koffi 2.16.1 → 3.0.0-alpha.2

package/src/koffi/src/{abi_arm64_asm.S → abi/arm64_asm.S}

@@ -22,9 +22,9 @@
 .global SYMBOL(ForwardCallDDDD)
 
 # The X variants are slightly slower, and are used when XMM arguments must be forwarded.
-.global SYMBOL(ForwardCallXGG)
-.global SYMBOL(ForwardCallXF)
-.global SYMBOL(ForwardCallXDDDD)
+.global SYMBOL(ForwardCallGGX)
+.global SYMBOL(ForwardCallFX)
+.global SYMBOL(ForwardCallDDDDX)
 
 # Copy function pointer to r9, in order to save it through argument forwarding.
 # Also make a copy of the SP to CallData::old_sp because the callback system might need it.
@@ -40,7 +40,7 @@
     .cfi_def_cfa x29, 16
     str x29, [x2, 0]
     mov x9, x0
-    add sp, x1, #136
+    add sp, x1, #152
 .endm
 
 # Call native function.
@@ -90,19 +90,19 @@ SYMBOL(ForwardCallDDDD):
     forward_gpr
     epilogue
 
-SYMBOL(ForwardCallXGG):
+SYMBOL(ForwardCallGGX):
     prologue
     forward_vec
     forward_gpr
     epilogue
 
-SYMBOL(ForwardCallXF):
+SYMBOL(ForwardCallFX):
     prologue
     forward_vec
     forward_gpr
     epilogue
 
-SYMBOL(ForwardCallXDDDD):
+SYMBOL(ForwardCallDDDDX):
     prologue
     forward_vec
     forward_gpr
@@ -114,7 +114,12 @@ SYMBOL(ForwardCallXDDDD):
 .global SYMBOL(RelayCallback)
 .global SYMBOL(SwitchAndRelay)
 .global SYMBOL(RelayDirect)
+#if defined(_WIN32)
+.global SYMBOL(FindTrampolineStart)
+.global SYMBOL(FindTrampolineEnd)
+#else
 .global SYMBOL(TrampolineEnd)
+#endif
 
 # First, make a copy of the argument registers.
 # Then call the C function RelayCallback with the following arguments:
@@ -137,21 +142,21 @@ RelayTrampoline:
     .cfi_offset 30, 8
     sub sp, sp, #192
     .cfi_def_cfa sp, 208
-    stp x0, x1, [sp, 0]
-    stp x2, x3, [sp, 16]
-    stp x4, x5, [sp, 32]
-    stp x6, x7, [sp, 48]
-    str x8, [sp, 64]
-    stp d0, d1, [sp, 72]
-    stp d2, d3, [sp, 88]
-    stp d4, d5, [sp, 104]
-    stp d6, d7, [sp, 120]
+    stp x0, x1, [sp, 48]
+    stp x2, x3, [sp, 64]
+    stp x4, x5, [sp, 80]
+    stp x6, x7, [sp, 96]
+    str x8, [sp, 112]
+    stp d0, d1, [sp, 120]
+    stp d2, d3, [sp, 136]
+    stp d4, d5, [sp, 152]
+    stp d6, d7, [sp, 168]
     mov x0, x9
     mov x1, sp
     bl SYMBOL(RelayCallback)
-    ldp x0, x1, [sp, 136]
-    ldp d0, d1, [sp, 152]
-    ldp d2, d3, [sp, 168]
+    ldp x0, x1, [sp, 0]
+    ldp d0, d1, [sp, 16]
+    ldp d2, d3, [sp, 32]
     add sp, sp, #192
     .cfi_def_cfa sp, 16
     ldp x29, x30, [sp], 16
@@ -174,10 +179,7 @@ SYMBOL(SwitchAndRelay):
     .cfi_offset 29, 16
     .cfi_offset 30, 8
     mov x29, sp
-    ldr x9, [x4, 0]
-    sub x9, sp, x9
-    and x9, x9, #-16
-    str x9, [x4, 8]
+    str x29, [x4, 8]
     mov sp, x3
     bl SYMBOL(RelayDirect)
     mov sp, x29
@@ -189,7 +191,26 @@ SYMBOL(SwitchAndRelay):
     ret
     .cfi_endproc
 
+#if defined(_WIN32)
+
+.balign 16
+SYMBOL(FindTrampolineStart):
+    adr x0, 16
+    and x0, x0, #-16
+    ret
+.balign 16
+
+#include "gnu.inc"
+
+SYMBOL(FindTrampolineEnd):
+    adr x0, 0
+    ret
+
+#else
+
 #include "gnu.inc"
 
 SYMBOL(TrampolineEnd):
     ret
+
+#endif

package/src/koffi/src/{abi_arm64_asm.asm → abi/arm64_asm.asm}

@@ -13,9 +13,9 @@
     EXPORT ForwardCallDDDD
 
     ; The X variants are slightly slower, and are used when XMM arguments must be forwarded.
-    EXPORT ForwardCallXGG
-    EXPORT ForwardCallXF
-    EXPORT ForwardCallXDDDD
+    EXPORT ForwardCallGGX
+    EXPORT ForwardCallFX
+    EXPORT ForwardCallDDDDX
 
     ; Copy function pointer to r9, in order to save it through argument forwarding.
     ; Save RSP in r29 (non-volatile), and use carefully assembled stack provided by caller.
@@ -26,7 +26,7 @@
     mov x29, sp
     str x29, [x2, 0]
     mov x9, x0
-    add sp, x1, #136
+    add sp, x1, #152
     MEND
 
     ; Call native function.
@@ -80,21 +80,21 @@ ForwardCallDDDD PROC
     epilogue
     ENDP
 
-ForwardCallXGG PROC
+ForwardCallGGX PROC
     prologue
     forward_vec
     forward_gpr
     epilogue
     ENDP
 
-ForwardCallXF PROC
+ForwardCallFX PROC
     prologue
     forward_vec
     forward_gpr
     epilogue
     ENDP
 
-ForwardCallXDDDD PROC
+ForwardCallDDDDX PROC
     prologue
     forward_vec
     forward_gpr
@@ -109,7 +109,8 @@ ForwardCallXDDDD PROC
     EXPORT SwitchAndRelay
     EXPORT RelayDirect
     EXTERN RelayDirect
-    EXPORT TrampolineEnd
+    EXPORT FindTrampolineStart
+    EXPORT FindTrampolineEnd
 
     ; First, make a copy of the argument registers.
     ; Then call the C function RelayCallback with the following arguments:
@@ -126,21 +127,21 @@ ForwardCallXDDDD PROC
 RelayTrampoline PROC
     stp x29, x30, [sp, -16]!
     sub sp, sp, #192
-    stp x0, x1, [sp, 0]
-    stp x2, x3, [sp, 16]
-    stp x4, x5, [sp, 32]
-    stp x6, x7, [sp, 48]
-    str x8, [sp, 64]
-    stp d0, d1, [sp, 72]
-    stp d2, d3, [sp, 88]
-    stp d4, d5, [sp, 104]
-    stp d6, d7, [sp, 120]
+    stp x0, x1, [sp, 48]
+    stp x2, x3, [sp, 64]
+    stp x4, x5, [sp, 80]
+    stp x6, x7, [sp, 96]
+    str x8, [sp, 112]
+    stp d0, d1, [sp, 120]
+    stp d2, d3, [sp, 136]
+    stp d4, d5, [sp, 152]
+    stp d6, d7, [sp, 168]
     mov x0, x9
     mov x1, sp
     bl RelayCallback
-    ldp x0, x1, [sp, 136]
-    ldp d0, d1, [sp, 152]
-    ldp d2, d3, [sp, 168]
+    ldp x0, x1, [sp, 0]
+    ldp d0, d1, [sp, 16]
+    ldp d2, d3, [sp, 32]
     add sp, sp, #192
     ldp x29, x30, [sp], 16
     ret
@@ -154,10 +155,7 @@ RelayTrampoline PROC
 SwitchAndRelay PROC
     stp x29, x30, [sp, -16]!
     mov x29, sp
-    ldr x9, [x4, 0]
-    sub x9, sp, x9
-    and x9, x9, #-16
-    str x9, [x4, 8]
+    str x29, [x4, 8]
     mov sp, x3
     bl RelayDirect
     mov sp, x29
@@ -165,9 +163,18 @@ SwitchAndRelay PROC
     ret
     ENDP
 
+    ALIGN 16
+FindTrampolineStart PROC
+    adr x0, 16
+    and x0, x0, #-16
+    ret
+    ENDP
+    ALIGN 16
+
     INCLUDE armasm.inc
 
-TrampolineEnd PROC
+FindTrampolineEnd PROC
+    adr x0, 0
     ret
     ENDP
 

package/src/koffi/src/{abi_loong64_asm.S → abi/loong64_asm.S}

@@ -184,11 +184,7 @@ SwitchAndRelay:
     st.d $ra, $sp, 0
     st.d $fp, $sp, 8
     move $fp, $sp
-    ld.d $t0, $a4, 0
-    sub.d $t0, $sp, $t0
-    li.d $t1, -16
-    and $t0, $t0, $t1
-    st.d $t0, $a4, 8
+    st.d $sp, $a4, 8
     move $sp, $a3
     bl RelayDirect
     move $sp, $fp