[Wasm RyuJIT] Add Instruction Encoding for Local Var Declarations and Emit Local Declarations #122425

adamperlin · 2025-12-11T00:53:30Z

This PR adds a new instrDesc subtype for encoding Wasm local declarations, which have the form count: u32 type:valtype, where valtype can be encoded as a single byte for number and vector types.

The PR also adds code to emit local count and local declarations as part of the prolog for a function. Some of this may need to change based on what we decide around calling convention and register (locals) allocation.

…al JIT, since the RyuJIT backend is now emitting Wasm

…-arg-initialization

dotnet-policy-service · 2025-12-11T00:54:43Z

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

Copilot

Pull request overview

This PR introduces support for WebAssembly local variable declarations in the JIT compiler by adding a new instruction descriptor subtype and emission logic for encoding local declarations in the function prolog.

Key Changes

Adds instWasmValueType enum and mapping functions to convert between JIT types and WASM types
Introduces instrDescLclVarDecl instruction descriptor for encoding local declarations with count and type
Implements local declaration emission in the function prolog via genWasmLocals()

Reviewed changes

Copilot reviewed 10 out of 10 changed files in this pull request and generated 12 comments.

Show a summary per file

File	Description
src/coreclr/jit/wasmtypesdef.h	New header defining WASM value types and type conversion functions
src/coreclr/jit/registeropswasm.h	Removes trailing blank line
src/coreclr/jit/instrswasm.h	Adds `local_cnt` and `local_decl` instructions
src/coreclr/jit/emitwasm.h	Declares new emission methods for local declarations
src/coreclr/jit/emitwasm.cpp	Implements local declaration emission and descriptor handling
src/coreclr/jit/emitfmtswasm.h	Adds `IF_LOCAL_CNT` and `IF_LOCAL_DECL` instruction formats
src/coreclr/jit/emit.h	Adds `instrDescLclVarDecl` structure and bitfield flag
src/coreclr/jit/codegenwasm.cpp	Adds `INS_end` instruction to epilog
src/coreclr/jit/codegencommon.cpp	Implements prolog local count and declaration emission
src/coreclr/jit/codegen.h	Declares `genWasmLocals()` method

src/coreclr/jit/emitwasm.cpp

src/coreclr/jit/codegencommon.cpp

src/coreclr/jit/emit.h

src/coreclr/jit/wasmtypesdef.h

src/coreclr/jit/emitfmtswasm.h

src/coreclr/jit/emitwasm.cpp

src/coreclr/jit/wasmtypesdef.h

src/coreclr/jit/emitwasm.cpp

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

src/coreclr/jit/codegen.h

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

adamperlin · 2025-12-11T17:31:07Z

src/coreclr/jit/codegencommon.cpp

+    for (unsigned i = 0; i < compiler->info.compLocalsCount - compiler->info.compArgsCount; i++)
+    {
+        LclVarDsc* varDsc = compiler->lvaGetDesc(i);
+        assert(!varDsc->lvIsParam);


TODO: This assert probably isn't necessary.

Looks like you are trying to enumerate the IL visible locals?

The jit can generate new locals for lots of reasons, and these may also end up as Wasm locals. So this emission needs to examine all the non-param locals.

Also the wasm register number will not be the local number, typically. So if we have to emit these in increasing wasm register number then more work is needed; we need to walk all the locals and collect their (reg,type) pairs, then sort by reg, then emit these.

Or maybe we can bias the register allocation to make this emission simpler?

If not, we probably need something like:

// set up a collection object jitstd::vector<...> localsig (...); // find enregistered locals -- collect reg and type for (unsigned i = 0; i < lvaCount; i++) { LclVarDsc* varDsc = compiler->lvaGetDesc(i); if (varDsc->IsInReg() && !varDsc->lvIsParam) { localsig.push_back(UnpackWasmReg(varDsc->GetRegNum()), emitter::instWasmValueType type = GetEmitter()->genWasmTypeFromVarType(varDsc->TypeGet()) } } // sort collection by increasing reg jitstd::sort(...); // paranoia: verify no dups or holes in the ordering // emit types in reg order for (... : localsig) { emit (type) }

Thanks for the feedback. I think I will try and implement something like you are suggesting re: enumerating all locals and ensuring that they emitted in sorted register order. Then maybe later, depending on the register allocation implementation, we can come back and change how this is handled?

This will need RA data structures to communicate the full set of used locals (the algorithm above won't discover locals used as e. g. temporaries). I would suggest doing something like emitting one local of each value type for now to validate the emitter APIs and we can add the real local counts later.

So you're suggesting we just declare locals that we never actually refer to, for now?

adamperlin · 2025-12-11T17:32:00Z

src/coreclr/jit/emit.h

+    struct instrDescLclVarDecl : instrDesc
+    {
+        instrDescLclVarDecl() = delete;
+        cnsval_ssize_t    lclCnt;


TODO: is it possible to reuse the constant field inside the parent instr desc for small size constants?

AndyAyersMS · 2025-12-11T20:13:42Z

src/coreclr/jit/wasmtypesdef.h

+        case TYP_INT:
+        case TYP_REF:
+        case TYP_BYREF:
+            return instWasmValueType::I32;


I think we're trying to stay flexible and not hard-wire to Wasm32.

For Wasm64 REF/BYREF will be I64.

AndyAyersMS · 2025-12-11T20:17:36Z

src/coreclr/jit/codegenwasm.cpp

    // TODO-WASM-CQ: do not emit "return" in case this is the last block

    instGen(INS_return);
+    instGen(INS_end);


Is this end required at the end of methods? We can have multiple epilogs (though we might want to reconsider)... and there's no guarantee even with one epilog that the epilog will be at the end of the instruction stream.

Yeah it is required. If the epilog isn't guaranteed to be the end, where would be a safe place to ensure that the end comes at the actual end of the instruction stream for a method?

Probably at the end of genCodeForBBlist or similar.

genFnEpilog is a BasicBlock* parameter. I am pretty sure you can just check block->IsLast() and switch whether you return return or end based on that. That would address the TODO above as well.

It might not be the actual last block when we have funclets.

That's a good point. This case can be detected by checking if the next block is a funclet entry.

Adding this at the end of genCodeForBBList proved a little bit tricky because of instruction groups. It does look like we track the last used instruction group and could grab that and append after we've finished all method code? Will the last used group be guaranteed to be at the end of the method though? Also, what is the best way to detect if the next block is a funclet? For now I've added an IsLast check to genFnEpilog with a comment that we need to revisit if we're doing funclets.

AndyAyersMS · 2025-12-11T20:34:27Z

src/coreclr/jit/codegencommon.cpp

+    for (unsigned i = 0; i < compiler->info.compLocalsCount - compiler->info.compArgsCount; i++)
+    {
+        LclVarDsc* varDsc = compiler->lvaGetDesc(i);
+        assert(!varDsc->lvIsParam);


Looks like you are trying to enumerate the IL visible locals?

The jit can generate new locals for lots of reasons, and these may also end up as Wasm locals. So this emission needs to examine all the non-param locals.

Also the wasm register number will not be the local number, typically. So if we have to emit these in increasing wasm register number then more work is needed; we need to walk all the locals and collect their (reg,type) pairs, then sort by reg, then emit these.

Or maybe we can bias the register allocation to make this emission simpler?

If not, we probably need something like:

// set up a collection object jitstd::vector<...> localsig (...); // find enregistered locals -- collect reg and type for (unsigned i = 0; i < lvaCount; i++) { LclVarDsc* varDsc = compiler->lvaGetDesc(i); if (varDsc->IsInReg() && !varDsc->lvIsParam) { localsig.push_back(UnpackWasmReg(varDsc->GetRegNum()), emitter::instWasmValueType type = GetEmitter()->genWasmTypeFromVarType(varDsc->TypeGet()) } } // sort collection by increasing reg jitstd::sort(...); // paranoia: verify no dups or holes in the ordering // emit types in reg order for (... : localsig) { emit (type) }

src/coreclr/jit/wasmtypesdef.h

src/coreclr/jit/emit.h

…pilog

SingleAccretion · 2025-12-13T18:25:19Z

src/coreclr/jit/codegenwasm.cpp

    instGen(INS_return);
+    // TODO-WASM: this condition will not be sufficient if we have funclet to determine if we're at the end of
+    // codegen for a method. Revisit later.
+    if (block->IsLast()) 
+    {
+        instGen(INS_end);
+    }


Suggested change

instGen(INS_return);

// TODO-WASM: this condition will not be sufficient if we have funclet to determine if we're at the end of

// codegen for a method. Revisit later.

if (block->IsLast())

{

instGen(INS_end);

}

if (block->IsLast() || bbIsFuncletBeg(block->Next()))

{

instGen(INS_end);

}

else

{

instGen(INS_return);

}

(And remove the TODO above).

Notably, this is still not quite enough because the method can end up on a non-epilogue-generating basic block (e. g. an infinite loop, or an always-throw call). To handle those cases we will need to add the logical equivalent of:

if (isEndOfFunctionOrFunclet(block) && !inspiredEpilogue(block)) instGen(INS_end);

To genEmitEndBlock. But no need to do this in this change.

SingleAccretion · 2025-12-13T18:25:41Z

src/coreclr/jit/codegenlinear.cpp

        // offset 21:
        //
-        //          (  0,  0) [000040] ------------                il_offset void   IL offset: 21
+        //          (  0,  0) [000040] ------------   /             il_offset void   IL offset: 21


Intentional?

SingleAccretion · 2025-12-13T18:25:59Z

src/coreclr/jit/codegencommon.cpp

+// genWasmLocals: generate wasm locals for all locals
+//
+// TODO-WASM: pre-declare all "register" locals
+void CodeGen::genWasmLocals()


This should be in codegenwasm.cpp.

SingleAccretion · 2025-12-13T18:26:20Z

src/coreclr/jit/codegencommon.cpp

-    // TODO-WASM: prolog zeroing, shadow stack maintenance
+    // TODO-WASM: proper local count, local declarations, and shadow stack maintenance
+    assert(compiler->info.compLocalsCount >= compiler->info.compArgsCount);
+    GetEmitter()->emitIns_I(INS_local_cnt, EA_8BYTE, (unsigned)WasmValueType::Count - 1);


This should also be moved into genWasmLocals. Also, no need for the assert.

SingleAccretion · 2025-12-13T18:26:56Z

src/coreclr/jit/codegen.h

 #include "jitgcinfo.h"

+#if defined(TARGET_WASM)
+#include "wasmtypesdef.h"


I think you missed deleting this file (now unused).

SingleAccretion · 2025-12-13T18:35:50Z

src/coreclr/jit/instrswasm.h

+INST(local_cnt,   "local.cnt",   0, IF_LOCAL_CNT, 0x00)
+INST(local_decl,  "local",       0, IF_LOCAL_DECL, 0x00)


Nit: maybe move these together with label, so that all our "pseudos" are together?

SingleAccretion · 2025-12-13T18:36:31Z

src/coreclr/jit/emitwasm.cpp

+    instrDesc* id  = this->emitNewInstrLclVarDecl(EA_8BYTE, imm, valType);
+    insFormat  fmt = this->emitInsFormat(ins);
+
+    id->idIns(ins);
+    id->idInsFmt(fmt);
+
+    this->dispIns(id);
+    this->appendToCurIG(id);


Remove all this->?

SingleAccretion · 2025-12-13T18:45:41Z

src/coreclr/jit/emitwasm.cpp

+            cnsval_ssize_t    imm     = emitGetLclVarDeclCount(id);
+            WasmValueType  valType = emitGetLclVarDeclType(id);
+            printf(" %llu %s", (uint64_t)imm, WasmValueTypeName(valType));


I've been trying to keep this disasm format the same as wasm-objdump; the wasm-objdump way to display this would be like so:

1c08cc: 14 7f | local[6..25] type=i32 1c08ce: 01 7e | local[26] type=i64 1c08d0: 01 7f | local[27] type=i32

It is useful since it also lists the indices instead of the raw encoding. However, here we only have counts. The easiest way to fix that would be to add another field to idDebugOnlyInfo that'd store the "base" index.

SingleAccretion · 2025-12-13T18:47:29Z

src/coreclr/jit/registeropswasm.h

+    "Count",
+};
+
+inline const char* WasmValueTypeName(WasmValueType type)


This (implementation) should be moved to registeropswasm.cpp (and WasmValueTypeNames can be made a static array local to the function). In general it is an anti-pattern to have arrays of data in headers (since they get duplicated into each object file and may even get duplicated in the binary depending on how careful are you with linkage).

SingleAccretion · 2025-12-13T18:48:31Z

src/coreclr/jit/registeropswasm.h

+
+inline const char* WasmValueTypeName(WasmValueType type)
+{
+    static_assert(ArrLen(WasmValueTypeNames) == static_cast<unsigned>(WasmValueType::Count) + 1);


We should never see a WasmValueType type equal to WasmValueType::Count. You can exclude that from printing (e. g. have an assert that type < WasmValueType::Count.

adamperlin added 6 commits December 8, 2025 20:20

Revert wasm method stub method in crossgen to allow invoking the actu…

e93e57d

…al JIT, since the RyuJIT backend is now emitting Wasm

WIP add args as locals to the beginning of method epilog in Wasm codegen

4b991da

Multiple fixes for emitting local declarations for incoming args

5933e67

jit-format

7b8e339

Merge branch 'main' of github.com:dotnet/runtime into adamperlin/wasm…

78b7c70

…-arg-initialization

Generate locals only for non-args

9fbfb86

Copilot AI review requested due to automatic review settings December 11, 2025 00:53

github-actions bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Dec 11, 2025

dotnet-policy-service bot assigned adamperlin Dec 11, 2025

Copilot started reviewing on behalf of adamperlin December 11, 2025 00:54 View session

Copilot AI reviewed Dec 11, 2025

View reviewed changes

adamperlin changed the title ~~Add Instruction Encoding for Local Var Declarations and Emit Local Declarations~~ [Wasm RyuJIT] Add Instruction Encoding for Local Var Declarations and Emit Local Declarations Dec 11, 2025

adamperlin and others added 5 commits December 11, 2025 09:16

Update src/coreclr/jit/emit.h

7d12af0

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Update src/coreclr/jit/emitfmtswasm.h

4b91b3a

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Update src/coreclr/jit/codegencommon.cpp

183851f

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Update src/coreclr/jit/emitwasm.cpp

c9d3046

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Remove incorrect noway_asserts

0482ee3

adamperlin commented Dec 11, 2025

View reviewed changes

src/coreclr/jit/codegen.h Show resolved Hide resolved

adamperlin and others added 2 commits December 11, 2025 09:22

Remove empty ifdef block

dfb0627

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Apply some copilot review suggestions

b237c17

adamperlin commented Dec 11, 2025

View reviewed changes

AndyAyersMS reviewed Dec 11, 2025

View reviewed changes

SingleAccretion reviewed Dec 11, 2025

View reviewed changes

src/coreclr/jit/wasmtypesdef.h Show resolved Hide resolved

SingleAccretion reviewed Dec 11, 2025

View reviewed changes

src/coreclr/jit/emit.h Outdated Show resolved Hide resolved

build-analysis bot mentioned this pull request Dec 11, 2025

"Over release of MOW - COM" assert in COM wrappers #122447

Closed

adamperlin added 3 commits December 12, 2025 10:04

Address some review feedback

19c9eaf

Finish addressing review feedback

76a15f2

Fix instGen placement

d32f695

SingleAccretion self-requested a review December 12, 2025 23:08

adamperlin added 2 commits December 12, 2025 18:45

Add block->IsLast() guard for emitting end instruction in Wasm genFnE…

f3e2f82

…pilog

Emit correct local count for current state of Wasm codegen

0820a24

SingleAccretion reviewed Dec 13, 2025

View reviewed changes

		INST(local_cnt, "local.cnt", 0, IF_LOCAL_CNT, 0x00)
		INST(local_decl, "local", 0, IF_LOCAL_DECL, 0x00)

[Wasm RyuJIT] Add Instruction Encoding for Local Var Declarations and Emit Local Declarations #122425

Are you sure you want to change the base?

[Wasm RyuJIT] Add Instruction Encoding for Local Var Declarations and Emit Local Declarations #122425

Uh oh!

Conversation

adamperlin commented Dec 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

dotnet-policy-service bot commented Dec 11, 2025

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Key Changes

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

adamperlin Dec 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SingleAccretion Dec 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

adamperlin Dec 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

SingleAccretion Dec 11, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

adamperlin Dec 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

SingleAccretion Dec 13, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

adamperlin commented Dec 11, 2025 •

edited

Loading

adamperlin Dec 11, 2025 •

edited

Loading

SingleAccretion Dec 11, 2025 •

edited

Loading

adamperlin Dec 11, 2025 •

edited

Loading

SingleAccretion Dec 11, 2025 •

edited

Loading

adamperlin Dec 13, 2025 •

edited

Loading

SingleAccretion Dec 13, 2025 •

edited

Loading