wasmtime 19.0.2 → 20.0.0

data/ext/cargo-vendor/wasmtime-cranelift-20.0.0/src/compiler.rs

@@ -0,0 +1,1369 @@
+use crate::debug::{DwarfSectionRelocTarget, ModuleMemoryOffset};
+use crate::func_environ::FuncEnvironment;
+use crate::{array_call_signature, native_call_signature, DEBUG_ASSERT_TRAP_CODE};
+use crate::{builder::LinkOptions, wasm_call_signature, BuiltinFunctionSignatures};
+use crate::{CompiledFunction, ModuleTextBuilder};
+use anyhow::{Context as _, Result};
+use cranelift_codegen::ir::{self, InstBuilder, MemFlags, UserExternalName, UserFuncName, Value};
+use cranelift_codegen::isa::{
+    unwind::{UnwindInfo, UnwindInfoKind},
+    OwnedTargetIsa, TargetIsa,
+};
+use cranelift_codegen::print_errors::pretty_error;
+use cranelift_codegen::Context;
+use cranelift_codegen::{CompiledCode, MachStackMap};
+use cranelift_entity::{EntityRef, PrimaryMap};
+use cranelift_frontend::FunctionBuilder;
+use cranelift_wasm::{
+    DefinedFuncIndex, FuncTranslator, MemoryIndex, OwnedMemoryIndex, WasmFuncType, WasmValType,
+};
+use object::write::{Object, StandardSegment, SymbolId};
+use object::{RelocationEncoding, RelocationFlags, RelocationKind, SectionKind};
+use std::any::Any;
+use std::cmp;
+use std::collections::HashMap;
+use std::mem;
+use std::path;
+use std::sync::{Arc, Mutex};
+use wasmparser::{FuncValidatorAllocations, FunctionBody};
+use wasmtime_environ::{
+    AddressMapSection, BuiltinFunctionIndex, CacheStore, CompileError, FlagValue, FunctionBodyData,
+    FunctionLoc, ModuleTranslation, ModuleTypesBuilder, PtrSize, RelocationTarget,
+    StackMapInformation, TrapEncodingBuilder, Tunables, VMOffsets, WasmFunctionInfo,
+};
+
+#[cfg(feature = "component-model")]
+mod component;
+
+struct IncrementalCacheContext {
+    #[cfg(feature = "incremental-cache")]
+    cache_store: Arc<dyn CacheStore>,
+    num_hits: usize,
+    num_cached: usize,
+}
+
+struct CompilerContext {
+    func_translator: FuncTranslator,
+    codegen_context: Context,
+    incremental_cache_ctx: Option<IncrementalCacheContext>,
+    validator_allocations: FuncValidatorAllocations,
+}
+
+impl Default for CompilerContext {
+    fn default() -> Self {
+        Self {
+            func_translator: FuncTranslator::new(),
+            codegen_context: Context::new(),
+            incremental_cache_ctx: None,
+            validator_allocations: Default::default(),
+        }
+    }
+}
+
+/// A compiler that compiles a WebAssembly module with Compiler, translating
+/// the Wasm to Compiler IR, optimizing it and then translating to assembly.
+pub struct Compiler {
+    tunables: Tunables,
+    contexts: Mutex<Vec<CompilerContext>>,
+    isa: OwnedTargetIsa,
+    linkopts: LinkOptions,
+    cache_store: Option<Arc<dyn CacheStore>>,
+    clif_dir: Option<path::PathBuf>,
+    wmemcheck: bool,
+}
+
+impl Drop for Compiler {
+    fn drop(&mut self) {
+        if self.cache_store.is_none() {
+            return;
+        }
+
+        let mut num_hits = 0;
+        let mut num_cached = 0;
+        for ctx in self.contexts.lock().unwrap().iter() {
+            if let Some(ref cache_ctx) = ctx.incremental_cache_ctx {
+                num_hits += cache_ctx.num_hits;
+                num_cached += cache_ctx.num_cached;
+            }
+        }
+
+        let total = num_hits + num_cached;
+        if num_hits + num_cached > 0 {
+            log::trace!(
+                "Incremental compilation cache stats: {}/{} = {}% (hits/lookup)\ncached: {}",
+                num_hits,
+                total,
+                (num_hits as f32) / (total as f32) * 100.0,
+                num_cached
+            );
+        }
+    }
+}
+
+impl Compiler {
+    pub fn new(
+        tunables: Tunables,
+        isa: OwnedTargetIsa,
+        cache_store: Option<Arc<dyn CacheStore>>,
+        linkopts: LinkOptions,
+        clif_dir: Option<path::PathBuf>,
+        wmemcheck: bool,
+    ) -> Compiler {
+        Compiler {
+            contexts: Default::default(),
+            tunables,
+            isa,
+            linkopts,
+            cache_store,
+            clif_dir,
+            wmemcheck,
+        }
+    }
+}
+
+impl wasmtime_environ::Compiler for Compiler {
+    fn compile_function(
+        &self,
+        translation: &ModuleTranslation<'_>,
+        func_index: DefinedFuncIndex,
+        input: FunctionBodyData<'_>,
+        types: &ModuleTypesBuilder,
+    ) -> Result<(WasmFunctionInfo, Box<dyn Any + Send>), CompileError> {
+        let isa = &*self.isa;
+        let module = &translation.module;
+        let func_index = module.func_index(func_index);
+        let sig = translation.module.functions[func_index].signature;
+        let wasm_func_ty = &types[sig];
+
+        let mut compiler = self.function_compiler();
+
+        let context = &mut compiler.cx.codegen_context;
+        context.func.signature = wasm_call_signature(isa, wasm_func_ty, &self.tunables);
+        context.func.name = UserFuncName::User(UserExternalName {
+            namespace: crate::NS_WASM_FUNC,
+            index: func_index.as_u32(),
+        });
+
+        if self.tunables.generate_native_debuginfo {
+            context.func.collect_debug_info();
+        }
+
+        let mut func_env =
+            FuncEnvironment::new(isa, translation, types, &self.tunables, self.wmemcheck);
+
+        // The `stack_limit` global value below is the implementation of stack
+        // overflow checks in Wasmtime.
+        //
+        // The Wasm spec defines that stack overflows will raise a trap, and
+        // there's also an added constraint where as an embedder you frequently
+        // are running host-provided code called from wasm. WebAssembly and
+        // native code currently share the same call stack, so Wasmtime needs to
+        // make sure that host-provided code will have enough call-stack
+        // available to it.
+        //
+        // The way that stack overflow is handled here is by adding a prologue
+        // check to all functions for how much native stack is remaining. The
+        // `VMContext` pointer is the first argument to all functions, and the
+        // first field of this structure is `*const VMRuntimeLimits` and the
+        // first field of that is the stack limit. Note that the stack limit in
+        // this case means "if the stack pointer goes below this, trap". Each
+        // function which consumes stack space or isn't a leaf function starts
+        // off by loading the stack limit, checking it against the stack
+        // pointer, and optionally traps.
+        //
+        // This manual check allows the embedder to give wasm a relatively
+        // precise amount of stack allocation. Using this scheme we reserve a
+        // chunk of stack for wasm code relative from where wasm code was
+        // called. This ensures that native code called by wasm should have
+        // native stack space to run, and the numbers of stack spaces here
+        // should all be configurable for various embeddings.
+        //
+        // Note that this check is independent of each thread's stack guard page
+        // here. If the stack guard page is reached that's still considered an
+        // abort for the whole program since the runtime limits configured by
+        // the embedder should cause wasm to trap before it reaches that
+        // (ensuring the host has enough space as well for its functionality).
+        let vmctx = context
+            .func
+            .create_global_value(ir::GlobalValueData::VMContext);
+        let interrupts_ptr = context.func.create_global_value(ir::GlobalValueData::Load {
+            base: vmctx,
+            offset: i32::try_from(func_env.offsets.vmctx_runtime_limits())
+                .unwrap()
+                .into(),
+            global_type: isa.pointer_type(),
+            flags: MemFlags::trusted().with_readonly(),
+        });
+        let stack_limit = context.func.create_global_value(ir::GlobalValueData::Load {
+            base: interrupts_ptr,
+            offset: i32::try_from(func_env.offsets.ptr.vmruntime_limits_stack_limit())
+                .unwrap()
+                .into(),
+            global_type: isa.pointer_type(),
+            flags: MemFlags::trusted(),
+        });
+        context.func.stack_limit = Some(stack_limit);
+        let FunctionBodyData { validator, body } = input;
+        let mut validator =
+            validator.into_validator(mem::take(&mut compiler.cx.validator_allocations));
+        compiler.cx.func_translator.translate_body(
+            &mut validator,
+            body.clone(),
+            &mut context.func,
+            &mut func_env,
+        )?;
+
+        if let Some(path) = &self.clif_dir {
+            use std::io::Write;
+
+            let mut path = path.to_path_buf();
+            path.push(format!("wasm_func_{}", func_index.as_u32()));
+            path.set_extension("clif");
+
+            let mut output = std::fs::File::create(path).unwrap();
+            write!(output, "{}", context.func.display()).unwrap();
+        }
+
+        let (info, func) = compiler.finish_with_info(Some((&body, &self.tunables)))?;
+
+        let timing = cranelift_codegen::timing::take_current();
+        log::debug!("{:?} translated in {:?}", func_index, timing.total());
+        log::trace!("{:?} timing info\n{}", func_index, timing);
+
+        Ok((info, Box::new(func)))
+    }
+
+    fn compile_array_to_wasm_trampoline(
+        &self,
+        translation: &ModuleTranslation<'_>,
+        types: &ModuleTypesBuilder,
+        def_func_index: DefinedFuncIndex,
+    ) -> Result<Box<dyn Any + Send>, CompileError> {
+        let func_index = translation.module.func_index(def_func_index);
+        let sig = translation.module.functions[func_index].signature;
+        let wasm_func_ty = &types[sig];
+
+        let isa = &*self.isa;
+        let pointer_type = isa.pointer_type();
+        let wasm_call_sig = wasm_call_signature(isa, wasm_func_ty, &self.tunables);
+        let array_call_sig = array_call_signature(isa);
+
+        let mut compiler = self.function_compiler();
+        let func = ir::Function::with_name_signature(Default::default(), array_call_sig);
+        let (mut builder, block0) = compiler.builder(func);
+
+        let (vmctx, caller_vmctx, values_vec_ptr, values_vec_len) = {
+            let params = builder.func.dfg.block_params(block0);
+            (params[0], params[1], params[2], params[3])
+        };
+
+        // First load the actual arguments out of the array.
+        let mut args = self.load_values_from_array(
+            wasm_func_ty.params(),
+            &mut builder,
+            values_vec_ptr,
+            values_vec_len,
+        );
+        args.insert(0, caller_vmctx);
+        args.insert(0, vmctx);
+
+        // Just before we enter Wasm, save our stack pointer.
+        //
+        // Assert that we were really given a core Wasm vmctx, since that's
+        // what we are assuming with our offsets below.
+        debug_assert_vmctx_kind(isa, &mut builder, vmctx, wasmtime_environ::VMCONTEXT_MAGIC);
+        let offsets = VMOffsets::new(isa.pointer_bytes(), &translation.module);
+        let vm_runtime_limits_offset = offsets.vmctx_runtime_limits();
+        save_last_wasm_entry_sp(
+            &mut builder,
+            pointer_type,
+            &offsets.ptr,
+            vm_runtime_limits_offset,
+            vmctx,
+        );
+
+        // Then call the Wasm function with those arguments.
+        let call = declare_and_call(&mut builder, wasm_call_sig, func_index.as_u32(), &args);
+        let results = builder.func.dfg.inst_results(call).to_vec();
+
+        // Then store the results back into the array.
+        self.store_values_to_array(
+            &mut builder,
+            wasm_func_ty.returns(),
+            &results,
+            values_vec_ptr,
+            values_vec_len,
+        );
+
+        builder.ins().return_(&[]);
+        builder.finalize();
+
+        Ok(Box::new(compiler.finish()?))
+    }
+
+    fn compile_native_to_wasm_trampoline(
+        &self,
+        translation: &ModuleTranslation<'_>,
+        types: &ModuleTypesBuilder,
+        def_func_index: DefinedFuncIndex,
+    ) -> Result<Box<dyn Any + Send>, CompileError> {
+        let func_index = translation.module.func_index(def_func_index);
+        let sig = translation.module.functions[func_index].signature;
+        let wasm_func_ty = &types[sig];
+
+        let isa = &*self.isa;
+        let pointer_type = isa.pointer_type();
+        let func_index = translation.module.func_index(def_func_index);
+        let wasm_call_sig = wasm_call_signature(isa, wasm_func_ty, &self.tunables);
+        let native_call_sig = native_call_signature(isa, wasm_func_ty);
+
+        let mut compiler = self.function_compiler();
+        let func = ir::Function::with_name_signature(Default::default(), native_call_sig);
+        let (mut builder, block0) = compiler.builder(func);
+
+        let args = builder.func.dfg.block_params(block0).to_vec();
+        let vmctx = args[0];
+
+        // Since we are entering Wasm, save our SP.
+        //
+        // Assert that we were really given a core Wasm vmctx, since that's
+        // what we are assuming with our offsets below.
+        debug_assert_vmctx_kind(isa, &mut builder, vmctx, wasmtime_environ::VMCONTEXT_MAGIC);
+        let offsets = VMOffsets::new(isa.pointer_bytes(), &translation.module);
+        let vm_runtime_limits_offset = offsets.vmctx_runtime_limits();
+        save_last_wasm_entry_sp(
+            &mut builder,
+            pointer_type,
+            &offsets.ptr,
+            vm_runtime_limits_offset,
+            vmctx,
+        );
+
+        let ret = NativeRet::classify(isa, wasm_func_ty);
+        let wasm_args = ret.native_args(&args);
+
+        // Then call into Wasm.
+        let call = declare_and_call(&mut builder, wasm_call_sig, func_index.as_u32(), wasm_args);
+
+        // Forward the results along.
+        let results = builder.func.dfg.inst_results(call).to_vec();
+        ret.native_return(&mut builder, block0, &results);
+        builder.finalize();
+
+        Ok(Box::new(compiler.finish()?))
+    }
+
+    fn compile_wasm_to_native_trampoline(
+        &self,
+        wasm_func_ty: &WasmFuncType,
+    ) -> Result<Box<dyn Any + Send>, CompileError> {
+        let isa = &*self.isa;
+        let pointer_type = isa.pointer_type();
+        let wasm_call_sig = wasm_call_signature(isa, wasm_func_ty, &self.tunables);
+        let native_call_sig = native_call_signature(isa, wasm_func_ty);
+
+        let mut compiler = self.function_compiler();
+        let func = ir::Function::with_name_signature(Default::default(), wasm_call_sig);
+        let (mut builder, block0) = compiler.builder(func);
+
+        let mut args = builder.func.dfg.block_params(block0).to_vec();
+        let callee_vmctx = args[0];
+        let caller_vmctx = args[1];
+
+        let ret = NativeRet::classify(isa, wasm_func_ty);
+
+        // We are exiting Wasm, so save our PC and FP.
+        //
+        // Assert that the caller vmctx really is a core Wasm vmctx, since
+        // that's what we are assuming with our offsets below.
+        debug_assert_vmctx_kind(
+            isa,
+            &mut builder,
+            caller_vmctx,
+            wasmtime_environ::VMCONTEXT_MAGIC,
+        );
+        let ptr = isa.pointer_bytes();
+        let limits = builder.ins().load(
+            pointer_type,
+            MemFlags::trusted(),
+            caller_vmctx,
+            i32::try_from(ptr.vmcontext_runtime_limits()).unwrap(),
+        );
+        save_last_wasm_exit_fp_and_pc(&mut builder, pointer_type, &ptr, limits);
+
+        // If the native call signature for this function uses a return pointer
+        // then allocate the return pointer here on the stack and pass it as the
+        // last argument.
+        let slot = match &ret {
+            NativeRet::Bare => None,
+            NativeRet::Retptr { size, .. } => Some(builder.func.create_sized_stack_slot(
+                ir::StackSlotData::new(ir::StackSlotKind::ExplicitSlot, *size),
+            )),
+        };
+        if let Some(slot) = slot {
+            args.push(builder.ins().stack_addr(pointer_type, slot, 0));
+        }
+
+        // Load the actual callee out of the
+        // `VMNativeCallHostFuncContext::host_func`.
+        let ptr_size = isa.pointer_bytes();
+        let callee = builder.ins().load(
+            pointer_type,
+            MemFlags::trusted(),
+            callee_vmctx,
+            ptr_size.vmnative_call_host_func_context_func_ref()
+                + ptr_size.vm_func_ref_native_call(),
+        );
+
+        // Do an indirect call to the callee.
+        let callee_signature = builder.func.import_signature(native_call_sig);
+        let call = builder.ins().call_indirect(callee_signature, callee, &args);
+
+        // Forward the results back to the caller. If a return pointer was in
+        // use for the native call then load the results from the return pointer
+        // to pass through as native return values in the wasm abi.
+        let mut results = builder.func.dfg.inst_results(call).to_vec();
+        if let NativeRet::Retptr { slots, .. } = ret {
+            let base = *args.last().unwrap();
+            assert_eq!(slots.len(), wasm_func_ty.returns().len() - 1);
+            for (offset, ty) in slots {
+                results.push(
+                    builder
+                        .ins()
+                        .load(ty, ir::MemFlags::trusted(), base, offset),
+                );
+            }
+        }
+        builder.ins().return_(&results);
+        builder.finalize();
+
+        Ok(Box::new(compiler.finish()?))
+    }
+
+    fn append_code(
+        &self,
+        obj: &mut Object<'static>,
+        funcs: &[(String, Box<dyn Any + Send>)],
+        resolve_reloc: &dyn Fn(usize, RelocationTarget) -> usize,
+    ) -> Result<Vec<(SymbolId, FunctionLoc)>> {
+        let mut builder =
+            ModuleTextBuilder::new(obj, self, self.isa.text_section_builder(funcs.len()));
+        if self.linkopts.force_jump_veneers {
+            builder.force_veneers();
+        }
+        let mut addrs = AddressMapSection::default();
+        let mut traps = TrapEncodingBuilder::default();
+
+        let mut ret = Vec::with_capacity(funcs.len());
+        for (i, (sym, func)) in funcs.iter().enumerate() {
+            let func = func.downcast_ref::<CompiledFunction>().unwrap();
+            let (sym, range) = builder.append_func(&sym, func, |idx| resolve_reloc(i, idx));
+            if self.tunables.generate_address_map {
+                let addr = func.address_map();
+                addrs.push(range.clone(), &addr.instructions);
+            }
+            traps.push(range.clone(), &func.traps().collect::<Vec<_>>());
+            builder.append_padding(self.linkopts.padding_between_functions);
+            let info = FunctionLoc {
+                start: u32::try_from(range.start).unwrap(),
+                length: u32::try_from(range.end - range.start).unwrap(),
+            };
+            ret.push((sym, info));
+        }
+
+        builder.finish();
+
+        if self.tunables.generate_address_map {
+            addrs.append_to(obj);
+        }
+        traps.append_to(obj);
+
+        Ok(ret)
+    }
+
+    fn emit_trampolines_for_array_call_host_func(
+        &self,
+        ty: &WasmFuncType,
+        host_fn: usize,
+        obj: &mut Object<'static>,
+    ) -> Result<(FunctionLoc, FunctionLoc)> {
+        let mut wasm_to_array = self.wasm_to_array_trampoline(ty, host_fn)?;
+        let mut native_to_array = self.native_to_array_trampoline(ty, host_fn)?;
+
+        let mut builder = ModuleTextBuilder::new(obj, self, self.isa.text_section_builder(2));
+
+        let (_, wasm_to_array) =
+            builder.append_func("wasm_to_array", &mut wasm_to_array, |_| unreachable!());
+        let (_, native_to_array) =
+            builder.append_func("native_to_array", &mut native_to_array, |_| unreachable!());
+
+        let wasm_to_array = FunctionLoc {
+            start: u32::try_from(wasm_to_array.start).unwrap(),
+            length: u32::try_from(wasm_to_array.end - wasm_to_array.start).unwrap(),
+        };
+        let native_to_array = FunctionLoc {
+            start: u32::try_from(native_to_array.start).unwrap(),
+            length: u32::try_from(native_to_array.end - native_to_array.start).unwrap(),
+        };
+
+        builder.finish();
+        Ok((wasm_to_array, native_to_array))
+    }
+
+    fn triple(&self) -> &target_lexicon::Triple {
+        self.isa.triple()
+    }
+
+    fn flags(&self) -> Vec<(&'static str, FlagValue<'static>)> {
+        crate::clif_flags_to_wasmtime(self.isa.flags().iter())
+    }
+
+    fn isa_flags(&self) -> Vec<(&'static str, FlagValue<'static>)> {
+        crate::clif_flags_to_wasmtime(self.isa.isa_flags())
+    }
+
+    fn is_branch_protection_enabled(&self) -> bool {
+        self.isa.is_branch_protection_enabled()
+    }
+
+    #[cfg(feature = "component-model")]
+    fn component_compiler(&self) -> &dyn wasmtime_environ::component::ComponentCompiler {
+        self
+    }
+
+    fn append_dwarf(
+        &self,
+        obj: &mut Object<'_>,
+        translation: &ModuleTranslation<'_>,
+        funcs: &PrimaryMap<DefinedFuncIndex, (SymbolId, &(dyn Any + Send))>,
+    ) -> Result<()> {
+        let ofs = VMOffsets::new(
+            self.isa
+                .triple()
+                .architecture
+                .pointer_width()
+                .unwrap()
+                .bytes(),
+            &translation.module,
+        );
+
+        let memory_offset = if ofs.num_imported_memories > 0 {
+            ModuleMemoryOffset::Imported(ofs.vmctx_vmmemory_import(MemoryIndex::new(0)))
+        } else if ofs.num_defined_memories > 0 {
+            // The addition of shared memory makes the following assumption,
+            // "owned memory index = 0", possibly false. If the first memory
+            // is a shared memory, the base pointer will not be stored in
+            // the `owned_memories` array. The following code should
+            // eventually be fixed to not only handle shared memories but
+            // also multiple memories.
+            assert_eq!(
+                ofs.num_defined_memories, ofs.num_owned_memories,
+                "the memory base pointer may be incorrect due to sharing memory"
+            );
+            ModuleMemoryOffset::Defined(
+                ofs.vmctx_vmmemory_definition_base(OwnedMemoryIndex::new(0)),
+            )
+        } else {
+            ModuleMemoryOffset::None
+        };
+        let functions_info = funcs
+            .iter()
+            .map(|(_, (_, func))| {
+                let f = func.downcast_ref::<CompiledFunction>().unwrap();
+                f.metadata()
+            })
+            .collect();
+        let dwarf_sections = crate::debug::emit_dwarf(
+            &*self.isa,
+            &translation.debuginfo,
+            &functions_info,
+            &memory_offset,
+        )
+        .with_context(|| "failed to emit DWARF debug information")?;
+
+        let (debug_bodies, debug_relocs): (Vec<_>, Vec<_>) = dwarf_sections
+            .iter()
+            .map(|s| ((s.name, &s.body), (s.name, &s.relocs)))
+            .unzip();
+        let mut dwarf_sections_ids = HashMap::new();
+        for (name, body) in debug_bodies {
+            let segment = obj.segment_name(StandardSegment::Debug).to_vec();
+            let section_id = obj.add_section(segment, name.as_bytes().to_vec(), SectionKind::Debug);
+            dwarf_sections_ids.insert(name, section_id);
+            obj.append_section_data(section_id, &body, 1);
+        }
+
+        // Write all debug data relocations.
+        for (name, relocs) in debug_relocs {
+            let section_id = *dwarf_sections_ids.get(name).unwrap();
+            for reloc in relocs {
+                let target_symbol = match reloc.target {
+                    DwarfSectionRelocTarget::Func(index) => funcs[DefinedFuncIndex::new(index)].0,
+                    DwarfSectionRelocTarget::Section(name) => {
+                        obj.section_symbol(dwarf_sections_ids[name])
+                    }
+                };
+                obj.add_relocation(
+                    section_id,
+                    object::write::Relocation {
+                        offset: u64::from(reloc.offset),
+                        symbol: target_symbol,
+                        addend: i64::from(reloc.addend),
+                        flags: RelocationFlags::Generic {
+                            size: reloc.size << 3,
+                            kind: RelocationKind::Absolute,
+                            encoding: RelocationEncoding::Generic,
+                        },
+                    },
+                )?;
+            }
+        }
+
+        Ok(())
+    }
+
+    fn create_systemv_cie(&self) -> Option<gimli::write::CommonInformationEntry> {
+        self.isa.create_systemv_cie()
+    }
+
+    fn compile_wasm_to_builtin(
+        &self,
+        index: BuiltinFunctionIndex,
+    ) -> Result<Box<dyn Any + Send>, CompileError> {
+        let isa = &*self.isa;
+        let ptr_size = isa.pointer_bytes();
+        let pointer_type = isa.pointer_type();
+        let sig = BuiltinFunctionSignatures::new(isa).signature(index);
+
+        let mut compiler = self.function_compiler();
+        let func = ir::Function::with_name_signature(Default::default(), sig.clone());
+        let (mut builder, block0) = compiler.builder(func);
+        let vmctx = builder.block_params(block0)[0];
+
+        // Debug-assert that this is the right kind of vmctx, and then
+        // additionally perform the "routine of the exit trampoline" of saving
+        // fp/pc/etc.
+        debug_assert_vmctx_kind(isa, &mut builder, vmctx, wasmtime_environ::VMCONTEXT_MAGIC);
+        let limits = builder.ins().load(
+            pointer_type,
+            MemFlags::trusted(),
+            vmctx,
+            ptr_size.vmcontext_runtime_limits(),
+        );
+        save_last_wasm_exit_fp_and_pc(&mut builder, pointer_type, &ptr_size, limits);
+
+        // Now it's time to delegate to the actual builtin. Builtins are stored
+        // in an array in all `VMContext`s. First load the base pointer of the
+        // array and then load the entry of the array that correspons to this
+        // builtin.
+        let mem_flags = ir::MemFlags::trusted().with_readonly();
+        let array_addr = builder.ins().load(
+            pointer_type,
+            mem_flags,
+            vmctx,
+            i32::try_from(ptr_size.vmcontext_builtin_functions()).unwrap(),
+        );
+        let body_offset = i32::try_from(index.index() * pointer_type.bytes()).unwrap();
+        let func_addr = builder
+            .ins()
+            .load(pointer_type, mem_flags, array_addr, body_offset);
+
+        // Forward all our own arguments to the libcall itself, and then return
+        // all the same results as the libcall.
+        let block_params = builder.block_params(block0).to_vec();
+        let sig = builder.func.import_signature(sig);
+        let call = builder.ins().call_indirect(sig, func_addr, &block_params);
+        let results = builder.func.dfg.inst_results(call).to_vec();
+        builder.ins().return_(&results);
+        builder.finalize();
+
+        Ok(Box::new(compiler.finish()?))
+    }
+
+    fn compiled_function_relocation_targets<'a>(
+        &'a self,
+        func: &'a dyn Any,
+    ) -> Box<dyn Iterator<Item = RelocationTarget> + 'a> {
+        let func = func.downcast_ref::<CompiledFunction>().unwrap();
+        Box::new(func.relocations().map(|r| r.reloc_target))
+    }
+}
+
+#[cfg(feature = "incremental-cache")]
+mod incremental_cache {
+    use super::*;
+
+    struct CraneliftCacheStore(Arc<dyn CacheStore>);
+
+    impl cranelift_codegen::incremental_cache::CacheKvStore for CraneliftCacheStore {
+        fn get(&self, key: &[u8]) -> Option<std::borrow::Cow<[u8]>> {
+            self.0.get(key)
+        }
+        fn insert(&mut self, key: &[u8], val: Vec<u8>) {
+            self.0.insert(key, val);
+        }
+    }
+
+    pub(super) fn compile_maybe_cached<'a>(
+        context: &'a mut Context,
+        isa: &dyn TargetIsa,
+        cache_ctx: Option<&mut IncrementalCacheContext>,
+    ) -> Result<(&'a CompiledCode, Vec<u8>), CompileError> {
+        let cache_ctx = match cache_ctx {
+            Some(ctx) => ctx,
+            None => return compile_uncached(context, isa),
+        };
+
+        let mut cache_store = CraneliftCacheStore(cache_ctx.cache_store.clone());
+        let (compiled_code, from_cache) = context
+            .compile_with_cache(isa, &mut cache_store, &mut Default::default())
+            .map_err(|error| CompileError::Codegen(pretty_error(&error.func, error.inner)))?;
+
+        if from_cache {
+            cache_ctx.num_hits += 1;
+        } else {
+            cache_ctx.num_cached += 1;
+        }
+
+        Ok((compiled_code, compiled_code.code_buffer().to_vec()))
+    }
+}
+
+#[cfg(feature = "incremental-cache")]
+use incremental_cache::*;
+
+#[cfg(not(feature = "incremental-cache"))]
+fn compile_maybe_cached<'a>(
+    context: &'a mut Context,
+    isa: &dyn TargetIsa,
+    _cache_ctx: Option<&mut IncrementalCacheContext>,
+) -> Result<(&'a CompiledCode, Vec<u8>), CompileError> {
+    compile_uncached(context, isa)
+}
+
+fn compile_uncached<'a>(
+    context: &'a mut Context,
+    isa: &dyn TargetIsa,
+) -> Result<(&'a CompiledCode, Vec<u8>), CompileError> {
+    let mut code_buf = Vec::new();
+    let compiled_code = context
+        .compile_and_emit(isa, &mut code_buf, &mut Default::default())
+        .map_err(|error| CompileError::Codegen(pretty_error(&error.func, error.inner)))?;
+    Ok((compiled_code, code_buf))
+}
+
+impl Compiler {
+    /// Creates a trampoline for calling a host function callee defined with the
+    /// "array" calling convention from a native calling convention caller.
+    ///
+    /// This style of trampoline is used with `Func::new`-style callees and
+    /// `TypedFunc::call`-style callers.
+    ///
+    /// Both callee and caller are on the host side, so there is no host/Wasm
+    /// transition and associated entry/exit state to maintain.
+    ///
+    /// The `host_fn` is a function pointer in this process with the following
+    /// signature:
+    ///
+    /// ```ignore
+    /// unsafe extern "C" fn(*mut VMContext, *mut VMContext, *mut ValRaw, usize)
+    /// ```
+    ///
+    /// where the first two arguments are forwarded from the trampoline
+    /// generated here itself, and the second two arguments are a pointer/length
+    /// into stack-space of this trampoline with storage for both the arguments
+    /// to the function and the results.
+    ///
+    /// Note that `host_fn` is an immediate which is an actual function pointer
+    /// in this process. As such this compiled trampoline is not suitable for
+    /// serialization.
+    fn native_to_array_trampoline(
+        &self,
+        ty: &WasmFuncType,
+        host_fn: usize,
+    ) -> Result<CompiledFunction, CompileError> {
+        let isa = &*self.isa;
+        let pointer_type = isa.pointer_type();
+        let native_call_sig = native_call_signature(isa, ty);
+        let array_call_sig = array_call_signature(isa);
+
+        let mut compiler = self.function_compiler();
+        let func = ir::Function::with_name_signature(Default::default(), native_call_sig);
+        let (mut builder, block0) = compiler.builder(func);
+        let args = builder.func.dfg.block_params(block0).to_vec();
+
+        let ret = NativeRet::classify(isa, ty);
+        let wasm_args = &ret.native_args(&args)[2..];
+
+        let (values_vec_ptr, values_vec_len) =
+            self.allocate_stack_array_and_spill_args(ty, &mut builder, wasm_args);
+        let values_vec_len = builder
+            .ins()
+            .iconst(pointer_type, i64::from(values_vec_len));
+
+        let callee_args = [args[0], args[1], values_vec_ptr, values_vec_len];
+
+        let new_sig = builder.import_signature(array_call_sig);
+        let callee_value = builder.ins().iconst(pointer_type, host_fn as i64);
+        builder
+            .ins()
+            .call_indirect(new_sig, callee_value, &callee_args);
+
+        let results =
+            self.load_values_from_array(ty.returns(), &mut builder, values_vec_ptr, values_vec_len);
+        ret.native_return(&mut builder, block0, &results);
+        builder.finalize();
+
+        compiler.finish()
+    }
+
+    /// Creates a trampoline for WebAssembly to call a host function defined
+    /// with the "array" calling convention: where all the arguments are spilled
+    /// to an array on the stack and results are loaded from the stack array.
+    ///
+    /// This style of trampoline is currently only used with the
+    /// `Func::new`-style created functions in the Wasmtime embedding API. The
+    /// generated trampoline has a function signature appropriate to the `ty`
+    /// specified (e.g. a System-V ABI) and will call a `host_fn` that has a
+    /// type signature of:
+    ///
+    /// ```ignore
+    /// unsafe extern "C" fn(*mut VMContext, *mut VMContext, *mut ValRaw, usize)
+    /// ```
+    ///
+    /// where the first two arguments are forwarded from the trampoline
+    /// generated here itself, and the second two arguments are a pointer/length
+    /// into stack-space of this trampoline with storage for both the arguments
+    /// to the function and the results.
+    ///
+    /// Note that `host_fn` is an immediate which is an actual function pointer
+    /// in this process, and `limits` is a pointer to `VMRuntimeLimits`. As such
+    /// this compiled trampoline is not suitable for serialization, and only
+    /// valid for a particular store.
+    fn wasm_to_array_trampoline(
+        &self,
+        ty: &WasmFuncType,
+        host_fn: usize,
+    ) -> Result<CompiledFunction, CompileError> {
+        let isa = &*self.isa;
+        let pointer_type = isa.pointer_type();
+        let wasm_call_sig = wasm_call_signature(isa, ty, &self.tunables);
+        let array_call_sig = array_call_signature(isa);
+
+        let mut compiler = self.function_compiler();
+        let func = ir::Function::with_name_signature(Default::default(), wasm_call_sig);
+        let (mut builder, block0) = compiler.builder(func);
+        let args = builder.func.dfg.block_params(block0).to_vec();
+        let caller_vmctx = args[1];
+
+        // Assert that we were really given a core Wasm vmctx, since that's
+        // what we are assuming with our offsets below.
+        debug_assert_vmctx_kind(
+            isa,
+            &mut builder,
+            caller_vmctx,
+            wasmtime_environ::VMCONTEXT_MAGIC,
+        );
+        let ptr_size = isa.pointer_bytes();
+        let limits = builder.ins().load(
+            pointer_type,
+            MemFlags::trusted(),
+            caller_vmctx,
+            ptr_size.vmcontext_runtime_limits(),
+        );
+        save_last_wasm_exit_fp_and_pc(&mut builder, pointer_type, &ptr_size, limits);
+
+        let (values_vec_ptr, values_vec_len) =
+            self.allocate_stack_array_and_spill_args(ty, &mut builder, &args[2..]);
+        let values_vec_len = builder
+            .ins()
+            .iconst(pointer_type, i64::from(values_vec_len));
+
+        let block_params = builder.func.dfg.block_params(block0);
+        let callee_args = [
+            block_params[0],
+            block_params[1],
+            values_vec_ptr,
+            values_vec_len,
+        ];
+
+        let new_sig = builder.import_signature(array_call_sig);
+        let callee_value = builder.ins().iconst(pointer_type, host_fn as i64);
+        builder
+            .ins()
+            .call_indirect(new_sig, callee_value, &callee_args);
+
+        let results =
+            self.load_values_from_array(ty.returns(), &mut builder, values_vec_ptr, values_vec_len);
+        builder.ins().return_(&results);
+        builder.finalize();
+
+        compiler.finish()
+    }
+
+    /// This function will allocate a stack slot suitable for storing both the
+    /// arguments and return values of the function, and then the arguments will
+    /// all be stored in this block.
+    ///
+    /// `block0` must be the entry block of the function and `ty` must be the
+    /// Wasm function type of the trampoline.
+    ///
+    /// The stack slot pointer is returned in addition to the size, in units of
+    /// `ValRaw`, of the stack slot.
+    fn allocate_stack_array_and_spill_args(
+        &self,
+        ty: &WasmFuncType,
+        builder: &mut FunctionBuilder,
+        args: &[ir::Value],
+    ) -> (Value, u32) {
+        let isa = &*self.isa;
+        let pointer_type = isa.pointer_type();
+
+        // Compute the size of the values vector.
+        let value_size = mem::size_of::<u128>();
+        let values_vec_len = cmp::max(ty.params().len(), ty.returns().len());
+        let values_vec_byte_size = u32::try_from(value_size * values_vec_len).unwrap();
+        let values_vec_len = u32::try_from(values_vec_len).unwrap();
+
+        let slot = builder.func.create_sized_stack_slot(ir::StackSlotData::new(
+            ir::StackSlotKind::ExplicitSlot,
+            values_vec_byte_size,
+        ));
+        let values_vec_ptr = builder.ins().stack_addr(pointer_type, slot, 0);
+
+        {
+            let values_vec_len = builder
+                .ins()
+                .iconst(ir::types::I32, i64::try_from(values_vec_len).unwrap());
+            self.store_values_to_array(builder, ty.params(), args, values_vec_ptr, values_vec_len);
+        }
+
+        (values_vec_ptr, values_vec_len)
+    }
+
+    /// Store values to an array in the array calling convention.
+    ///
+    /// Used either to store arguments to the array when calling a function
+    /// using the array calling convention, or used to store results to the
+    /// array when implementing a function that exposes the array calling
+    /// convention.
+    fn store_values_to_array(
+        &self,
+        builder: &mut FunctionBuilder,
+        types: &[WasmValType],
+        values: &[Value],
+        values_vec_ptr: Value,
+        values_vec_capacity: Value,
+    ) {
+        debug_assert_eq!(types.len(), values.len());
+        debug_assert_enough_capacity_for_length(builder, types.len(), values_vec_capacity);
+
+        // Note that loads and stores are unconditionally done in the
+        // little-endian format rather than the host's native-endianness,
+        // despite this load/store being unrelated to execution in wasm itself.
+        // For more details on this see the `ValRaw` type in the
+        // `wasmtime-runtime` crate.
+        let flags = ir::MemFlags::new()
+            .with_notrap()
+            .with_endianness(ir::Endianness::Little);
+
+        let value_size = mem::size_of::<u128>();
+        for (i, (val, ty)) in values.iter().copied().zip(types).enumerate() {
+            crate::unbarriered_store_type_at_offset(
+                &*self.isa,
+                &mut builder.cursor(),
+                *ty,
+                flags,
+                values_vec_ptr,
+                i32::try_from(i * value_size).unwrap(),
+                val,
+            );
+        }
+    }
+
+    /// Used for loading the values of an array-call host function's value
+    /// array.
+    ///
+    /// This can be used to load arguments out of the array if the trampoline we
+    /// are building exposes the array calling convention, or it can be used to
+    /// laod results out of the array if the trampoline we are building calls a
+    /// function that uses the array calling convention.
+    fn load_values_from_array(
+        &self,
+        types: &[WasmValType],
+        builder: &mut FunctionBuilder,
+        values_vec_ptr: Value,
+        values_vec_capacity: Value,
+    ) -> Vec<ir::Value> {
+        let isa = &*self.isa;
+        let value_size = mem::size_of::<u128>();
+
+        debug_assert_enough_capacity_for_length(builder, types.len(), values_vec_capacity);
+
+        // Note that this is little-endian like `store_values_to_array` above,
+        // see notes there for more information.
+        let flags = MemFlags::new()
+            .with_notrap()
+            .with_endianness(ir::Endianness::Little);
+
+        let mut results = Vec::new();
+        for (i, ty) in types.iter().enumerate() {
+            results.push(crate::unbarriered_load_type_at_offset(
+                isa,
+                &mut builder.cursor(),
+                *ty,
+                flags,
+                values_vec_ptr,
+                i32::try_from(i * value_size).unwrap(),
+            ));
+        }
+        results
+    }
+
+    fn function_compiler(&self) -> FunctionCompiler<'_> {
+        let saved_context = self.contexts.lock().unwrap().pop();
+        FunctionCompiler {
+            compiler: self,
+            cx: saved_context
+                .map(|mut ctx| {
+                    ctx.codegen_context.clear();
+                    ctx
+                })
+                .unwrap_or_else(|| CompilerContext {
+                    #[cfg(feature = "incremental-cache")]
+                    incremental_cache_ctx: self.cache_store.as_ref().map(|cache_store| {
+                        IncrementalCacheContext {
+                            cache_store: cache_store.clone(),
+                            num_hits: 0,
+                            num_cached: 0,
+                        }
+                    }),
+                    ..Default::default()
+                }),
+        }
+    }
+}
+
+struct FunctionCompiler<'a> {
+    compiler: &'a Compiler,
+    cx: CompilerContext,
+}
+
+impl FunctionCompiler<'_> {
+    fn builder(&mut self, func: ir::Function) -> (FunctionBuilder<'_>, ir::Block) {
+        self.cx.codegen_context.func = func;
+        let mut builder = FunctionBuilder::new(
+            &mut self.cx.codegen_context.func,
+            self.cx.func_translator.context(),
+        );
+
+        let block0 = builder.create_block();
+        builder.append_block_params_for_function_params(block0);
+        builder.switch_to_block(block0);
+        builder.seal_block(block0);
+        (builder, block0)
+    }
+
+    fn finish(self) -> Result<CompiledFunction, CompileError> {
+        let (info, func) = self.finish_with_info(None)?;
+        assert!(info.stack_maps.is_empty());
+        Ok(func)
+    }
+
+    fn finish_with_info(
+        mut self,
+        body_and_tunables: Option<(&FunctionBody<'_>, &Tunables)>,
+    ) -> Result<(WasmFunctionInfo, CompiledFunction), CompileError> {
+        let context = &mut self.cx.codegen_context;
+        let isa = &*self.compiler.isa;
+        let (_, _code_buf) =
+            compile_maybe_cached(context, isa, self.cx.incremental_cache_ctx.as_mut())?;
+        let compiled_code = context.compiled_code().unwrap();
+
+        // Give wasm functions, user defined code, a "preferred" alignment
+        // instead of the minimum alignment as this can help perf in niche
+        // situations.
+        let preferred_alignment = if body_and_tunables.is_some() {
+            self.compiler.isa.function_alignment().preferred
+        } else {
+            1
+        };
+
+        let alignment = compiled_code.buffer.alignment.max(preferred_alignment);
+        let mut compiled_function = CompiledFunction::new(
+            compiled_code.buffer.clone(),
+            context.func.params.user_named_funcs().clone(),
+            alignment,
+        );
+
+        if let Some((body, tunables)) = body_and_tunables {
+            let data = body.get_binary_reader();
+            let offset = data.original_position();
+            let len = data.bytes_remaining();
+            compiled_function.set_address_map(
+                offset as u32,
+                len as u32,
+                tunables.generate_address_map,
+            );
+        }
+
+        if isa.flags().unwind_info() {
+            let unwind = compiled_code
+                .create_unwind_info(isa)
+                .map_err(|error| CompileError::Codegen(pretty_error(&context.func, error)))?;
+
+            if let Some(unwind_info) = unwind {
+                compiled_function.set_unwind_info(unwind_info);
+            }
+        }
+
+        if body_and_tunables
+            .map(|(_, t)| t.generate_native_debuginfo)
+            .unwrap_or(false)
+        {
+            compiled_function.set_value_labels_ranges(compiled_code.value_labels_ranges.clone());
+
+            // DWARF debugging needs the CFA-based unwind information even on Windows.
+            if !matches!(
+                compiled_function.metadata().unwind_info,
+                Some(UnwindInfo::SystemV(_))
+            ) {
+                let cfa_unwind = compiled_code
+                    .create_unwind_info_of_kind(isa, UnwindInfoKind::SystemV)
+                    .map_err(|error| CompileError::Codegen(pretty_error(&context.func, error)))?;
+
+                if let Some(UnwindInfo::SystemV(cfa_unwind_info)) = cfa_unwind {
+                    compiled_function.set_cfa_unwind_info(cfa_unwind_info);
+                }
+            }
+        }
+
+        let stack_maps = mach_stack_maps_to_stack_maps(compiled_code.buffer.stack_maps());
+        compiled_function
+            .set_sized_stack_slots(std::mem::take(&mut context.func.sized_stack_slots));
+        self.compiler.contexts.lock().unwrap().push(self.cx);
+
+        Ok((
+            WasmFunctionInfo {
+                start_srcloc: compiled_function.metadata().address_map.start_srcloc,
+                stack_maps: stack_maps.into(),
+            },
+            compiled_function,
+        ))
+    }
+}
+
+fn mach_stack_maps_to_stack_maps(mach_stack_maps: &[MachStackMap]) -> Vec<StackMapInformation> {
+    // This is converting from Cranelift's representation of a stack map to
+    // Wasmtime's representation. They happen to align today but that may
+    // not always be true in the future.
+    let mut stack_maps = Vec::new();
+    for &MachStackMap {
+        offset_end,
+        ref stack_map,
+        ..
+    } in mach_stack_maps
+    {
+        let stack_map = wasmtime_environ::StackMap::new(
+            stack_map.mapped_words(),
+            stack_map.as_slice().iter().map(|a| a.0),
+        );
+        stack_maps.push(StackMapInformation {
+            code_offset: offset_end,
+            stack_map,
+        });
+    }
+    stack_maps.sort_unstable_by_key(|info| info.code_offset);
+    stack_maps
+}
+
+fn declare_and_call(
+    builder: &mut FunctionBuilder,
+    signature: ir::Signature,
+    func_index: u32,
+    args: &[ir::Value],
+) -> ir::Inst {
+    let name = ir::ExternalName::User(builder.func.declare_imported_user_function(
+        ir::UserExternalName {
+            namespace: crate::NS_WASM_FUNC,
+            index: func_index,
+        },
+    ));
+    let signature = builder.func.import_signature(signature);
+    let callee = builder.func.dfg.ext_funcs.push(ir::ExtFuncData {
+        name,
+        signature,
+        colocated: true,
+    });
+    builder.ins().call(callee, &args)
+}
+
+fn debug_assert_enough_capacity_for_length(
+    builder: &mut FunctionBuilder,
+    length: usize,
+    capacity: ir::Value,
+) {
+    if cfg!(debug_assertions) {
+        let enough_capacity = builder.ins().icmp_imm(
+            ir::condcodes::IntCC::UnsignedGreaterThanOrEqual,
+            capacity,
+            ir::immediates::Imm64::new(length.try_into().unwrap()),
+        );
+        builder
+            .ins()
+            .trapz(enough_capacity, ir::TrapCode::User(DEBUG_ASSERT_TRAP_CODE));
+    }
+}
+
+fn debug_assert_vmctx_kind(
+    isa: &dyn TargetIsa,
+    builder: &mut FunctionBuilder,
+    vmctx: ir::Value,
+    expected_vmctx_magic: u32,
+) {
+    if cfg!(debug_assertions) {
+        let magic = builder.ins().load(
+            ir::types::I32,
+            MemFlags::trusted().with_endianness(isa.endianness()),
+            vmctx,
+            0,
+        );
+        let is_expected_vmctx = builder.ins().icmp_imm(
+            ir::condcodes::IntCC::Equal,
+            magic,
+            i64::from(expected_vmctx_magic),
+        );
+        builder.ins().trapz(
+            is_expected_vmctx,
+            ir::TrapCode::User(DEBUG_ASSERT_TRAP_CODE),
+        );
+    }
+}
+
+fn save_last_wasm_entry_sp(
+    builder: &mut FunctionBuilder,
+    pointer_type: ir::Type,
+    ptr_size: &impl PtrSize,
+    vm_runtime_limits_offset: u32,
+    vmctx: Value,
+) {
+    // First we need to get the `VMRuntimeLimits`.
+    let limits = builder.ins().load(
+        pointer_type,
+        MemFlags::trusted(),
+        vmctx,
+        i32::try_from(vm_runtime_limits_offset).unwrap(),
+    );
+
+    // Then store our current stack pointer into the appropriate slot.
+    let sp = builder.ins().get_stack_pointer(pointer_type);
+    builder.ins().store(
+        MemFlags::trusted(),
+        sp,
+        limits,
+        ptr_size.vmruntime_limits_last_wasm_entry_sp(),
+    );
+}
+
+fn save_last_wasm_exit_fp_and_pc(
+    builder: &mut FunctionBuilder,
+    pointer_type: ir::Type,
+    ptr: &impl PtrSize,
+    limits: Value,
+) {
+    // Save the exit Wasm FP to the limits. We dereference the current FP to get
+    // the previous FP because the current FP is the trampoline's FP, and we
+    // want the Wasm function's FP, which is the caller of this trampoline.
+    let trampoline_fp = builder.ins().get_frame_pointer(pointer_type);
+    let wasm_fp = builder.ins().load(
+        pointer_type,
+        MemFlags::trusted(),
+        trampoline_fp,
+        // The FP always points to the next older FP for all supported
+        // targets. See assertion in
+        // `crates/runtime/src/traphandlers/backtrace.rs`.
+        0,
+    );
+    builder.ins().store(
+        MemFlags::trusted(),
+        wasm_fp,
+        limits,
+        ptr.vmruntime_limits_last_wasm_exit_fp(),
+    );
+    // Finally save the Wasm return address to the limits.
+    let wasm_pc = builder.ins().get_return_address(pointer_type);
+    builder.ins().store(
+        MemFlags::trusted(),
+        wasm_pc,
+        limits,
+        ptr.vmruntime_limits_last_wasm_exit_pc(),
+    );
+}
+
+enum NativeRet {
+    Bare,
+    Retptr {
+        slots: Vec<(i32, ir::Type)>,
+        size: u32,
+    },
+}
+
+impl NativeRet {
+    fn classify(isa: &dyn TargetIsa, ty: &WasmFuncType) -> NativeRet {
+        fn align_to(val: i32, align: i32) -> i32 {
+            (val + (align - 1)) & !(align - 1)
+        }
+
+        match ty.returns() {
+            [] | [_] => NativeRet::Bare,
+            other => {
+                let mut offset = 0;
+                let mut offsets = Vec::new();
+                let mut max_align = 1;
+                for ty in other[1..].iter() {
+                    let ty = crate::value_type(isa, *ty);
+                    let size = ty.bytes();
+                    let size = i32::try_from(size).unwrap();
+                    offset = align_to(offset, size);
+                    offsets.push((offset, ty));
+                    offset += size;
+                    max_align = max_align.max(size);
+                }
+                NativeRet::Retptr {
+                    slots: offsets,
+                    size: u32::try_from(align_to(offset, max_align)).unwrap(),
+                }
+            }
+        }
+    }
+
+    fn native_args<'a>(&self, args: &'a [ir::Value]) -> &'a [ir::Value] {
+        match self {
+            NativeRet::Bare => args,
+            NativeRet::Retptr { .. } => &args[..args.len() - 1],
+        }
+    }
+
+    fn native_return(
+        &self,
+        builder: &mut FunctionBuilder<'_>,
+        block0: ir::Block,
+        results: &[ir::Value],
+    ) {
+        match self {
+            NativeRet::Bare => {
+                builder.ins().return_(&results);
+            }
+            NativeRet::Retptr { slots, .. } => {
+                let ptr = *builder.func.dfg.block_params(block0).last().unwrap();
+                let (first, rest) = results.split_first().unwrap();
+                assert_eq!(rest.len(), slots.len());
+                for (arg, (offset, ty)) in rest.iter().zip(slots) {
+                    assert_eq!(builder.func.dfg.value_type(*arg), *ty);
+                    builder.ins().store(MemFlags::trusted(), *arg, ptr, *offset);
+                }
+                builder.ins().return_(&[*first]);
+            }
+        }
+    }
+}