compiler: bound aggregate call signatures

[jakebailey]

Fixes #5615

This is a sort-of more involved alternative to #5618; the gist is that we ensure we limit everything to up to 1000 scalar elements across the board. Back when I did #5526, I had something like this, but scaled it back because I didn't think exactness would matter too much so long as we didn't overload LLVM. But, I guess it does matter.

[Copilot]

Pull request overview

This PR limits scalarized aggregate function signatures to 1,000 parameters.

Changes:

• Selects aggregate parameters for indirect passing.
• Handles calls, interfaces, defers, goroutines, and exported ABI validation.
• Adds ABI and global-reference tests.

Reviewed changes

Copilot reviewed 15 out of 15 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
builder/build.go	Validates WebAssembly signatures.
compiler/calls.go	Centralizes ABI argument handling.
compiler/compiler.go	Applies ABI choices to calls and functions.
compiler/compiler_test.go	Tests aggregate ABI limits.
compiler/defer.go	Applies the ABI to deferred calls.
compiler/func.go	Calculates and validates function ABIs.
compiler/goroutine.go	Applies the ABI to goroutine arguments.
compiler/interface.go	Handles interface invoke wrappers.
compiler/llvmutil/llvm.go	Removes temporary global references.
compiler/llvmutil/llvm_test.go	Tests global-reference removal.
compiler/symbol.go	Protects indirect signatures during optimization.
compiler/testdata/aggregate-abi.go	Adds internal ABI fixtures.
compiler/testdata/aggregate-export-abi.go	Adds exported ABI fixtures.
transform/interface-lowering.go	Reports incompatible interface ABIs.
transform/optimizer.go	Removes temporary ABI roots.

Suppressed comments (2)

compiler/compiler.go:2326

• This ABI calculation omits the synthetic typecode parameter that is appended below. At the 1000 parameter boundary, the invoke function spills an aggregate but this call passes it directly, which gives the callee a different ABI.

		abi = b.getInterfaceFunctionABI(instr.Signature())

compiler/interface.go:1350

• This uses the concrete receiver when it selects indirect parameters. The interface invoke ABI replaces that receiver with a pointer, so it can select different nonreceiver parameters and the wrapper then forwards values with the wrong ABI.

	receiverType := abi.params[0].llvmType
	var expandedReceiverType []llvm.Type
	receiverIndirect := abi.params[0].indirect
	var receiverInfos []paramInfo
	if receiverIndirect {
		receiverInfos = []paramInfo{{llvmType: c.dataPtrType}}
	} else {
		receiverInfos = c.expandDirectFormalParamType(receiverType, "", nil)

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 15 out of 15 changed files in this pull request and generated no new comments.

Suppressed comments (1)

Previously missed (1) — in code that hasn't changed since the last review.

compiler/compiler_test.go:249

• This test checks only that the compiler emits the marker. It does not run LowerInterfaces, so the new diagnostic path is not tested. Add a transform or build test that checks the expected error and a compatible exported interface call.

	if markedWrappers != 1 {

[deadprogram]

Thank you for this work @jakebailey. Here are the edited results from an automated review.

Findings

Dead code and stale text

• compiler/func.go:198 - aggregateValueCount now has only one caller, compiler_test.go:505. No production code uses it.
• compiler/calls.go:175 - the comment refers to expandFormalParamType, which this PR removes.

Unrelated behavior change

• transform/optimizer.go - the second commit moves the config.Scheduler() == "none" goroutine check from before the ThinLTO pre-link pipeline to after it. Only RemoveGlobalReferences and the new globaldce must move. The check now runs after dead code elimination, so a goroutine start that the optimizer removes no longer reports attempted to start a goroutine without a scheduler. The commit message does not mention this. Is the move necessary?

Polish

• transform/optimizer.go - the added globaldce runs on every build, also when no tinygo.indirect-abi global exists. A guard prevents the unnecessary pass.
• transform/interface-lowering.go:279 - fmt.Errorf("%s", ...) gives an error with no source position. The other code in this pass uses errorAt. The loop also returns at the first error, so the user sees one diagnostic at a time.
• compiler/compiler.go:1446 - SetInsertPointBefore in the phi loop leaves the insert point in a predecessor block after the loop ends. This is correct now, but it can break later changes.
• The validation message function bigExport has more than 1000 WebAssembly parameters after ABI lowering does not tell the user that //export prevented the lowering, or what to change.

Process

• The branch is 36 commits behind dev. Please rebase before merge.

[jakebailey]

Debased and updated, PTAL

[dgryski]

Passes my test corpus.

[deadprogram]

Thank you for the update @jakebailey. Here are the edited results from an automated review. All findings from the previous review are addressed.

Findings

1. compiler/interface.go:1345 - an exported method that an interface can reach stops the build with an LLVM verifier error, not with the new diagnostic:

function declaration may only have a unique !dbg attachment
ptr @"tcall$invoke"
verification error after compiling package t

attachDebugInfoRaw makes a DISubprogram with IsDefinition: true, but the ABI-error wrapper stays a declaration. LLVM permits only a unique, non-definition subprogram on a declaration (LLVM Language Reference, "Source Level Debugging", DISubprogram). The verifier stops the build before LowerInterfaces reports the error. A declaration-style DISubprogram corrects this and keeps the source position.

2. compiler/compiler_test.go:236 - TestAggregateExportedInterfaceABI does not call llvm.VerifyModule, so it does not find finding 1. Add a verification step for both debug settings.

3. compiler/defer.go:632 - the first getFunctionABI result is discarded for the invoke case. Move it into the else branch. Also, params = abi.params and params = params[1:] can be one line.

[deadprogram]

All looking good now @jakebailey thank you very much. Also thank you @dgryski for review. Now merging.