mm_maskz_srl_epi64(__mmask8 __U, __m128i __A, __m128i __B)
|
|
4916
|
-
{
|
|
4702
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4703
|
+
_mm_maskz_srl_epi64(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
4917
4704
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
4918
4705
|
(__v2di)_mm_srl_epi64(__A, __B),
|
|
4919
4706
|
(__v2di)_mm_setzero_si128());
|
|
4920
4707
|
}
|
|
4921
4708
|
|
|
4922
|
-
static __inline__ __m256i
|
|
4923
|
-
_mm256_mask_srl_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B)
|
|
4924
|
-
{
|
|
4709
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4710
|
+
_mm256_mask_srl_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) {
|
|
4925
4711
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
4926
4712
|
(__v4di)_mm256_srl_epi64(__A, __B),
|
|
4927
4713
|
(__v4di)__W);
|
|
4928
4714
|
}
|
|
4929
4715
|
|
|
4930
|
-
static __inline__ __m256i
|
|
4931
|
-
_mm256_maskz_srl_epi64(__mmask8 __U, __m256i __A, __m128i __B)
|
|
4932
|
-
{
|
|
4716
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4717
|
+
_mm256_maskz_srl_epi64(__mmask8 __U, __m256i __A, __m128i __B) {
|
|
4933
4718
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
4934
4719
|
(__v4di)_mm256_srl_epi64(__A, __B),
|
|
4935
4720
|
(__v4di)_mm256_setzero_si256());
|
|
@@ -4951,23 +4736,22 @@ _mm_maskz_srli_epi64(__mmask8 __U, __m128i __A, unsigned int __B)
|
|
|
4951
4736
|
(__v2di)_mm_setzero_si128());
|
|
4952
4737
|
}
|
|
4953
4738
|
|
|
4954
|
-
static __inline__ __m256i
|
|
4955
|
-
_mm256_mask_srli_epi64(__m256i __W, __mmask8 __U, __m256i __A,
|
|
4956
|
-
{
|
|
4739
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4740
|
+
_mm256_mask_srli_epi64(__m256i __W, __mmask8 __U, __m256i __A,
|
|
4741
|
+
unsigned int __B) {
|
|
4957
4742
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
4958
4743
|
(__v4di)_mm256_srli_epi64(__A, (int)__B),
|
|
4959
4744
|
(__v4di)__W);
|
|
4960
4745
|
}
|
|
4961
4746
|
|
|
4962
|
-
static __inline__ __m256i
|
|
4963
|
-
_mm256_maskz_srli_epi64(__mmask8 __U, __m256i __A, unsigned int __B)
|
|
4964
|
-
{
|
|
4747
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4748
|
+
_mm256_maskz_srli_epi64(__mmask8 __U, __m256i __A, unsigned int __B) {
|
|
4965
4749
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
4966
4750
|
(__v4di)_mm256_srli_epi64(__A, (int)__B),
|
|
4967
4751
|
(__v4di)_mm256_setzero_si256());
|
|
4968
4752
|
}
|
|
4969
4753
|
|
|
4970
|
-
static __inline__ __m128i
|
|
4754
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4971
4755
|
_mm_mask_srav_epi32(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
4972
4756
|
{
|
|
4973
4757
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
@@ -4975,7 +4759,7 @@ _mm_mask_srav_epi32(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
4975
4759
|
(__v4si)__W);
|
|
4976
4760
|
}
|
|
4977
4761
|
|
|
4978
|
-
static __inline__ __m128i
|
|
4762
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
4979
4763
|
_mm_maskz_srav_epi32(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
4980
4764
|
{
|
|
4981
4765
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
@@ -4983,7 +4767,7 @@ _mm_maskz_srav_epi32(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
4983
4767
|
(__v4si)_mm_setzero_si128());
|
|
4984
4768
|
}
|
|
4985
4769
|
|
|
4986
|
-
static __inline__ __m256i
|
|
4770
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4987
4771
|
_mm256_mask_srav_epi32(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
4988
4772
|
{
|
|
4989
4773
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
@@ -4991,7 +4775,7 @@ _mm256_mask_srav_epi32(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
|
4991
4775
|
(__v8si)__W);
|
|
4992
4776
|
}
|
|
4993
4777
|
|
|
4994
|
-
static __inline__ __m256i
|
|
4778
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
4995
4779
|
_mm256_maskz_srav_epi32(__mmask8 __U, __m256i __X, __m256i __Y)
|
|
4996
4780
|
{
|
|
4997
4781
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
@@ -4999,13 +4783,13 @@ _mm256_maskz_srav_epi32(__mmask8 __U, __m256i __X, __m256i __Y)
|
|
|
4999
4783
|
(__v8si)_mm256_setzero_si256());
|
|
5000
4784
|
}
|
|
5001
4785
|
|
|
5002
|
-
static __inline__ __m128i
|
|
4786
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5003
4787
|
_mm_srav_epi64(__m128i __X, __m128i __Y)
|
|
5004
4788
|
{
|
|
5005
4789
|
return (__m128i)__builtin_ia32_psravq128((__v2di)__X, (__v2di)__Y);
|
|
5006
4790
|
}
|
|
5007
4791
|
|
|
5008
|
-
static __inline__ __m128i
|
|
4792
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5009
4793
|
_mm_mask_srav_epi64(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
5010
4794
|
{
|
|
5011
4795
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
@@ -5013,7 +4797,7 @@ _mm_mask_srav_epi64(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
5013
4797
|
(__v2di)__W);
|
|
5014
4798
|
}
|
|
5015
4799
|
|
|
5016
|
-
static __inline__ __m128i
|
|
4800
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5017
4801
|
_mm_maskz_srav_epi64(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
5018
4802
|
{
|
|
5019
4803
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
@@ -5021,13 +4805,13 @@ _mm_maskz_srav_epi64(__mmask8 __U, __m128i __X, __m128i __Y)
|
|
|
5021
4805
|
(__v2di)_mm_setzero_si128());
|
|
5022
4806
|
}
|
|
5023
4807
|
|
|
5024
|
-
static __inline__ __m256i
|
|
4808
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5025
4809
|
_mm256_srav_epi64(__m256i __X, __m256i __Y)
|
|
5026
4810
|
{
|
|
5027
4811
|
return (__m256i)__builtin_ia32_psravq256((__v4di)__X, (__v4di) __Y);
|
|
5028
4812
|
}
|
|
5029
4813
|
|
|
5030
|
-
static __inline__ __m256i
|
|
4814
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5031
4815
|
_mm256_mask_srav_epi64(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
5032
4816
|
{
|
|
5033
4817
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
@@ -5035,7 +4819,7 @@ _mm256_mask_srav_epi64(__m256i __W, __mmask8 __U, __m256i __X, __m256i __Y)
|
|
|
5035
4819
|
(__v4di)__W);
|
|
5036
4820
|
}
|
|
5037
4821
|
|
|
5038
|
-
static __inline__ __m256i
|
|
4822
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5039
4823
|
_mm256_maskz_srav_epi64 (__mmask8 __U, __m256i __X, __m256i __Y)
|
|
5040
4824
|
{
|
|
5041
4825
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
@@ -5296,69 +5080,55 @@ _mm256_maskz_movedup_pd (__mmask8 __U, __m256d __A)
|
|
|
5296
5080
|
(__v4df)_mm256_setzero_pd());
|
|
5297
5081
|
}
|
|
5298
5082
|
|
|
5299
|
-
static __inline__ __m128i
|
|
5300
|
-
_mm_mask_set1_epi32(__m128i __O, __mmask8 __M, int __A)
|
|
5301
|
-
|
|
5302
|
-
|
|
5303
|
-
(__v4si) _mm_set1_epi32(__A),
|
|
5304
|
-
(__v4si)__O);
|
|
5083
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5084
|
+
_mm_mask_set1_epi32(__m128i __O, __mmask8 __M, int __A) {
|
|
5085
|
+
return (__m128i)__builtin_ia32_selectd_128(__M, (__v4si)_mm_set1_epi32(__A),
|
|
5086
|
+
(__v4si)__O);
|
|
5305
5087
|
}
|
|
5306
5088
|
|
|
5307
|
-
static __inline__ __m128i
|
|
5308
|
-
_mm_maskz_set1_epi32(
|
|
5309
|
-
|
|
5310
|
-
|
|
5311
|
-
(__v4si) _mm_set1_epi32(__A),
|
|
5312
|
-
(__v4si)_mm_setzero_si128());
|
|
5089
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5090
|
+
_mm_maskz_set1_epi32(__mmask8 __M, int __A) {
|
|
5091
|
+
return (__m128i)__builtin_ia32_selectd_128(__M, (__v4si)_mm_set1_epi32(__A),
|
|
5092
|
+
(__v4si)_mm_setzero_si128());
|
|
5313
5093
|
}
|
|
5314
5094
|
|
|
5315
|
-
static __inline__ __m256i
|
|
5316
|
-
_mm256_mask_set1_epi32(__m256i __O, __mmask8 __M, int __A)
|
|
5317
|
-
|
|
5318
|
-
|
|
5319
|
-
(__v8si) _mm256_set1_epi32(__A),
|
|
5320
|
-
(__v8si)__O);
|
|
5095
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5096
|
+
_mm256_mask_set1_epi32(__m256i __O, __mmask8 __M, int __A) {
|
|
5097
|
+
return (__m256i)__builtin_ia32_selectd_256(
|
|
5098
|
+
__M, (__v8si)_mm256_set1_epi32(__A), (__v8si)__O);
|
|
5321
5099
|
}
|
|
5322
5100
|
|
|
5323
|
-
static __inline__ __m256i
|
|
5324
|
-
_mm256_maskz_set1_epi32(
|
|
5325
|
-
|
|
5326
|
-
|
|
5327
|
-
(__v8si) _mm256_set1_epi32(__A),
|
|
5328
|
-
(__v8si)_mm256_setzero_si256());
|
|
5101
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5102
|
+
_mm256_maskz_set1_epi32(__mmask8 __M, int __A) {
|
|
5103
|
+
return (__m256i)__builtin_ia32_selectd_256(
|
|
5104
|
+
__M, (__v8si)_mm256_set1_epi32(__A), (__v8si)_mm256_setzero_si256());
|
|
5329
5105
|
}
|
|
5330
5106
|
|
|
5331
|
-
|
|
5332
|
-
|
|
5333
|
-
_mm_mask_set1_epi64 (__m128i __O, __mmask8 __M, long long __A)
|
|
5334
|
-
{
|
|
5107
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5108
|
+
_mm_mask_set1_epi64(__m128i __O, __mmask8 __M, long long __A) {
|
|
5335
5109
|
return (__m128i) __builtin_ia32_selectq_128(__M,
|
|
5336
5110
|
(__v2di) _mm_set1_epi64x(__A),
|
|
5337
5111
|
(__v2di) __O);
|
|
5338
5112
|
}
|
|
5339
5113
|
|
|
5340
|
-
static __inline__ __m128i
|
|
5341
|
-
_mm_maskz_set1_epi64
|
|
5342
|
-
{
|
|
5114
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5115
|
+
_mm_maskz_set1_epi64(__mmask8 __M, long long __A) {
|
|
5343
5116
|
return (__m128i) __builtin_ia32_selectq_128(__M,
|
|
5344
5117
|
(__v2di) _mm_set1_epi64x(__A),
|
|
5345
5118
|
(__v2di) _mm_setzero_si128());
|
|
5346
5119
|
}
|
|
5347
5120
|
|
|
5348
|
-
static __inline__ __m256i
|
|
5349
|
-
_mm256_mask_set1_epi64
|
|
5350
|
-
{
|
|
5121
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5122
|
+
_mm256_mask_set1_epi64(__m256i __O, __mmask8 __M, long long __A) {
|
|
5351
5123
|
return (__m256i) __builtin_ia32_selectq_256(__M,
|
|
5352
5124
|
(__v4di) _mm256_set1_epi64x(__A),
|
|
5353
5125
|
(__v4di) __O) ;
|
|
5354
5126
|
}
|
|
5355
5127
|
|
|
5356
|
-
static __inline__ __m256i
|
|
5357
|
-
_mm256_maskz_set1_epi64
|
|
5358
|
-
|
|
5359
|
-
|
|
5360
|
-
(__v4di) _mm256_set1_epi64x(__A),
|
|
5361
|
-
(__v4di) _mm256_setzero_si256());
|
|
5128
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5129
|
+
_mm256_maskz_set1_epi64(__mmask8 __M, long long __A) {
|
|
5130
|
+
return (__m256i)__builtin_ia32_selectq_256(
|
|
5131
|
+
__M, (__v4di)_mm256_set1_epi64x(__A), (__v4di)_mm256_setzero_si256());
|
|
5362
5132
|
}
|
|
5363
5133
|
|
|
5364
5134
|
#define _mm_fixupimm_pd(A, B, C, imm) \
|
|
@@ -5805,130 +5575,113 @@ _mm256_mask_storeu_ps (void *__P, __mmask8 __U, __m256 __A)
|
|
|
5805
5575
|
(__mmask8) __U);
|
|
5806
5576
|
}
|
|
5807
5577
|
|
|
5808
|
-
|
|
5809
|
-
|
|
5810
|
-
_mm_mask_unpackhi_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
|
|
5811
|
-
{
|
|
5578
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5579
|
+
_mm_mask_unpackhi_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
|
|
5812
5580
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
5813
5581
|
(__v2df)_mm_unpackhi_pd(__A, __B),
|
|
5814
5582
|
(__v2df)__W);
|
|
5815
5583
|
}
|
|
5816
5584
|
|
|
5817
|
-
static __inline__ __m128d
|
|
5818
|
-
_mm_maskz_unpackhi_pd(__mmask8 __U, __m128d __A, __m128d __B)
|
|
5819
|
-
{
|
|
5585
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5586
|
+
_mm_maskz_unpackhi_pd(__mmask8 __U, __m128d __A, __m128d __B) {
|
|
5820
5587
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
5821
5588
|
(__v2df)_mm_unpackhi_pd(__A, __B),
|
|
5822
5589
|
(__v2df)_mm_setzero_pd());
|
|
5823
5590
|
}
|
|
5824
5591
|
|
|
5825
|
-
static __inline__ __m256d
|
|
5826
|
-
_mm256_mask_unpackhi_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B)
|
|
5827
|
-
{
|
|
5592
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5593
|
+
_mm256_mask_unpackhi_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
|
|
5828
5594
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
5829
5595
|
(__v4df)_mm256_unpackhi_pd(__A, __B),
|
|
5830
5596
|
(__v4df)__W);
|
|
5831
5597
|
}
|
|
5832
5598
|
|
|
5833
|
-
static __inline__ __m256d
|
|
5834
|
-
_mm256_maskz_unpackhi_pd(__mmask8 __U, __m256d __A, __m256d __B)
|
|
5835
|
-
{
|
|
5599
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5600
|
+
_mm256_maskz_unpackhi_pd(__mmask8 __U, __m256d __A, __m256d __B) {
|
|
5836
5601
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
5837
5602
|
(__v4df)_mm256_unpackhi_pd(__A, __B),
|
|
5838
5603
|
(__v4df)_mm256_setzero_pd());
|
|
5839
5604
|
}
|
|
5840
5605
|
|
|
5841
|
-
static __inline__ __m128
|
|
5842
|
-
_mm_mask_unpackhi_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
|
|
5843
|
-
{
|
|
5606
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5607
|
+
_mm_mask_unpackhi_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
|
|
5844
5608
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
5845
5609
|
(__v4sf)_mm_unpackhi_ps(__A, __B),
|
|
5846
5610
|
(__v4sf)__W);
|
|
5847
5611
|
}
|
|
5848
5612
|
|
|
5849
|
-
static __inline__ __m128
|
|
5850
|
-
_mm_maskz_unpackhi_ps(__mmask8 __U, __m128 __A, __m128 __B)
|
|
5851
|
-
{
|
|
5613
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5614
|
+
_mm_maskz_unpackhi_ps(__mmask8 __U, __m128 __A, __m128 __B) {
|
|
5852
5615
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
5853
5616
|
(__v4sf)_mm_unpackhi_ps(__A, __B),
|
|
5854
5617
|
(__v4sf)_mm_setzero_ps());
|
|
5855
5618
|
}
|
|
5856
5619
|
|
|
5857
|
-
static __inline__ __m256
|
|
5858
|
-
_mm256_mask_unpackhi_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
|
|
5859
|
-
{
|
|
5620
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5621
|
+
_mm256_mask_unpackhi_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
|
|
5860
5622
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
5861
5623
|
(__v8sf)_mm256_unpackhi_ps(__A, __B),
|
|
5862
5624
|
(__v8sf)__W);
|
|
5863
5625
|
}
|
|
5864
5626
|
|
|
5865
|
-
static __inline__ __m256
|
|
5866
|
-
_mm256_maskz_unpackhi_ps(__mmask8 __U, __m256 __A, __m256 __B)
|
|
5867
|
-
{
|
|
5627
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5628
|
+
_mm256_maskz_unpackhi_ps(__mmask8 __U, __m256 __A, __m256 __B) {
|
|
5868
5629
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
5869
5630
|
(__v8sf)_mm256_unpackhi_ps(__A, __B),
|
|
5870
5631
|
(__v8sf)_mm256_setzero_ps());
|
|
5871
5632
|
}
|
|
5872
5633
|
|
|
5873
|
-
static __inline__ __m128d
|
|
5874
|
-
_mm_mask_unpacklo_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B)
|
|
5875
|
-
{
|
|
5634
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5635
|
+
_mm_mask_unpacklo_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) {
|
|
5876
5636
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
5877
5637
|
(__v2df)_mm_unpacklo_pd(__A, __B),
|
|
5878
5638
|
(__v2df)__W);
|
|
5879
5639
|
}
|
|
5880
5640
|
|
|
5881
|
-
static __inline__ __m128d
|
|
5882
|
-
_mm_maskz_unpacklo_pd(__mmask8 __U, __m128d __A, __m128d __B)
|
|
5883
|
-
{
|
|
5641
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5642
|
+
_mm_maskz_unpacklo_pd(__mmask8 __U, __m128d __A, __m128d __B) {
|
|
5884
5643
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
5885
5644
|
(__v2df)_mm_unpacklo_pd(__A, __B),
|
|
5886
5645
|
(__v2df)_mm_setzero_pd());
|
|
5887
5646
|
}
|
|
5888
5647
|
|
|
5889
|
-
static __inline__ __m256d
|
|
5890
|
-
_mm256_mask_unpacklo_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B)
|
|
5891
|
-
{
|
|
5648
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5649
|
+
_mm256_mask_unpacklo_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) {
|
|
5892
5650
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
5893
5651
|
(__v4df)_mm256_unpacklo_pd(__A, __B),
|
|
5894
5652
|
(__v4df)__W);
|
|
5895
5653
|
}
|
|
5896
5654
|
|
|
5897
|
-
static __inline__ __m256d
|
|
5898
|
-
_mm256_maskz_unpacklo_pd(__mmask8 __U, __m256d __A, __m256d __B)
|
|
5899
|
-
{
|
|
5655
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5656
|
+
_mm256_maskz_unpacklo_pd(__mmask8 __U, __m256d __A, __m256d __B) {
|
|
5900
5657
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
5901
5658
|
(__v4df)_mm256_unpacklo_pd(__A, __B),
|
|
5902
5659
|
(__v4df)_mm256_setzero_pd());
|
|
5903
5660
|
}
|
|
5904
5661
|
|
|
5905
|
-
static __inline__ __m128
|
|
5906
|
-
_mm_mask_unpacklo_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B)
|
|
5907
|
-
{
|
|
5662
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5663
|
+
_mm_mask_unpacklo_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) {
|
|
5908
5664
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
5909
5665
|
(__v4sf)_mm_unpacklo_ps(__A, __B),
|
|
5910
5666
|
(__v4sf)__W);
|
|
5911
5667
|
}
|
|
5912
5668
|
|
|
5913
|
-
static __inline__ __m128
|
|
5914
|
-
_mm_maskz_unpacklo_ps(__mmask8 __U, __m128 __A, __m128 __B)
|
|
5915
|
-
{
|
|
5669
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5670
|
+
_mm_maskz_unpacklo_ps(__mmask8 __U, __m128 __A, __m128 __B) {
|
|
5916
5671
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
5917
5672
|
(__v4sf)_mm_unpacklo_ps(__A, __B),
|
|
5918
5673
|
(__v4sf)_mm_setzero_ps());
|
|
5919
5674
|
}
|
|
5920
5675
|
|
|
5921
|
-
static __inline__ __m256
|
|
5922
|
-
_mm256_mask_unpacklo_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B)
|
|
5923
|
-
{
|
|
5676
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5677
|
+
_mm256_mask_unpacklo_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) {
|
|
5924
5678
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
5925
5679
|
(__v8sf)_mm256_unpacklo_ps(__A, __B),
|
|
5926
5680
|
(__v8sf)__W);
|
|
5927
5681
|
}
|
|
5928
5682
|
|
|
5929
|
-
static __inline__ __m256
|
|
5930
|
-
_mm256_maskz_unpacklo_ps(__mmask8 __U, __m256 __A, __m256 __B)
|
|
5931
|
-
{
|
|
5683
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5684
|
+
_mm256_maskz_unpacklo_ps(__mmask8 __U, __m256 __A, __m256 __B) {
|
|
5932
5685
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
5933
5686
|
(__v8sf)_mm256_unpacklo_ps(__A, __B),
|
|
5934
5687
|
(__v8sf)_mm256_setzero_ps());
|
|
@@ -6078,65 +5831,57 @@ _mm256_maskz_rcp14_ps (__mmask8 __U, __m256 __A)
|
|
|
6078
5831
|
(__v8sf)_mm256_permute_ps((X), (C)), \
|
|
6079
5832
|
(__v8sf)_mm256_setzero_ps()))
|
|
6080
5833
|
|
|
6081
|
-
static __inline__ __m128d
|
|
6082
|
-
_mm_mask_permutevar_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128i __C)
|
|
6083
|
-
{
|
|
5834
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5835
|
+
_mm_mask_permutevar_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128i __C) {
|
|
6084
5836
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
6085
5837
|
(__v2df)_mm_permutevar_pd(__A, __C),
|
|
6086
5838
|
(__v2df)__W);
|
|
6087
5839
|
}
|
|
6088
5840
|
|
|
6089
|
-
static __inline__ __m128d
|
|
6090
|
-
_mm_maskz_permutevar_pd(__mmask8 __U, __m128d __A, __m128i __C)
|
|
6091
|
-
{
|
|
5841
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5842
|
+
_mm_maskz_permutevar_pd(__mmask8 __U, __m128d __A, __m128i __C) {
|
|
6092
5843
|
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U,
|
|
6093
5844
|
(__v2df)_mm_permutevar_pd(__A, __C),
|
|
6094
5845
|
(__v2df)_mm_setzero_pd());
|
|
6095
5846
|
}
|
|
6096
5847
|
|
|
6097
|
-
static __inline__ __m256d
|
|
6098
|
-
_mm256_mask_permutevar_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256i __C)
|
|
6099
|
-
{
|
|
5848
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5849
|
+
_mm256_mask_permutevar_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256i __C) {
|
|
6100
5850
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
6101
5851
|
(__v4df)_mm256_permutevar_pd(__A, __C),
|
|
6102
5852
|
(__v4df)__W);
|
|
6103
5853
|
}
|
|
6104
5854
|
|
|
6105
|
-
static __inline__ __m256d
|
|
6106
|
-
_mm256_maskz_permutevar_pd(__mmask8 __U, __m256d __A, __m256i __C)
|
|
6107
|
-
{
|
|
5855
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5856
|
+
_mm256_maskz_permutevar_pd(__mmask8 __U, __m256d __A, __m256i __C) {
|
|
6108
5857
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
6109
5858
|
(__v4df)_mm256_permutevar_pd(__A, __C),
|
|
6110
5859
|
(__v4df)_mm256_setzero_pd());
|
|
6111
5860
|
}
|
|
6112
5861
|
|
|
6113
|
-
static __inline__ __m128
|
|
6114
|
-
_mm_mask_permutevar_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128i __C)
|
|
6115
|
-
{
|
|
5862
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5863
|
+
_mm_mask_permutevar_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128i __C) {
|
|
6116
5864
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
6117
5865
|
(__v4sf)_mm_permutevar_ps(__A, __C),
|
|
6118
5866
|
(__v4sf)__W);
|
|
6119
5867
|
}
|
|
6120
5868
|
|
|
6121
|
-
static __inline__ __m128
|
|
6122
|
-
_mm_maskz_permutevar_ps(__mmask8 __U, __m128 __A, __m128i __C)
|
|
6123
|
-
{
|
|
5869
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5870
|
+
_mm_maskz_permutevar_ps(__mmask8 __U, __m128 __A, __m128i __C) {
|
|
6124
5871
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
6125
5872
|
(__v4sf)_mm_permutevar_ps(__A, __C),
|
|
6126
5873
|
(__v4sf)_mm_setzero_ps());
|
|
6127
5874
|
}
|
|
6128
5875
|
|
|
6129
|
-
static __inline__ __m256
|
|
6130
|
-
_mm256_mask_permutevar_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256i __C)
|
|
6131
|
-
{
|
|
5876
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5877
|
+
_mm256_mask_permutevar_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256i __C) {
|
|
6132
5878
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
6133
5879
|
(__v8sf)_mm256_permutevar_ps(__A, __C),
|
|
6134
5880
|
(__v8sf)__W);
|
|
6135
5881
|
}
|
|
6136
5882
|
|
|
6137
|
-
static __inline__ __m256
|
|
6138
|
-
_mm256_maskz_permutevar_ps(__mmask8 __U, __m256 __A, __m256i __C)
|
|
6139
|
-
{
|
|
5883
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
5884
|
+
_mm256_maskz_permutevar_ps(__mmask8 __U, __m256 __A, __m256i __C) {
|
|
6140
5885
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
6141
5886
|
(__v8sf)_mm256_permutevar_ps(__A, __C),
|
|
6142
5887
|
(__v8sf)_mm256_setzero_ps());
|
|
@@ -6250,282 +5995,251 @@ _mm256_mask_testn_epi64_mask (__mmask8 __U, __m256i __A, __m256i __B)
|
|
|
6250
5995
|
_mm256_setzero_si256());
|
|
6251
5996
|
}
|
|
6252
5997
|
|
|
6253
|
-
static __inline__ __m128i
|
|
6254
|
-
_mm_mask_unpackhi_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
6255
|
-
{
|
|
5998
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
5999
|
+
_mm_mask_unpackhi_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
6256
6000
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
6257
6001
|
(__v4si)_mm_unpackhi_epi32(__A, __B),
|
|
6258
6002
|
(__v4si)__W);
|
|
6259
6003
|
}
|
|
6260
6004
|
|
|
6261
|
-
static __inline__ __m128i
|
|
6262
|
-
_mm_maskz_unpackhi_epi32(__mmask8 __U, __m128i __A, __m128i __B)
|
|
6263
|
-
{
|
|
6005
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6006
|
+
_mm_maskz_unpackhi_epi32(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
6264
6007
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
6265
6008
|
(__v4si)_mm_unpackhi_epi32(__A, __B),
|
|
6266
6009
|
(__v4si)_mm_setzero_si128());
|
|
6267
6010
|
}
|
|
6268
6011
|
|
|
6269
|
-
static __inline__ __m256i
|
|
6270
|
-
_mm256_mask_unpackhi_epi32(__m256i __W, __mmask8 __U, __m256i __A,
|
|
6271
|
-
{
|
|
6012
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6013
|
+
_mm256_mask_unpackhi_epi32(__m256i __W, __mmask8 __U, __m256i __A,
|
|
6014
|
+
__m256i __B) {
|
|
6272
6015
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
6273
6016
|
(__v8si)_mm256_unpackhi_epi32(__A, __B),
|
|
6274
6017
|
(__v8si)__W);
|
|
6275
6018
|
}
|
|
6276
6019
|
|
|
6277
|
-
static __inline__ __m256i
|
|
6278
|
-
_mm256_maskz_unpackhi_epi32(__mmask8 __U, __m256i __A, __m256i __B)
|
|
6279
|
-
{
|
|
6020
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6021
|
+
_mm256_maskz_unpackhi_epi32(__mmask8 __U, __m256i __A, __m256i __B) {
|
|
6280
6022
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
6281
6023
|
(__v8si)_mm256_unpackhi_epi32(__A, __B),
|
|
6282
6024
|
(__v8si)_mm256_setzero_si256());
|
|
6283
6025
|
}
|
|
6284
6026
|
|
|
6285
|
-
static __inline__ __m128i
|
|
6286
|
-
_mm_mask_unpackhi_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
6287
|
-
{
|
|
6027
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6028
|
+
_mm_mask_unpackhi_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
6288
6029
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
6289
6030
|
(__v2di)_mm_unpackhi_epi64(__A, __B),
|
|
6290
6031
|
(__v2di)__W);
|
|
6291
6032
|
}
|
|
6292
6033
|
|
|
6293
|
-
static __inline__ __m128i
|
|
6294
|
-
_mm_maskz_unpackhi_epi64(__mmask8 __U, __m128i __A, __m128i __B)
|
|
6295
|
-
{
|
|
6034
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6035
|
+
_mm_maskz_unpackhi_epi64(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
6296
6036
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
6297
6037
|
(__v2di)_mm_unpackhi_epi64(__A, __B),
|
|
6298
6038
|
(__v2di)_mm_setzero_si128());
|
|
6299
6039
|
}
|
|
6300
6040
|
|
|
6301
|
-
static __inline__ __m256i
|
|
6302
|
-
_mm256_mask_unpackhi_epi64(__m256i __W, __mmask8 __U, __m256i __A,
|
|
6303
|
-
{
|
|
6041
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6042
|
+
_mm256_mask_unpackhi_epi64(__m256i __W, __mmask8 __U, __m256i __A,
|
|
6043
|
+
__m256i __B) {
|
|
6304
6044
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
6305
6045
|
(__v4di)_mm256_unpackhi_epi64(__A, __B),
|
|
6306
6046
|
(__v4di)__W);
|
|
6307
6047
|
}
|
|
6308
6048
|
|
|
6309
|
-
static __inline__ __m256i
|
|
6310
|
-
_mm256_maskz_unpackhi_epi64(__mmask8 __U, __m256i __A, __m256i __B)
|
|
6311
|
-
{
|
|
6049
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6050
|
+
_mm256_maskz_unpackhi_epi64(__mmask8 __U, __m256i __A, __m256i __B) {
|
|
6312
6051
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
6313
6052
|
(__v4di)_mm256_unpackhi_epi64(__A, __B),
|
|
6314
6053
|
(__v4di)_mm256_setzero_si256());
|
|
6315
6054
|
}
|
|
6316
6055
|
|
|
6317
|
-
static __inline__ __m128i
|
|
6318
|
-
_mm_mask_unpacklo_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
6319
|
-
{
|
|
6056
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6057
|
+
_mm_mask_unpacklo_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
6320
6058
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
6321
6059
|
(__v4si)_mm_unpacklo_epi32(__A, __B),
|
|
6322
6060
|
(__v4si)__W);
|
|
6323
6061
|
}
|
|
6324
6062
|
|
|
6325
|
-
static __inline__ __m128i
|
|
6326
|
-
_mm_maskz_unpacklo_epi32(__mmask8 __U, __m128i __A, __m128i __B)
|
|
6327
|
-
{
|
|
6063
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6064
|
+
_mm_maskz_unpacklo_epi32(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
6328
6065
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
6329
6066
|
(__v4si)_mm_unpacklo_epi32(__A, __B),
|
|
6330
6067
|
(__v4si)_mm_setzero_si128());
|
|
6331
6068
|
}
|
|
6332
6069
|
|
|
6333
|
-
static __inline__ __m256i
|
|
6334
|
-
_mm256_mask_unpacklo_epi32(__m256i __W, __mmask8 __U, __m256i __A,
|
|
6335
|
-
{
|
|
6070
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6071
|
+
_mm256_mask_unpacklo_epi32(__m256i __W, __mmask8 __U, __m256i __A,
|
|
6072
|
+
__m256i __B) {
|
|
6336
6073
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
6337
6074
|
(__v8si)_mm256_unpacklo_epi32(__A, __B),
|
|
6338
6075
|
(__v8si)__W);
|
|
6339
6076
|
}
|
|
6340
6077
|
|
|
6341
|
-
static __inline__ __m256i
|
|
6342
|
-
_mm256_maskz_unpacklo_epi32(__mmask8 __U, __m256i __A, __m256i __B)
|
|
6343
|
-
{
|
|
6078
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6079
|
+
_mm256_maskz_unpacklo_epi32(__mmask8 __U, __m256i __A, __m256i __B) {
|
|
6344
6080
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
6345
6081
|
(__v8si)_mm256_unpacklo_epi32(__A, __B),
|
|
6346
6082
|
(__v8si)_mm256_setzero_si256());
|
|
6347
6083
|
}
|
|
6348
6084
|
|
|
6349
|
-
static __inline__ __m128i
|
|
6350
|
-
_mm_mask_unpacklo_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
6351
|
-
{
|
|
6085
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6086
|
+
_mm_mask_unpacklo_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
6352
6087
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
6353
6088
|
(__v2di)_mm_unpacklo_epi64(__A, __B),
|
|
6354
6089
|
(__v2di)__W);
|
|
6355
6090
|
}
|
|
6356
6091
|
|
|
6357
|
-
static __inline__ __m128i
|
|
6358
|
-
_mm_maskz_unpacklo_epi64(__mmask8 __U, __m128i __A, __m128i __B)
|
|
6359
|
-
{
|
|
6092
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6093
|
+
_mm_maskz_unpacklo_epi64(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
6360
6094
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,
|
|
6361
6095
|
(__v2di)_mm_unpacklo_epi64(__A, __B),
|
|
6362
6096
|
(__v2di)_mm_setzero_si128());
|
|
6363
6097
|
}
|
|
6364
6098
|
|
|
6365
|
-
static __inline__ __m256i
|
|
6366
|
-
_mm256_mask_unpacklo_epi64(__m256i __W, __mmask8 __U, __m256i __A,
|
|
6367
|
-
{
|
|
6099
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6100
|
+
_mm256_mask_unpacklo_epi64(__m256i __W, __mmask8 __U, __m256i __A,
|
|
6101
|
+
__m256i __B) {
|
|
6368
6102
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
6369
6103
|
(__v4di)_mm256_unpacklo_epi64(__A, __B),
|
|
6370
6104
|
(__v4di)__W);
|
|
6371
6105
|
}
|
|
6372
6106
|
|
|
6373
|
-
static __inline__ __m256i
|
|
6374
|
-
_mm256_maskz_unpacklo_epi64(__mmask8 __U, __m256i __A, __m256i __B)
|
|
6375
|
-
{
|
|
6107
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6108
|
+
_mm256_maskz_unpacklo_epi64(__mmask8 __U, __m256i __A, __m256i __B) {
|
|
6376
6109
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,
|
|
6377
6110
|
(__v4di)_mm256_unpacklo_epi64(__A, __B),
|
|
6378
6111
|
(__v4di)_mm256_setzero_si256());
|
|
6379
6112
|
}
|
|
6380
6113
|
|
|
6381
|
-
static __inline__ __m128i
|
|
6382
|
-
_mm_mask_sra_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
6383
|
-
{
|
|
6114
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6115
|
+
_mm_mask_sra_epi32(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
6384
6116
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
6385
6117
|
(__v4si)_mm_sra_epi32(__A, __B),
|
|
6386
6118
|
(__v4si)__W);
|
|
6387
6119
|
}
|
|
6388
6120
|
|
|
6389
|
-
static __inline__ __m128i
|
|
6390
|
-
_mm_maskz_sra_epi32(__mmask8 __U, __m128i __A, __m128i __B)
|
|
6391
|
-
{
|
|
6121
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6122
|
+
_mm_maskz_sra_epi32(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
6392
6123
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
6393
6124
|
(__v4si)_mm_sra_epi32(__A, __B),
|
|
6394
6125
|
(__v4si)_mm_setzero_si128());
|
|
6395
6126
|
}
|
|
6396
6127
|
|
|
6397
|
-
static __inline__ __m256i
|
|
6398
|
-
_mm256_mask_sra_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B)
|
|
6399
|
-
{
|
|
6128
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6129
|
+
_mm256_mask_sra_epi32(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) {
|
|
6400
6130
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
6401
6131
|
(__v8si)_mm256_sra_epi32(__A, __B),
|
|
6402
6132
|
(__v8si)__W);
|
|
6403
6133
|
}
|
|
6404
6134
|
|
|
6405
|
-
static __inline__ __m256i
|
|
6406
|
-
_mm256_maskz_sra_epi32(__mmask8 __U, __m256i __A, __m128i __B)
|
|
6407
|
-
{
|
|
6135
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6136
|
+
_mm256_maskz_sra_epi32(__mmask8 __U, __m256i __A, __m128i __B) {
|
|
6408
6137
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
6409
6138
|
(__v8si)_mm256_sra_epi32(__A, __B),
|
|
6410
6139
|
(__v8si)_mm256_setzero_si256());
|
|
6411
6140
|
}
|
|
6412
6141
|
|
|
6413
|
-
static __inline__ __m128i
|
|
6414
|
-
_mm_mask_srai_epi32(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B)
|
|
6415
|
-
{
|
|
6142
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6143
|
+
_mm_mask_srai_epi32(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B) {
|
|
6416
6144
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
6417
6145
|
(__v4si)_mm_srai_epi32(__A, (int)__B),
|
|
6418
6146
|
(__v4si)__W);
|
|
6419
6147
|
}
|
|
6420
6148
|
|
|
6421
|
-
static __inline__ __m128i
|
|
6422
|
-
_mm_maskz_srai_epi32(__mmask8 __U, __m128i __A, unsigned int __B)
|
|
6423
|
-
{
|
|
6149
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6150
|
+
_mm_maskz_srai_epi32(__mmask8 __U, __m128i __A, unsigned int __B) {
|
|
6424
6151
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,
|
|
6425
6152
|
(__v4si)_mm_srai_epi32(__A, (int)__B),
|
|
6426
6153
|
(__v4si)_mm_setzero_si128());
|
|
6427
6154
|
}
|
|
6428
6155
|
|
|
6429
|
-
static __inline__ __m256i
|
|
6430
|
-
_mm256_mask_srai_epi32(__m256i __W, __mmask8 __U, __m256i __A,
|
|
6431
|
-
{
|
|
6156
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6157
|
+
_mm256_mask_srai_epi32(__m256i __W, __mmask8 __U, __m256i __A,
|
|
6158
|
+
unsigned int __B) {
|
|
6432
6159
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
6433
6160
|
(__v8si)_mm256_srai_epi32(__A, (int)__B),
|
|
6434
6161
|
(__v8si)__W);
|
|
6435
6162
|
}
|
|
6436
6163
|
|
|
6437
|
-
static __inline__ __m256i
|
|
6438
|
-
_mm256_maskz_srai_epi32(__mmask8 __U, __m256i __A, unsigned int __B)
|
|
6439
|
-
{
|
|
6164
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6165
|
+
_mm256_maskz_srai_epi32(__mmask8 __U, __m256i __A, unsigned int __B) {
|
|
6440
6166
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,
|
|
6441
6167
|
(__v8si)_mm256_srai_epi32(__A, (int)__B),
|
|
6442
6168
|
(__v8si)_mm256_setzero_si256());
|
|
6443
6169
|
}
|
|
6444
6170
|
|
|
6445
|
-
static __inline__ __m128i
|
|
6446
|
-
_mm_sra_epi64(__m128i __A, __m128i __B)
|
|
6447
|
-
{
|
|
6171
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6172
|
+
_mm_sra_epi64(__m128i __A, __m128i __B) {
|
|
6448
6173
|
return (__m128i)__builtin_ia32_psraq128((__v2di)__A, (__v2di)__B);
|
|
6449
6174
|
}
|
|
6450
6175
|
|
|
6451
|
-
static __inline__ __m128i
|
|
6452
|
-
_mm_mask_sra_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B)
|
|
6453
|
-
{
|
|
6176
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6177
|
+
_mm_mask_sra_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) {
|
|
6454
6178
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, \
|
|
6455
6179
|
(__v2di)_mm_sra_epi64(__A, __B), \
|
|
6456
6180
|
(__v2di)__W);
|
|
6457
6181
|
}
|
|
6458
6182
|
|
|
6459
|
-
static __inline__ __m128i
|
|
6460
|
-
_mm_maskz_sra_epi64(__mmask8 __U, __m128i __A, __m128i __B)
|
|
6461
|
-
{
|
|
6183
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6184
|
+
_mm_maskz_sra_epi64(__mmask8 __U, __m128i __A, __m128i __B) {
|
|
6462
6185
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, \
|
|
6463
6186
|
(__v2di)_mm_sra_epi64(__A, __B), \
|
|
6464
6187
|
(__v2di)_mm_setzero_si128());
|
|
6465
6188
|
}
|
|
6466
6189
|
|
|
6467
|
-
static __inline__ __m256i
|
|
6468
|
-
_mm256_sra_epi64(__m256i __A, __m128i __B)
|
|
6469
|
-
{
|
|
6190
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6191
|
+
_mm256_sra_epi64(__m256i __A, __m128i __B) {
|
|
6470
6192
|
return (__m256i)__builtin_ia32_psraq256((__v4di) __A, (__v2di) __B);
|
|
6471
6193
|
}
|
|
6472
6194
|
|
|
6473
|
-
static __inline__ __m256i
|
|
6474
|
-
_mm256_mask_sra_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B)
|
|
6475
|
-
{
|
|
6195
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6196
|
+
_mm256_mask_sra_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m128i __B) {
|
|
6476
6197
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, \
|
|
6477
6198
|
(__v4di)_mm256_sra_epi64(__A, __B), \
|
|
6478
6199
|
(__v4di)__W);
|
|
6479
6200
|
}
|
|
6480
6201
|
|
|
6481
|
-
static __inline__ __m256i
|
|
6482
|
-
_mm256_maskz_sra_epi64(__mmask8 __U, __m256i __A, __m128i __B)
|
|
6483
|
-
{
|
|
6202
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6203
|
+
_mm256_maskz_sra_epi64(__mmask8 __U, __m256i __A, __m128i __B) {
|
|
6484
6204
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, \
|
|
6485
6205
|
(__v4di)_mm256_sra_epi64(__A, __B), \
|
|
6486
6206
|
(__v4di)_mm256_setzero_si256());
|
|
6487
6207
|
}
|
|
6488
6208
|
|
|
6489
|
-
static __inline__ __m128i
|
|
6490
|
-
_mm_srai_epi64(__m128i __A, unsigned int __imm)
|
|
6491
|
-
{
|
|
6209
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6210
|
+
_mm_srai_epi64(__m128i __A, unsigned int __imm) {
|
|
6492
6211
|
return (__m128i)__builtin_ia32_psraqi128((__v2di)__A, (int)__imm);
|
|
6493
6212
|
}
|
|
6494
6213
|
|
|
6495
|
-
static __inline__ __m128i
|
|
6496
|
-
|
|
6497
|
-
{
|
|
6214
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR _mm_mask_srai_epi64(
|
|
6215
|
+
__m128i __W, __mmask8 __U, __m128i __A, unsigned int __imm) {
|
|
6498
6216
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, \
|
|
6499
6217
|
(__v2di)_mm_srai_epi64(__A, __imm), \
|
|
6500
6218
|
(__v2di)__W);
|
|
6501
6219
|
}
|
|
6502
6220
|
|
|
6503
|
-
static __inline__ __m128i
|
|
6504
|
-
_mm_maskz_srai_epi64(__mmask8 __U, __m128i __A, unsigned int __imm)
|
|
6505
|
-
{
|
|
6221
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6222
|
+
_mm_maskz_srai_epi64(__mmask8 __U, __m128i __A, unsigned int __imm) {
|
|
6506
6223
|
return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, \
|
|
6507
6224
|
(__v2di)_mm_srai_epi64(__A, __imm), \
|
|
6508
6225
|
(__v2di)_mm_setzero_si128());
|
|
6509
6226
|
}
|
|
6510
6227
|
|
|
6511
|
-
static __inline__ __m256i
|
|
6512
|
-
_mm256_srai_epi64(__m256i __A, unsigned int __imm)
|
|
6513
|
-
{
|
|
6228
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6229
|
+
_mm256_srai_epi64(__m256i __A, unsigned int __imm) {
|
|
6514
6230
|
return (__m256i)__builtin_ia32_psraqi256((__v4di)__A, (int)__imm);
|
|
6515
6231
|
}
|
|
6516
6232
|
|
|
6517
|
-
static __inline__ __m256i
|
|
6233
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6518
6234
|
_mm256_mask_srai_epi64(__m256i __W, __mmask8 __U, __m256i __A,
|
|
6519
|
-
unsigned int __imm)
|
|
6520
|
-
{
|
|
6235
|
+
unsigned int __imm) {
|
|
6521
6236
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, \
|
|
6522
6237
|
(__v4di)_mm256_srai_epi64(__A, __imm), \
|
|
6523
6238
|
(__v4di)__W);
|
|
6524
6239
|
}
|
|
6525
6240
|
|
|
6526
|
-
static __inline__ __m256i
|
|
6527
|
-
_mm256_maskz_srai_epi64(__mmask8 __U, __m256i __A, unsigned int __imm)
|
|
6528
|
-
{
|
|
6241
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6242
|
+
_mm256_maskz_srai_epi64(__mmask8 __U, __m256i __A, unsigned int __imm) {
|
|
6529
6243
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, \
|
|
6530
6244
|
(__v4di)_mm256_srai_epi64(__A, __imm), \
|
|
6531
6245
|
(__v4di)_mm256_setzero_si256());
|
|
@@ -6792,159 +6506,139 @@ _mm256_maskz_rsqrt14_ps (__mmask8 __U, __m256 __A)
|
|
|
6792
6506
|
(__mmask8) __U);
|
|
6793
6507
|
}
|
|
6794
6508
|
|
|
6795
|
-
static __inline__ __m256
|
|
6796
|
-
_mm256_broadcast_f32x4(__m128 __A)
|
|
6797
|
-
{
|
|
6509
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6510
|
+
_mm256_broadcast_f32x4(__m128 __A) {
|
|
6798
6511
|
return (__m256)__builtin_shufflevector((__v4sf)__A, (__v4sf)__A,
|
|
6799
6512
|
0, 1, 2, 3, 0, 1, 2, 3);
|
|
6800
6513
|
}
|
|
6801
6514
|
|
|
6802
|
-
static __inline__ __m256
|
|
6803
|
-
_mm256_mask_broadcast_f32x4(__m256 __O, __mmask8 __M, __m128 __A)
|
|
6804
|
-
{
|
|
6515
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6516
|
+
_mm256_mask_broadcast_f32x4(__m256 __O, __mmask8 __M, __m128 __A) {
|
|
6805
6517
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__M,
|
|
6806
6518
|
(__v8sf)_mm256_broadcast_f32x4(__A),
|
|
6807
6519
|
(__v8sf)__O);
|
|
6808
6520
|
}
|
|
6809
6521
|
|
|
6810
|
-
static __inline__ __m256
|
|
6811
|
-
_mm256_maskz_broadcast_f32x4
|
|
6812
|
-
{
|
|
6522
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6523
|
+
_mm256_maskz_broadcast_f32x4(__mmask8 __M, __m128 __A) {
|
|
6813
6524
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__M,
|
|
6814
6525
|
(__v8sf)_mm256_broadcast_f32x4(__A),
|
|
6815
6526
|
(__v8sf)_mm256_setzero_ps());
|
|
6816
6527
|
}
|
|
6817
6528
|
|
|
6818
|
-
static __inline__ __m256i
|
|
6819
|
-
_mm256_broadcast_i32x4(__m128i __A)
|
|
6820
|
-
{
|
|
6529
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6530
|
+
_mm256_broadcast_i32x4(__m128i __A) {
|
|
6821
6531
|
return (__m256i)__builtin_shufflevector((__v4si)__A, (__v4si)__A,
|
|
6822
6532
|
0, 1, 2, 3, 0, 1, 2, 3);
|
|
6823
6533
|
}
|
|
6824
6534
|
|
|
6825
|
-
static __inline__ __m256i
|
|
6826
|
-
_mm256_mask_broadcast_i32x4(__m256i __O, __mmask8 __M, __m128i __A)
|
|
6827
|
-
{
|
|
6535
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6536
|
+
_mm256_mask_broadcast_i32x4(__m256i __O, __mmask8 __M, __m128i __A) {
|
|
6828
6537
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
6829
6538
|
(__v8si)_mm256_broadcast_i32x4(__A),
|
|
6830
6539
|
(__v8si)__O);
|
|
6831
6540
|
}
|
|
6832
6541
|
|
|
6833
|
-
static __inline__ __m256i
|
|
6834
|
-
_mm256_maskz_broadcast_i32x4(__mmask8 __M, __m128i __A)
|
|
6835
|
-
{
|
|
6542
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6543
|
+
_mm256_maskz_broadcast_i32x4(__mmask8 __M, __m128i __A) {
|
|
6836
6544
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
6837
6545
|
(__v8si)_mm256_broadcast_i32x4(__A),
|
|
6838
6546
|
(__v8si)_mm256_setzero_si256());
|
|
6839
6547
|
}
|
|
6840
6548
|
|
|
6841
|
-
static __inline__ __m256d
|
|
6842
|
-
_mm256_mask_broadcastsd_pd
|
|
6843
|
-
{
|
|
6549
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6550
|
+
_mm256_mask_broadcastsd_pd(__m256d __O, __mmask8 __M, __m128d __A) {
|
|
6844
6551
|
return (__m256d)__builtin_ia32_selectpd_256(__M,
|
|
6845
6552
|
(__v4df) _mm256_broadcastsd_pd(__A),
|
|
6846
6553
|
(__v4df) __O);
|
|
6847
6554
|
}
|
|
6848
6555
|
|
|
6849
|
-
static __inline__ __m256d
|
|
6850
|
-
_mm256_maskz_broadcastsd_pd
|
|
6851
|
-
{
|
|
6556
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6557
|
+
_mm256_maskz_broadcastsd_pd(__mmask8 __M, __m128d __A) {
|
|
6852
6558
|
return (__m256d)__builtin_ia32_selectpd_256(__M,
|
|
6853
6559
|
(__v4df) _mm256_broadcastsd_pd(__A),
|
|
6854
6560
|
(__v4df) _mm256_setzero_pd());
|
|
6855
6561
|
}
|
|
6856
6562
|
|
|
6857
|
-
static __inline__ __m128
|
|
6858
|
-
_mm_mask_broadcastss_ps
|
|
6859
|
-
{
|
|
6563
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6564
|
+
_mm_mask_broadcastss_ps(__m128 __O, __mmask8 __M, __m128 __A) {
|
|
6860
6565
|
return (__m128)__builtin_ia32_selectps_128(__M,
|
|
6861
6566
|
(__v4sf) _mm_broadcastss_ps(__A),
|
|
6862
6567
|
(__v4sf) __O);
|
|
6863
6568
|
}
|
|
6864
6569
|
|
|
6865
|
-
static __inline__ __m128
|
|
6866
|
-
_mm_maskz_broadcastss_ps
|
|
6867
|
-
{
|
|
6570
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6571
|
+
_mm_maskz_broadcastss_ps(__mmask8 __M, __m128 __A) {
|
|
6868
6572
|
return (__m128)__builtin_ia32_selectps_128(__M,
|
|
6869
6573
|
(__v4sf) _mm_broadcastss_ps(__A),
|
|
6870
6574
|
(__v4sf) _mm_setzero_ps());
|
|
6871
6575
|
}
|
|
6872
6576
|
|
|
6873
|
-
static __inline__ __m256
|
|
6874
|
-
_mm256_mask_broadcastss_ps
|
|
6875
|
-
{
|
|
6577
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6578
|
+
_mm256_mask_broadcastss_ps(__m256 __O, __mmask8 __M, __m128 __A) {
|
|
6876
6579
|
return (__m256)__builtin_ia32_selectps_256(__M,
|
|
6877
6580
|
(__v8sf) _mm256_broadcastss_ps(__A),
|
|
6878
6581
|
(__v8sf) __O);
|
|
6879
6582
|
}
|
|
6880
6583
|
|
|
6881
|
-
static __inline__ __m256
|
|
6882
|
-
_mm256_maskz_broadcastss_ps
|
|
6883
|
-
{
|
|
6584
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6585
|
+
_mm256_maskz_broadcastss_ps(__mmask8 __M, __m128 __A) {
|
|
6884
6586
|
return (__m256)__builtin_ia32_selectps_256(__M,
|
|
6885
6587
|
(__v8sf) _mm256_broadcastss_ps(__A),
|
|
6886
6588
|
(__v8sf) _mm256_setzero_ps());
|
|
6887
6589
|
}
|
|
6888
6590
|
|
|
6889
|
-
static __inline__ __m128i
|
|
6890
|
-
_mm_mask_broadcastd_epi32
|
|
6891
|
-
{
|
|
6591
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6592
|
+
_mm_mask_broadcastd_epi32(__m128i __O, __mmask8 __M, __m128i __A) {
|
|
6892
6593
|
return (__m128i)__builtin_ia32_selectd_128(__M,
|
|
6893
6594
|
(__v4si) _mm_broadcastd_epi32(__A),
|
|
6894
6595
|
(__v4si) __O);
|
|
6895
6596
|
}
|
|
6896
6597
|
|
|
6897
|
-
static __inline__ __m128i
|
|
6898
|
-
_mm_maskz_broadcastd_epi32
|
|
6899
|
-
{
|
|
6598
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6599
|
+
_mm_maskz_broadcastd_epi32(__mmask8 __M, __m128i __A) {
|
|
6900
6600
|
return (__m128i)__builtin_ia32_selectd_128(__M,
|
|
6901
6601
|
(__v4si) _mm_broadcastd_epi32(__A),
|
|
6902
6602
|
(__v4si) _mm_setzero_si128());
|
|
6903
6603
|
}
|
|
6904
6604
|
|
|
6905
|
-
static __inline__ __m256i
|
|
6906
|
-
_mm256_mask_broadcastd_epi32
|
|
6907
|
-
{
|
|
6605
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6606
|
+
_mm256_mask_broadcastd_epi32(__m256i __O, __mmask8 __M, __m128i __A) {
|
|
6908
6607
|
return (__m256i)__builtin_ia32_selectd_256(__M,
|
|
6909
6608
|
(__v8si) _mm256_broadcastd_epi32(__A),
|
|
6910
6609
|
(__v8si) __O);
|
|
6911
6610
|
}
|
|
6912
6611
|
|
|
6913
|
-
static __inline__ __m256i
|
|
6914
|
-
_mm256_maskz_broadcastd_epi32
|
|
6915
|
-
{
|
|
6612
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6613
|
+
_mm256_maskz_broadcastd_epi32(__mmask8 __M, __m128i __A) {
|
|
6916
6614
|
return (__m256i)__builtin_ia32_selectd_256(__M,
|
|
6917
6615
|
(__v8si) _mm256_broadcastd_epi32(__A),
|
|
6918
6616
|
(__v8si) _mm256_setzero_si256());
|
|
6919
6617
|
}
|
|
6920
6618
|
|
|
6921
|
-
static __inline__ __m128i
|
|
6922
|
-
_mm_mask_broadcastq_epi64
|
|
6923
|
-
{
|
|
6619
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6620
|
+
_mm_mask_broadcastq_epi64(__m128i __O, __mmask8 __M, __m128i __A) {
|
|
6924
6621
|
return (__m128i)__builtin_ia32_selectq_128(__M,
|
|
6925
6622
|
(__v2di) _mm_broadcastq_epi64(__A),
|
|
6926
6623
|
(__v2di) __O);
|
|
6927
6624
|
}
|
|
6928
6625
|
|
|
6929
|
-
static __inline__ __m128i
|
|
6930
|
-
_mm_maskz_broadcastq_epi64
|
|
6931
|
-
{
|
|
6626
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
6627
|
+
_mm_maskz_broadcastq_epi64(__mmask8 __M, __m128i __A) {
|
|
6932
6628
|
return (__m128i)__builtin_ia32_selectq_128(__M,
|
|
6933
6629
|
(__v2di) _mm_broadcastq_epi64(__A),
|
|
6934
6630
|
(__v2di) _mm_setzero_si128());
|
|
6935
6631
|
}
|
|
6936
6632
|
|
|
6937
|
-
static __inline__ __m256i
|
|
6938
|
-
_mm256_mask_broadcastq_epi64
|
|
6939
|
-
{
|
|
6633
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6634
|
+
_mm256_mask_broadcastq_epi64(__m256i __O, __mmask8 __M, __m128i __A) {
|
|
6940
6635
|
return (__m256i)__builtin_ia32_selectq_256(__M,
|
|
6941
6636
|
(__v4di) _mm256_broadcastq_epi64(__A),
|
|
6942
6637
|
(__v4di) __O);
|
|
6943
6638
|
}
|
|
6944
6639
|
|
|
6945
|
-
static __inline__ __m256i
|
|
6946
|
-
_mm256_maskz_broadcastq_epi64
|
|
6947
|
-
{
|
|
6640
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
6641
|
+
_mm256_maskz_broadcastq_epi64(__mmask8 __M, __m128i __A) {
|
|
6948
6642
|
return (__m256i)__builtin_ia32_selectq_256(__M,
|
|
6949
6643
|
(__v4di) _mm256_broadcastq_epi64(__A),
|
|
6950
6644
|
(__v4di) _mm256_setzero_si256());
|
|
@@ -7536,9 +7230,8 @@ _mm256_mask_cvtusepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
|
|
|
7536
7230
|
__builtin_ia32_pmovusqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M);
|
|
7537
7231
|
}
|
|
7538
7232
|
|
|
7539
|
-
static __inline__ __m128i
|
|
7540
|
-
_mm_cvtepi32_epi8
|
|
7541
|
-
{
|
|
7233
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7234
|
+
_mm_cvtepi32_epi8(__m128i __A) {
|
|
7542
7235
|
return (__m128i)__builtin_shufflevector(
|
|
7543
7236
|
__builtin_convertvector((__v4si)__A, __v4qi), (__v4qi){0, 0, 0, 0}, 0, 1,
|
|
7544
7237
|
2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7);
|
|
@@ -7566,9 +7259,8 @@ _mm_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
|
|
|
7566
7259
|
__builtin_ia32_pmovdb128mem_mask ((__v16qi *) __P, (__v4si) __A, __M);
|
|
7567
7260
|
}
|
|
7568
7261
|
|
|
7569
|
-
static __inline__ __m128i
|
|
7570
|
-
_mm256_cvtepi32_epi8
|
|
7571
|
-
{
|
|
7262
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7263
|
+
_mm256_cvtepi32_epi8(__m256i __A) {
|
|
7572
7264
|
return (__m128i)__builtin_shufflevector(
|
|
7573
7265
|
__builtin_convertvector((__v8si)__A, __v8qi),
|
|
7574
7266
|
(__v8qi){0, 0, 0, 0, 0, 0, 0, 0}, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
|
|
@@ -7576,8 +7268,7 @@ _mm256_cvtepi32_epi8 (__m256i __A)
|
|
|
7576
7268
|
}
|
|
7577
7269
|
|
|
7578
7270
|
static __inline__ __m128i __DEFAULT_FN_ATTRS256
|
|
7579
|
-
_mm256_mask_cvtepi32_epi8
|
|
7580
|
-
{
|
|
7271
|
+
_mm256_mask_cvtepi32_epi8(__m128i __O, __mmask8 __M, __m256i __A) {
|
|
7581
7272
|
return (__m128i) __builtin_ia32_pmovdb256_mask ((__v8si) __A,
|
|
7582
7273
|
(__v16qi) __O, __M);
|
|
7583
7274
|
}
|
|
@@ -7596,9 +7287,8 @@ _mm256_mask_cvtepi32_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
|
|
|
7596
7287
|
__builtin_ia32_pmovdb256mem_mask ((__v16qi *) __P, (__v8si) __A, __M);
|
|
7597
7288
|
}
|
|
7598
7289
|
|
|
7599
|
-
static __inline__ __m128i
|
|
7600
|
-
_mm_cvtepi32_epi16
|
|
7601
|
-
{
|
|
7290
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7291
|
+
_mm_cvtepi32_epi16(__m128i __A) {
|
|
7602
7292
|
return (__m128i)__builtin_shufflevector(
|
|
7603
7293
|
__builtin_convertvector((__v4si)__A, __v4hi), (__v4hi){0, 0, 0, 0}, 0, 1,
|
|
7604
7294
|
2, 3, 4, 5, 6, 7);
|
|
@@ -7625,9 +7315,8 @@ _mm_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
|
|
|
7625
7315
|
__builtin_ia32_pmovdw128mem_mask ((__v8hi *) __P, (__v4si) __A, __M);
|
|
7626
7316
|
}
|
|
7627
7317
|
|
|
7628
|
-
static __inline__ __m128i
|
|
7629
|
-
_mm256_cvtepi32_epi16
|
|
7630
|
-
{
|
|
7318
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7319
|
+
_mm256_cvtepi32_epi16(__m256i __A) {
|
|
7631
7320
|
return (__m128i)__builtin_convertvector((__v8si)__A, __v8hi);
|
|
7632
7321
|
}
|
|
7633
7322
|
|
|
@@ -7652,9 +7341,8 @@ _mm256_mask_cvtepi32_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
|
|
|
7652
7341
|
__builtin_ia32_pmovdw256mem_mask ((__v8hi *) __P, (__v8si) __A, __M);
|
|
7653
7342
|
}
|
|
7654
7343
|
|
|
7655
|
-
static __inline__ __m128i
|
|
7656
|
-
_mm_cvtepi64_epi8
|
|
7657
|
-
{
|
|
7344
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7345
|
+
_mm_cvtepi64_epi8(__m128i __A) {
|
|
7658
7346
|
return (__m128i)__builtin_shufflevector(
|
|
7659
7347
|
__builtin_convertvector((__v2di)__A, __v2qi), (__v2qi){0, 0}, 0, 1, 2, 3,
|
|
7660
7348
|
3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3);
|
|
@@ -7681,9 +7369,8 @@ _mm_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A)
|
|
|
7681
7369
|
__builtin_ia32_pmovqb128mem_mask ((__v16qi *) __P, (__v2di) __A, __M);
|
|
7682
7370
|
}
|
|
7683
7371
|
|
|
7684
|
-
static __inline__ __m128i
|
|
7685
|
-
_mm256_cvtepi64_epi8
|
|
7686
|
-
{
|
|
7372
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7373
|
+
_mm256_cvtepi64_epi8(__m256i __A) {
|
|
7687
7374
|
return (__m128i)__builtin_shufflevector(
|
|
7688
7375
|
__builtin_convertvector((__v4di)__A, __v4qi), (__v4qi){0, 0, 0, 0}, 0, 1,
|
|
7689
7376
|
2, 3, 4, 5, 6, 7, 7, 7, 7, 7, 7, 7, 7, 7);
|
|
@@ -7710,9 +7397,8 @@ _mm256_mask_cvtepi64_storeu_epi8 (void * __P, __mmask8 __M, __m256i __A)
|
|
|
7710
7397
|
__builtin_ia32_pmovqb256mem_mask ((__v16qi *) __P, (__v4di) __A, __M);
|
|
7711
7398
|
}
|
|
7712
7399
|
|
|
7713
|
-
static __inline__ __m128i
|
|
7714
|
-
_mm_cvtepi64_epi32
|
|
7715
|
-
{
|
|
7400
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7401
|
+
_mm_cvtepi64_epi32(__m128i __A) {
|
|
7716
7402
|
return (__m128i)__builtin_shufflevector(
|
|
7717
7403
|
__builtin_convertvector((__v2di)__A, __v2si), (__v2si){0, 0}, 0, 1, 2, 3);
|
|
7718
7404
|
}
|
|
@@ -7738,23 +7424,20 @@ _mm_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m128i __A)
|
|
|
7738
7424
|
__builtin_ia32_pmovqd128mem_mask ((__v4si *) __P, (__v2di) __A, __M);
|
|
7739
7425
|
}
|
|
7740
7426
|
|
|
7741
|
-
static __inline__ __m128i
|
|
7742
|
-
_mm256_cvtepi64_epi32
|
|
7743
|
-
{
|
|
7427
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7428
|
+
_mm256_cvtepi64_epi32(__m256i __A) {
|
|
7744
7429
|
return (__m128i)__builtin_convertvector((__v4di)__A, __v4si);
|
|
7745
7430
|
}
|
|
7746
7431
|
|
|
7747
|
-
static __inline__ __m128i
|
|
7748
|
-
_mm256_mask_cvtepi64_epi32
|
|
7749
|
-
{
|
|
7432
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7433
|
+
_mm256_mask_cvtepi64_epi32(__m128i __O, __mmask8 __M, __m256i __A) {
|
|
7750
7434
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
|
|
7751
7435
|
(__v4si)_mm256_cvtepi64_epi32(__A),
|
|
7752
7436
|
(__v4si)__O);
|
|
7753
7437
|
}
|
|
7754
7438
|
|
|
7755
|
-
static __inline__ __m128i
|
|
7756
|
-
_mm256_maskz_cvtepi64_epi32
|
|
7757
|
-
{
|
|
7439
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7440
|
+
_mm256_maskz_cvtepi64_epi32(__mmask8 __M, __m256i __A) {
|
|
7758
7441
|
return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M,
|
|
7759
7442
|
(__v4si)_mm256_cvtepi64_epi32(__A),
|
|
7760
7443
|
(__v4si)_mm_setzero_si128());
|
|
@@ -7766,9 +7449,8 @@ _mm256_mask_cvtepi64_storeu_epi32 (void * __P, __mmask8 __M, __m256i __A)
|
|
|
7766
7449
|
__builtin_ia32_pmovqd256mem_mask ((__v4si *) __P, (__v4di) __A, __M);
|
|
7767
7450
|
}
|
|
7768
7451
|
|
|
7769
|
-
static __inline__ __m128i
|
|
7770
|
-
_mm_cvtepi64_epi16
|
|
7771
|
-
{
|
|
7452
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7453
|
+
_mm_cvtepi64_epi16(__m128i __A) {
|
|
7772
7454
|
return (__m128i)__builtin_shufflevector(
|
|
7773
7455
|
__builtin_convertvector((__v2di)__A, __v2hi), (__v2hi){0, 0}, 0, 1, 2, 3,
|
|
7774
7456
|
3, 3, 3, 3);
|
|
@@ -7796,9 +7478,8 @@ _mm_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m128i __A)
|
|
|
7796
7478
|
__builtin_ia32_pmovqw128mem_mask ((__v8hi *) __P, (__v2di) __A, __M);
|
|
7797
7479
|
}
|
|
7798
7480
|
|
|
7799
|
-
static __inline__ __m128i
|
|
7800
|
-
_mm256_cvtepi64_epi16
|
|
7801
|
-
{
|
|
7481
|
+
static __inline__ __m128i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7482
|
+
_mm256_cvtepi64_epi16(__m256i __A) {
|
|
7802
7483
|
return (__m128i)__builtin_shufflevector(
|
|
7803
7484
|
__builtin_convertvector((__v4di)__A, __v4hi), (__v4hi){0, 0, 0, 0}, 0, 1,
|
|
7804
7485
|
2, 3, 4, 5, 6, 7);
|
|
@@ -7825,11 +7506,10 @@ _mm256_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
|
|
|
7825
7506
|
__builtin_ia32_pmovqw256mem_mask ((__v8hi *) __P, (__v4di) __A, __M);
|
|
7826
7507
|
}
|
|
7827
7508
|
|
|
7828
|
-
#define _mm256_extractf32x4_ps(A, imm)
|
|
7829
|
-
((__m128)__builtin_ia32_extractf32x4_256_mask(
|
|
7830
|
-
|
|
7831
|
-
|
|
7832
|
-
(__mmask8)-1))
|
|
7509
|
+
#define _mm256_extractf32x4_ps(A, imm) \
|
|
7510
|
+
((__m128)__builtin_ia32_extractf32x4_256_mask( \
|
|
7511
|
+
(__v8sf)(__m256)(A), (int)(imm), (__v4sf)_mm_setzero_ps(), \
|
|
7512
|
+
(__mmask8) - 1))
|
|
7833
7513
|
|
|
7834
7514
|
#define _mm256_mask_extractf32x4_ps(W, U, A, imm) \
|
|
7835
7515
|
((__m128)__builtin_ia32_extractf32x4_256_mask((__v8sf)(__m256)(A), \
|
|
@@ -7843,11 +7523,10 @@ _mm256_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
|
|
|
7843
7523
|
(__v4sf)_mm_setzero_ps(), \
|
|
7844
7524
|
(__mmask8)(U)))
|
|
7845
7525
|
|
|
7846
|
-
#define _mm256_extracti32x4_epi32(A, imm)
|
|
7847
|
-
((__m128i)__builtin_ia32_extracti32x4_256_mask(
|
|
7848
|
-
|
|
7849
|
-
|
|
7850
|
-
(__mmask8)-1))
|
|
7526
|
+
#define _mm256_extracti32x4_epi32(A, imm) \
|
|
7527
|
+
((__m128i)__builtin_ia32_extracti32x4_256_mask( \
|
|
7528
|
+
(__v8si)(__m256i)(A), (int)(imm), (__v4si)_mm_setzero_si128(), \
|
|
7529
|
+
(__mmask8) - 1))
|
|
7851
7530
|
|
|
7852
7531
|
#define _mm256_mask_extracti32x4_epi32(W, U, A, imm) \
|
|
7853
7532
|
((__m128i)__builtin_ia32_extracti32x4_256_mask((__v8si)(__m256i)(A), \
|
|
@@ -8083,47 +7762,41 @@ _mm256_mask_cvtepi64_storeu_epi16 (void * __P, __mmask8 __M, __m256i __A)
|
|
|
8083
7762
|
(__v4di)_mm256_permutex_epi64((X), (C)), \
|
|
8084
7763
|
(__v4di)_mm256_setzero_si256()))
|
|
8085
7764
|
|
|
8086
|
-
static __inline__ __m256d
|
|
8087
|
-
_mm256_permutexvar_pd
|
|
8088
|
-
{
|
|
7765
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7766
|
+
_mm256_permutexvar_pd(__m256i __X, __m256d __Y) {
|
|
8089
7767
|
return (__m256d)__builtin_ia32_permvardf256((__v4df)__Y, (__v4di)__X);
|
|
8090
7768
|
}
|
|
8091
7769
|
|
|
8092
|
-
static __inline__ __m256d
|
|
8093
|
-
_mm256_mask_permutexvar_pd
|
|
8094
|
-
|
|
8095
|
-
{
|
|
7770
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7771
|
+
_mm256_mask_permutexvar_pd(__m256d __W, __mmask8 __U, __m256i __X,
|
|
7772
|
+
__m256d __Y) {
|
|
8096
7773
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
8097
7774
|
(__v4df)_mm256_permutexvar_pd(__X, __Y),
|
|
8098
7775
|
(__v4df)__W);
|
|
8099
7776
|
}
|
|
8100
7777
|
|
|
8101
|
-
static __inline__ __m256d
|
|
8102
|
-
_mm256_maskz_permutexvar_pd
|
|
8103
|
-
{
|
|
7778
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7779
|
+
_mm256_maskz_permutexvar_pd(__mmask8 __U, __m256i __X, __m256d __Y) {
|
|
8104
7780
|
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U,
|
|
8105
7781
|
(__v4df)_mm256_permutexvar_pd(__X, __Y),
|
|
8106
7782
|
(__v4df)_mm256_setzero_pd());
|
|
8107
7783
|
}
|
|
8108
7784
|
|
|
8109
|
-
static __inline__ __m256i
|
|
8110
|
-
_mm256_permutexvar_epi64
|
|
8111
|
-
{
|
|
7785
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7786
|
+
_mm256_permutexvar_epi64(__m256i __X, __m256i __Y) {
|
|
8112
7787
|
return (__m256i)__builtin_ia32_permvardi256((__v4di) __Y, (__v4di) __X);
|
|
8113
7788
|
}
|
|
8114
7789
|
|
|
8115
|
-
static __inline__ __m256i
|
|
8116
|
-
_mm256_maskz_permutexvar_epi64
|
|
8117
|
-
{
|
|
7790
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7791
|
+
_mm256_maskz_permutexvar_epi64(__mmask8 __M, __m256i __X, __m256i __Y) {
|
|
8118
7792
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
8119
7793
|
(__v4di)_mm256_permutexvar_epi64(__X, __Y),
|
|
8120
7794
|
(__v4di)_mm256_setzero_si256());
|
|
8121
7795
|
}
|
|
8122
7796
|
|
|
8123
|
-
static __inline__ __m256i
|
|
8124
|
-
_mm256_mask_permutexvar_epi64
|
|
8125
|
-
|
|
8126
|
-
{
|
|
7797
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7798
|
+
_mm256_mask_permutexvar_epi64(__m256i __W, __mmask8 __M, __m256i __X,
|
|
7799
|
+
__m256i __Y) {
|
|
8127
7800
|
return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M,
|
|
8128
7801
|
(__v4di)_mm256_permutexvar_epi64(__X, __Y),
|
|
8129
7802
|
(__v4di)__W);
|
|
@@ -8131,17 +7804,15 @@ _mm256_mask_permutexvar_epi64 (__m256i __W, __mmask8 __M, __m256i __X,
|
|
|
8131
7804
|
|
|
8132
7805
|
#define _mm256_permutexvar_ps(A, B) _mm256_permutevar8x32_ps((B), (A))
|
|
8133
7806
|
|
|
8134
|
-
static __inline__ __m256
|
|
8135
|
-
_mm256_mask_permutexvar_ps(__m256 __W, __mmask8 __U, __m256i __X, __m256 __Y)
|
|
8136
|
-
{
|
|
7807
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7808
|
+
_mm256_mask_permutexvar_ps(__m256 __W, __mmask8 __U, __m256i __X, __m256 __Y) {
|
|
8137
7809
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
8138
7810
|
(__v8sf)_mm256_permutexvar_ps(__X, __Y),
|
|
8139
7811
|
(__v8sf)__W);
|
|
8140
7812
|
}
|
|
8141
7813
|
|
|
8142
|
-
static __inline__ __m256
|
|
8143
|
-
_mm256_maskz_permutexvar_ps(__mmask8 __U, __m256i __X, __m256 __Y)
|
|
8144
|
-
{
|
|
7814
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7815
|
+
_mm256_maskz_permutexvar_ps(__mmask8 __U, __m256i __X, __m256 __Y) {
|
|
8145
7816
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
8146
7817
|
(__v8sf)_mm256_permutexvar_ps(__X, __Y),
|
|
8147
7818
|
(__v8sf)_mm256_setzero_ps());
|
|
@@ -8149,18 +7820,16 @@ _mm256_maskz_permutexvar_ps(__mmask8 __U, __m256i __X, __m256 __Y)
|
|
|
8149
7820
|
|
|
8150
7821
|
#define _mm256_permutexvar_epi32(A, B) _mm256_permutevar8x32_epi32((B), (A))
|
|
8151
7822
|
|
|
8152
|
-
static __inline__ __m256i
|
|
7823
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
8153
7824
|
_mm256_mask_permutexvar_epi32(__m256i __W, __mmask8 __M, __m256i __X,
|
|
8154
|
-
__m256i __Y)
|
|
8155
|
-
{
|
|
7825
|
+
__m256i __Y) {
|
|
8156
7826
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
8157
7827
|
(__v8si)_mm256_permutexvar_epi32(__X, __Y),
|
|
8158
7828
|
(__v8si)__W);
|
|
8159
7829
|
}
|
|
8160
7830
|
|
|
8161
|
-
static __inline__ __m256i
|
|
8162
|
-
_mm256_maskz_permutexvar_epi32(__mmask8 __M, __m256i __X, __m256i __Y)
|
|
8163
|
-
{
|
|
7831
|
+
static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7832
|
+
_mm256_maskz_permutexvar_epi32(__mmask8 __M, __m256i __X, __m256i __Y) {
|
|
8164
7833
|
return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M,
|
|
8165
7834
|
(__v8si)_mm256_permutexvar_epi32(__X, __Y),
|
|
8166
7835
|
(__v8si)_mm256_setzero_si256());
|
|
@@ -8222,65 +7891,57 @@ _mm256_maskz_permutexvar_epi32(__mmask8 __M, __m256i __X, __m256i __Y)
|
|
|
8222
7891
|
(__v4di)_mm256_alignr_epi64((A), (B), (imm)), \
|
|
8223
7892
|
(__v4di)_mm256_setzero_si256()))
|
|
8224
7893
|
|
|
8225
|
-
static __inline__ __m128
|
|
8226
|
-
_mm_mask_movehdup_ps
|
|
8227
|
-
{
|
|
7894
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7895
|
+
_mm_mask_movehdup_ps(__m128 __W, __mmask8 __U, __m128 __A) {
|
|
8228
7896
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
8229
7897
|
(__v4sf)_mm_movehdup_ps(__A),
|
|
8230
7898
|
(__v4sf)__W);
|
|
8231
7899
|
}
|
|
8232
7900
|
|
|
8233
|
-
static __inline__ __m128
|
|
8234
|
-
_mm_maskz_movehdup_ps
|
|
8235
|
-
{
|
|
7901
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7902
|
+
_mm_maskz_movehdup_ps(__mmask8 __U, __m128 __A) {
|
|
8236
7903
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
8237
7904
|
(__v4sf)_mm_movehdup_ps(__A),
|
|
8238
7905
|
(__v4sf)_mm_setzero_ps());
|
|
8239
7906
|
}
|
|
8240
7907
|
|
|
8241
|
-
static __inline__ __m256
|
|
8242
|
-
_mm256_mask_movehdup_ps
|
|
8243
|
-
{
|
|
7908
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7909
|
+
_mm256_mask_movehdup_ps(__m256 __W, __mmask8 __U, __m256 __A) {
|
|
8244
7910
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
8245
7911
|
(__v8sf)_mm256_movehdup_ps(__A),
|
|
8246
7912
|
(__v8sf)__W);
|
|
8247
7913
|
}
|
|
8248
7914
|
|
|
8249
|
-
static __inline__ __m256
|
|
8250
|
-
_mm256_maskz_movehdup_ps
|
|
8251
|
-
{
|
|
7915
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7916
|
+
_mm256_maskz_movehdup_ps(__mmask8 __U, __m256 __A) {
|
|
8252
7917
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
8253
7918
|
(__v8sf)_mm256_movehdup_ps(__A),
|
|
8254
7919
|
(__v8sf)_mm256_setzero_ps());
|
|
8255
7920
|
}
|
|
8256
7921
|
|
|
8257
|
-
static __inline__ __m128
|
|
8258
|
-
_mm_mask_moveldup_ps
|
|
8259
|
-
{
|
|
7922
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7923
|
+
_mm_mask_moveldup_ps(__m128 __W, __mmask8 __U, __m128 __A) {
|
|
8260
7924
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
8261
7925
|
(__v4sf)_mm_moveldup_ps(__A),
|
|
8262
7926
|
(__v4sf)__W);
|
|
8263
7927
|
}
|
|
8264
7928
|
|
|
8265
|
-
static __inline__ __m128
|
|
8266
|
-
_mm_maskz_moveldup_ps
|
|
8267
|
-
{
|
|
7929
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7930
|
+
_mm_maskz_moveldup_ps(__mmask8 __U, __m128 __A) {
|
|
8268
7931
|
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U,
|
|
8269
7932
|
(__v4sf)_mm_moveldup_ps(__A),
|
|
8270
7933
|
(__v4sf)_mm_setzero_ps());
|
|
8271
7934
|
}
|
|
8272
7935
|
|
|
8273
|
-
static __inline__ __m256
|
|
8274
|
-
_mm256_mask_moveldup_ps
|
|
8275
|
-
{
|
|
7936
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7937
|
+
_mm256_mask_moveldup_ps(__m256 __W, __mmask8 __U, __m256 __A) {
|
|
8276
7938
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
8277
7939
|
(__v8sf)_mm256_moveldup_ps(__A),
|
|
8278
7940
|
(__v8sf)__W);
|
|
8279
7941
|
}
|
|
8280
7942
|
|
|
8281
|
-
static __inline__ __m256
|
|
8282
|
-
_mm256_maskz_moveldup_ps
|
|
8283
|
-
{
|
|
7943
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7944
|
+
_mm256_maskz_moveldup_ps(__mmask8 __U, __m256 __A) {
|
|
8284
7945
|
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U,
|
|
8285
7946
|
(__v8sf)_mm256_moveldup_ps(__A),
|
|
8286
7947
|
(__v8sf)_mm256_setzero_ps());
|
|
@@ -8306,68 +7967,52 @@ _mm256_maskz_moveldup_ps (__mmask8 __U, __m256 __A)
|
|
|
8306
7967
|
(__v4si)_mm_shuffle_epi32((A), (I)), \
|
|
8307
7968
|
(__v4si)_mm_setzero_si128()))
|
|
8308
7969
|
|
|
8309
|
-
static __inline__ __m128d
|
|
8310
|
-
_mm_mask_mov_pd
|
|
8311
|
-
|
|
8312
|
-
|
|
8313
|
-
(__v2df) __A,
|
|
8314
|
-
(__v2df) __W);
|
|
7970
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7971
|
+
_mm_mask_mov_pd(__m128d __W, __mmask8 __U, __m128d __A) {
|
|
7972
|
+
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, (__v2df)__A,
|
|
7973
|
+
(__v2df)__W);
|
|
8315
7974
|
}
|
|
8316
7975
|
|
|
8317
|
-
static __inline__ __m128d
|
|
8318
|
-
_mm_maskz_mov_pd
|
|
8319
|
-
|
|
8320
|
-
|
|
8321
|
-
(__v2df) __A,
|
|
8322
|
-
(__v2df) _mm_setzero_pd ());
|
|
7976
|
+
static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7977
|
+
_mm_maskz_mov_pd(__mmask8 __U, __m128d __A) {
|
|
7978
|
+
return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, (__v2df)__A,
|
|
7979
|
+
(__v2df)_mm_setzero_pd());
|
|
8323
7980
|
}
|
|
8324
7981
|
|
|
8325
|
-
static __inline__ __m256d
|
|
8326
|
-
_mm256_mask_mov_pd
|
|
8327
|
-
|
|
8328
|
-
|
|
8329
|
-
(__v4df) __A,
|
|
8330
|
-
(__v4df) __W);
|
|
7982
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7983
|
+
_mm256_mask_mov_pd(__m256d __W, __mmask8 __U, __m256d __A) {
|
|
7984
|
+
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, (__v4df)__A,
|
|
7985
|
+
(__v4df)__W);
|
|
8331
7986
|
}
|
|
8332
7987
|
|
|
8333
|
-
static __inline__ __m256d
|
|
8334
|
-
_mm256_maskz_mov_pd
|
|
8335
|
-
|
|
8336
|
-
|
|
8337
|
-
(__v4df) __A,
|
|
8338
|
-
(__v4df) _mm256_setzero_pd ());
|
|
7988
|
+
static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
7989
|
+
_mm256_maskz_mov_pd(__mmask8 __U, __m256d __A) {
|
|
7990
|
+
return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, (__v4df)__A,
|
|
7991
|
+
(__v4df)_mm256_setzero_pd());
|
|
8339
7992
|
}
|
|
8340
7993
|
|
|
8341
|
-
static __inline__ __m128
|
|
8342
|
-
_mm_mask_mov_ps
|
|
8343
|
-
|
|
8344
|
-
|
|
8345
|
-
(__v4sf) __A,
|
|
8346
|
-
(__v4sf) __W);
|
|
7994
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
7995
|
+
_mm_mask_mov_ps(__m128 __W, __mmask8 __U, __m128 __A) {
|
|
7996
|
+
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, (__v4sf)__A,
|
|
7997
|
+
(__v4sf)__W);
|
|
8347
7998
|
}
|
|
8348
7999
|
|
|
8349
|
-
static __inline__ __m128
|
|
8350
|
-
_mm_maskz_mov_ps
|
|
8351
|
-
|
|
8352
|
-
|
|
8353
|
-
(__v4sf) __A,
|
|
8354
|
-
(__v4sf) _mm_setzero_ps ());
|
|
8000
|
+
static __inline__ __m128 __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
8001
|
+
_mm_maskz_mov_ps(__mmask8 __U, __m128 __A) {
|
|
8002
|
+
return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, (__v4sf)__A,
|
|
8003
|
+
(__v4sf)_mm_setzero_ps());
|
|
8355
8004
|
}
|
|
8356
8005
|
|
|
8357
|
-
static __inline__ __m256
|
|
8358
|
-
_mm256_mask_mov_ps
|
|
8359
|
-
|
|
8360
|
-
|
|
8361
|
-
(__v8sf) __A,
|
|
8362
|
-
(__v8sf) __W);
|
|
8006
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
8007
|
+
_mm256_mask_mov_ps(__m256 __W, __mmask8 __U, __m256 __A) {
|
|
8008
|
+
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, (__v8sf)__A,
|
|
8009
|
+
(__v8sf)__W);
|
|
8363
8010
|
}
|
|
8364
8011
|
|
|
8365
|
-
static __inline__ __m256
|
|
8366
|
-
_mm256_maskz_mov_ps
|
|
8367
|
-
|
|
8368
|
-
|
|
8369
|
-
(__v8sf) __A,
|
|
8370
|
-
(__v8sf) _mm256_setzero_ps ());
|
|
8012
|
+
static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
8013
|
+
_mm256_maskz_mov_ps(__mmask8 __U, __m256 __A) {
|
|
8014
|
+
return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, (__v8sf)__A,
|
|
8015
|
+
(__v8sf)_mm256_setzero_ps());
|
|
8371
8016
|
}
|
|
8372
8017
|
|
|
8373
8018
|
static __inline__ __m128 __DEFAULT_FN_ATTRS128
|
|
@@ -8430,8 +8075,9 @@ _mm256_maskz_cvtph_ps (__mmask8 __U, __m128i __A)
|
|
|
8430
8075
|
#define _mm256_mask_cvtps_ph _mm256_mask_cvt_roundps_ph
|
|
8431
8076
|
#define _mm256_maskz_cvtps_ph _mm256_maskz_cvt_roundps_ph
|
|
8432
8077
|
|
|
8433
|
-
|
|
8434
8078
|
#undef __DEFAULT_FN_ATTRS128
|
|
8435
8079
|
#undef __DEFAULT_FN_ATTRS256
|
|
8080
|
+
#undef __DEFAULT_FN_ATTRS256_CONSTEXPR
|
|
8081
|
+
#undef __DEFAULT_FN_ATTRS128_CONSTEXPR
|
|
8436
8082
|
|
|
8437
8083
|
#endif /* __AVX512VLINTRIN_H */
|