[Wasm RyuJit] localloc#124250

AndyAyersMS · 2026-02-11T00:12:11Z

Implement localloc.

Add the ability to have a simple block sig (an if/else that pushes a value).
Ensure there is a frame pointer reg
Save the frame pointer reg at the bottom of the stack after moving the SP (but not yet doing the analogous thing in the prolog to store the "method info")
Align both the localloc value and the SP to STACK_ALIGN (16 bytes).
Ignore possible stack overflow

Implement localloc when the size is known at jit time. The runtime case requires internal registers so can't be implemented yet. Remove the NYT around GS checks, we will rely on the TODO and issue tracking to decide if we want to do something here later on.

Copilot

Pull request overview

This PR implements fixed-sized localloc support for WASM RyuJIT by adding the memory.fill instruction and the genLclHeap function to generate code for GT_LCLHEAP nodes when the size is known at JIT time.

Changes:

Added memory.fill WASM instruction definition for zeroing allocated memory
Implemented genLclHeap() to handle compile-time constant localloc operations
Added GT_LCLHEAP case to the main codegen switch statement

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 5 comments.

File	Description
src/coreclr/jit/instrswasm.h	Adds `memory.fill` instruction definition for zeroing memory blocks
src/coreclr/jit/codegenwasm.cpp	Implements GT_LCLHEAP codegen with stack adjustment, optional zeroing, and frame pointer saving; variable-sized case marked NYI

src/coreclr/jit/codegenwasm.cpp

AndyAyersMS · 2026-02-11T00:18:56Z

This PR could (should?) also set up the sp[0] invariant in the prolog for non-leaf methods. Offline we discussed that this could be either the fp or else the memory address of the "method info", For methods with localloc it seems simpler to always have it be fp -- otherwise the localloc expansion doesn't know if it can overwrite that slot or not.

For frames where sp doesn't move sp[0] can be the method info pointer. Seems like this could almost just be the gc info but we probably also need to be able to use it to get to the method desc. We could also store the PEP here since we have it as an arg.

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

Copilot

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated 3 comments.

src/coreclr/jit/codegenwasm.cpp

dotnet-policy-service · 2026-02-11T13:36:46Z

Tagging subscribers to 'arch-wasm': @lewing, @pavelsavara
See info in area-owners.md if you want to be subscribed.

AndyAyersMS · 2026-02-12T16:01:41Z

Per ecma,

System.StackOverflowException is thrown if there is insufficient memory to service the
request.

So we'll need to do an explicit overflow check too. On native platforms we tend to get this "for free" via stack probing.

Seems like we thus need a runtime helper to access to the stack limit? Or else maybe emscripten_stack_get_free (may not be accurate since we don't resync the global SP).

SingleAccretion · 2026-02-12T19:49:51Z

So we'll need to do an explicit overflow check too. On native platforms we tend to get this "for free" via stack probing.

Do you mean that we should also insert such checks in the prolog? Otherwise it's not a complete solution.

AndyAyersMS · 2026-02-12T20:21:11Z

So we'll need to do an explicit overflow check too. On native platforms we tend to get this "for free" via stack probing.

Do you mean that we should also insert such checks in the prolog? Otherwise it's not a complete solution.

I am not sure... how do other languages handle stack overflow on Wasm?

SingleAccretion · 2026-02-12T20:28:04Z

I am not sure... how do other languages handle stack overflow on Wasm?

The C/C++ approach is to "let things be".

The default memory layout for debug code (you are presumed to test things with debug code, like we do with checked Jits) is to place the linear memory stack just after __global_base, so that in the event of overflow, your pointers become close to uint.MaxValue and any subsequent memory access traps.

In optimized code, the stack is placed after static data (so that relocations to that static data can be made smaller), so an overflow would still eventually reach the bottom of the memory range, but only after potentially overwriting that static data. It is a somewhat frequent point of contention.

Finally, there is language in the spec itself that limits the number of WASM frames (not linear memory, proper function call frames) that can exist, effectively meaning that if you can choose a value for the linear memory stack that would be large enough for this engine-determined limit to run out (this is what we've observed to be empirically the case in NAOT-LLVM with 2 stacks totaling ~2 MB between them).

Copilot

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated 6 comments.

src/coreclr/jit/codegenwasm.cpp

src/coreclr/jit/regallocwasm.cpp

src/coreclr/jit/codegenwasm.cpp

AndyAyersMS · 2026-02-20T22:07:41Z

AfterLowerLclHeap may be a bit too hacky. I wanted to simplify life by having the size always be TYP_I_IMPL... suggestions welcome.

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

Copilot

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated no new comments.

AndyAyersMS · 2026-02-20T23:14:04Z

PTAL @dotnet/jit-contrib

Copilot

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated 1 comment.

src/coreclr/jit/codegenwasm.cpp

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

Copilot

Pull request overview

Copilot reviewed 9 out of 9 changed files in this pull request and generated no new comments.

SingleAccretion · 2026-02-21T20:21:46Z

src/coreclr/jit/codegenwasm.cpp

+//   * whether the size operand is a constant or not
+//   * whether the allocated memory needs to be zeroed or not (info.compInitMem)
+//
+//   If the sp is changed, the value at sp[0] must be the frame pointer


I think we should add the exact unwinding scheme for variable-sized frames we settled on here into the ABI document.

Yes, will do.

SingleAccretion · 2026-02-21T20:25:23Z

src/coreclr/jit/codegenwasm.cpp

+    bool const     needsZeroing = m_compiler->info.compInitMem;
+    GenTree* const size         = tree->AsOp()->gtOp1;
+
+    assert(genActualType(size->gtType) == TYP_I_IMPL);


Suggested change

assert(genActualType(size->gtType) == TYP_I_IMPL);

assert(genActualType(size) == TYP_I_IMPL);

SingleAccretion · 2026-02-21T20:26:09Z

src/coreclr/jit/lower.cpp


        case GT_LCLHEAP:
            LowerLclHeap(node);
+            AfterLowerLclHeap(node);


Call this in LowerLclHeap instead?

SingleAccretion · 2026-02-21T20:28:17Z

src/coreclr/jit/emitwasm.h


 public:
 void emitIns(instruction ins);
+void emitIns_B(instruction ins, WasmValueType valType = WasmValueType::Invalid);


Nit: emitIns_BlockTy would be more immediately obvious (we already use Ty to mean 'value type').

SingleAccretion · 2026-02-21T20:30:31Z

src/coreclr/jit/lowerwasm.cpp

+
+    if (!size->isContainedIntOrIImmed())
+    {
+        size = m_compiler->gtNewCastNode(TYP_I_IMPL, size, false, TYP_I_IMPL);


For 32 bit WASM TYP_I_IMPL == TYP_INT. So a no-op.

Also, should this be an unsigned cast?

Also, this would deserve a comment as to why we're deciding to add a cast here (| why is it better than the alternatives).

SingleAccretion · 2026-02-21T20:33:32Z

src/coreclr/jit/emitwasm.cpp

+    instrDesc* id  = emitNewInstrSC(EA_4BYTE, (cnsval_ssize_t)valType);
+    insFormat  fmt = emitInsFormat(ins);
+
+    id->idIns(ins);
+    id->idInsFmt(fmt);
+
+    dispIns(id);
+    appendToCurIG(id);


Just:

Suggested change

instrDesc* id = emitNewInstrSC(EA_4BYTE, (cnsval_ssize_t)valType);

insFormat fmt = emitInsFormat(ins);

id->idIns(ins);

id->idInsFmt(fmt);

dispIns(id);

appendToCurIG(id);

emitIns_I(ins, EA_4BYTE, (cnsval_ssize_t)valType);

?

SingleAccretion · 2026-02-21T20:35:26Z

src/coreclr/jit/regallocwasm.cpp

    }

-    if (anyFrameLocals)
+    if (anyFrameLocals || m_codeGen->isFramePointerRequired())


I think it would be better to use the narrower compLocallocUsed here instead of isFramePointerRequired, with a comment to the effect of "LCLHEAP codegen uses FP for unwind info establishment..." (we wouldn't need the FP with a different scheme).

isFramePointerRequired is set with lots of conditions that are deeply irrelevant for WASM and probably should be ifdef-ed out. There are no FP chains on WASM.

SingleAccretion · 2026-02-21T20:36:35Z

src/coreclr/jit/regallocwasm.cpp

+//
+void WasmRegAlloc::CollectReferencesForLclHeap(GenTreeOp* lclHeapNode)
+{
+    ConsumeTemporaryRegForOperand(lclHeapNode->gtGetOp1() DEBUGARG("lcl heap"));


Suggested change

ConsumeTemporaryRegForOperand(lclHeapNode->gtGetOp1() DEBUGARG("lcl heap"));

ConsumeTemporaryRegForOperand(lclHeapNode->gtGetOp1() DEBUGARG("lcl heap size"));

SingleAccretion · 2026-02-21T20:38:25Z

src/coreclr/jit/codegenwasm.cpp

+
+                // TODO-WASM-CQ: possibly do small fills directly
+                //
+                GetEmitter()->emitIns_I(INS_memory_fill, EA_4BYTE, 0);


(This should rendezvous with the block store PR on using a named constant for the memory index)

SingleAccretion · 2026-02-21T20:42:47Z

src/coreclr/jit/codegenwasm.cpp

+            // Save the aligned size for the zeroing call below if needed.
+            if (needsZeroing)
+            {
+                GetEmitter()->emitIns_I(INS_local_tee, EA_PTRSIZE, WasmRegToIndex(sizeReg));


GetMultiUseOperandReg regs cannot be modified like this in the general case (they may still be live if they're candidates).

Ah, good point.

I will revise this to use an internal register. Doing that I can also avoid the need for the extra casting logic in lower.

Copilot AI review requested due to automatic review settings February 11, 2026 00:12

dotnet-policy-service bot assigned AndyAyersMS Feb 11, 2026

github-actions bot added the area-CodeGen-coreclr CLR JIT compiler in src/coreclr/src/jit and related components such as SuperPMI label Feb 11, 2026

Copilot started reviewing on behalf of AndyAyersMS February 11, 2026 00:13 View session

Copilot AI reviewed Feb 11, 2026

View reviewed changes

Update src/coreclr/jit/codegenwasm.cpp

7084c8e

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

Copilot AI review requested due to automatic review settings February 11, 2026 00:19

Copilot started reviewing on behalf of AndyAyersMS February 11, 2026 00:19 View session

fix comments

fc8f339

Copilot AI reviewed Feb 11, 2026

View reviewed changes

src/coreclr/jit/codegenwasm.cpp Show resolved Hide resolved

src/coreclr/jit/codegenwasm.cpp Show resolved Hide resolved

src/coreclr/jit/codegenwasm.cpp Outdated Show resolved Hide resolved

build-analysis bot mentioned this pull request Feb 11, 2026

Test failure: baseservices/exceptions/stackoverflow/stackoverflowtester/stackoverflowtester.cmd #110173

Open

am11 added the arch-wasm WebAssembly architecture label Feb 11, 2026

kg approved these changes Feb 12, 2026

View reviewed changes

AndyAyersMS added 5 commits February 19, 2026 16:45

merge main

b3f5601

wip

9298e1a

more wip

ba84fc5

fixes

c80daa4

Merge branch 'main' into WasmFixedLocalloc

03d7f61

Copilot AI review requested due to automatic review settings February 20, 2026 22:00

Copilot started reviewing on behalf of AndyAyersMS February 20, 2026 22:01 View session

AndyAyersMS changed the title ~~[Wasm RyuJit] fixed-sized locallocs~~ [Wasm RyuJit] localloc Feb 20, 2026

Copilot AI reviewed Feb 20, 2026

View reviewed changes

Apply suggestions from code review

02b29d5

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

Copilot AI review requested due to automatic review settings February 20, 2026 22:12

Update src/coreclr/jit/regallocwasm.cpp

fefc127

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

Copilot started reviewing on behalf of AndyAyersMS February 20, 2026 22:13 View session

Copilot AI reviewed Feb 20, 2026

View reviewed changes

dont try and reuse the reserved space

0eefad8

AndyAyersMS marked this pull request as ready for review February 20, 2026 23:13

Copilot AI review requested due to automatic review settings February 20, 2026 23:13

Copilot started reviewing on behalf of AndyAyersMS February 20, 2026 23:14 View session

Copilot AI reviewed Feb 20, 2026

View reviewed changes

src/coreclr/jit/codegenwasm.cpp Outdated Show resolved Hide resolved

Update src/coreclr/jit/codegenwasm.cpp

993d436

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

Copilot AI review requested due to automatic review settings February 20, 2026 23:44

Copilot started reviewing on behalf of AndyAyersMS February 20, 2026 23:45 View session

Copilot AI reviewed Feb 20, 2026

View reviewed changes

merge main

9659394

build-analysis bot mentioned this pull request Feb 21, 2026

Unable to pull image from mcr.microsoft.com #117164

Open

SingleAccretion reviewed Feb 21, 2026

View reviewed changes

	assert(genActualType(size->gtType) == TYP_I_IMPL);
	assert(genActualType(size) == TYP_I_IMPL);

	ConsumeTemporaryRegForOperand(lclHeapNode->gtGetOp1() DEBUGARG("lcl heap"));
	ConsumeTemporaryRegForOperand(lclHeapNode->gtGetOp1() DEBUGARG("lcl heap size"));

Comments

Conversation

AndyAyersMS commented Feb 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

AndyAyersMS commented Feb 11, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Uh oh!

dotnet-policy-service bot commented Feb 11, 2026

Uh oh!

AndyAyersMS commented Feb 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

SingleAccretion commented Feb 12, 2026

Uh oh!

AndyAyersMS commented Feb 12, 2026

Uh oh!

SingleAccretion commented Feb 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

AndyAyersMS commented Feb 20, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

AndyAyersMS commented Feb 20, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

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!

SingleAccretion Feb 21, 2026 • 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

AndyAyersMS commented Feb 11, 2026 •

edited

Loading

AndyAyersMS commented Feb 12, 2026 •

edited

Loading

SingleAccretion commented Feb 12, 2026 •

edited

Loading

SingleAccretion Feb 21, 2026 •

edited

Loading