all: bump golang.org/x/tools to v0.49.0, fold composite literals, and add LTO baseline benchmarks

[cpunion]

Summary

Bumps github.com/goplus/mod to v0.22.0 (which brings golang.org/x/tools v0.49.0, golang.org/x/mod v0.40.0, golang.org/x/sys v0.47.0, and golang.org/x/sync v0.22.0), adapts compiler lowering to go/ssa v0.49.0, enhances static initializer folding for package-level composite literals, and adds LTO measurements and PR merge-base baseline tracking to the benchmark suite.

---

Key Changes

1. Dependency Upgrade:

   • github.com/goplus/mod from v0.21.2 to v0.22.0
   • golang.org/x/tools to v0.49.0
   • golang.org/x/mod to v0.40.0, golang.org/x/sys to v0.47.0, golang.org/x/sync to v0.22.0

2. Composite Literal Static Folding (cl/static_init.go):

   • In x/tools v0.49.0, go/ssa lowers composite literals in init() using temporary local stack variables (bottom-up *ssa.Alloc + *ssa.Store + *ssa.UnOp(token.MUL)) rather than direct global field stores.
   • Enhanced cl/static_init.go (collectAllocStores) to recursively trace and fold these local temporary alloc stores into static LLVM global constants (avoiding runtime init() stack allocation, memset, and element-wise store instructions, saving ~52.6 KB of binary bloat).
   • Path Isolation: Used appendStaticInitPath to allocate isolated slice paths, preventing slice-aliasing hazards during DAG path collection.
   • Single-Consumer Validation: Added strict validation for *ssa.UnOp(MUL) to ensure dereferenced values are consumed solely by the folded store before instruction suppression.
   • Documentation: Added comprehensive GoDoc comments documenting the recursive DAG traversal and cycle protection.

3. Zero-Sized Global Sentinel Guard (ssa/decl.go):

   • Prevented packages with zero-sized global variables (such as internal/poll) from overwriting the module's shared sentinel @"__llgo.moduleZeroSizedAlloc$" with non-i8 initializers during package initialization, avoiding LLVM module verification panics.

4. Benchmark Enhancements (.github/workflows/benchmark.yml, benchmark/baseline/):

   • PR Merge-Base Baseline Tracking: Automatically computes and compares against the true git merge-base commit between base and head, eliminating base drift from concurrent main branch merges.
   • LTO Baseline Measurements: Added -lto=full workload measurements for cprintf_lto, println_lto, and fmtprintf_lto.

5. Test Fixtures & Coverage (cl/rewrite_internal_test.go, cl/_test*):

   • Refreshed FileCheck LITTEST fixtures in cl/_testgo/ and cl/_testrt/ for x/tools v0.49.0 lowering (explicit store zeroinitializer following AllocZ, stack slot adjustments).
   • Added unit test suites TestStaticGlobalPointerIndirectionLiteralInit and TestCollectAllocStoresFromSSA in cl/rewrite_internal_test.go covering nested composite literals, pointer indirections, and edge cases.

---

Verification

• go test -v ./cl -run 'TestStaticGlobal|TestBlankField|TestCollectAllocStores|TestRunAndTestFromTestrt/qsort' (PASS)
• go test -v ./benchmark/baseline (PASS)
• go test -v ./internal/packages (PASS)
• bash test/buildcache/test.sh (18/18 PASS)

[fennoai]

Review: bump goplus/mod to 0.22.0 + loader rewrite

Solid change overall. The dependency bump (goplus/mod 0.22.0, golang.org/x/{mod,sys,tools,sync}) forced dropping the go:linkname hooks into x/tools' internal loader symbols, and the internal/packages/load.go rewrite replaces them with a self-contained implementation over the public packages.Load + packages.Visit. go.sum hashes verify against the checksum database. The ssa/decl.go moduleZeroName guard and the new pointer/nested-alloc handling in cl/static_init.go are reasonable additions.

One issue worth fixing before merge — a slice-aliasing hazard in the new collectAllocStores that can silently miscompile static initializers (inline comment below).

Secondary notes (non-blocking):

• internal/packages/load.go — serial typecheck. The old loader parsed/typechecked dependencies in parallel goroutines; the new packages.Visit post-order loop runs typecheckPackage serially. This is an acceptable trade-off (correctness/simplicity, and post-order guarantees imported Types are complete before dependents), but it's a real compile-time cost on large dependency graphs — worth noting as known.
• internal/packages/load.go behavior parity. The rewrite drops IgnoreFuncBodies, the FakeImportC import "C" ignored diagnostic, and the requestedMode field-zeroing that trimmed returned Package fields to the requested mode. Both in-repo callers use the full loadSyntax mode so they're unaffected, but LoadEx/LoadExWithGoVersion are exported — returned fields are no longer trimmed to the requested mode. Consider documenting this in the doc comment.
• internal/packages/load.go:262 — targetCompilerAndArch(pkg *Package) never uses its pkg parameter; drop it.
• internal/packages/load.go:156 — doc-comment nit. The LoadEx comment (inherited from upstream packages.Load) refers to "the PrintErrors function is provided" and "Load returns an error", but this package re-exports neither PrintErrors nor a Load symbol. Minor wording drift.

[fennoai]

[P1] Slice-aliasing corrupts static-init paths in collectAllocStores

Slice-aliasing hazard: append(basePath, subPath...) results are stored long-term.

handleStoreVal stores the passed path slice directly into out[].path (staticInitStore.path), which is consumed much later in buildStaticGlobalInit → staticInitNode.add. But append(basePath, subPath...) reuses basePath's backing array whenever cap(basePath) > len(basePath). When an alloc has two or more FieldAddr/IndexAddr referrers (a struct with ≥2 fields, an array with ≥2 indices), the second iteration's append overwrites the tail of the path recorded for the first sibling — the retained path in out is silently corrupted.

basePath carries spare capacity in the common case: the nested-alloc recursion at line 283 passes fullPath (itself an append result, so cap > len), and the top-level call also passes a built-up path when the deref store targets a sub-path of the global. The result is a silent miscompilation (wrong field/element gets the value), not a crash.

Fix by copying per element, e.g.:

full := append(append([]staticInitPathElem(nil), basePath...), subPath...)
if !handleStoreVal(elemStore, full, out, instrs, visited) {

(same at line 253). staticInitStorePath and staticInitStorePathToAlloc build fresh slices along a single linear recursion, so they're safe.

[github-actions]

LLGo baseline benchmarks

4fc79b727911 | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs base	Build	vs base	Run	vs base
Linux	cprintf	19288 B	0 B / +0.0%	327.193 ms	-14.54 ms / -4.3% (better)	1.275 ms	-21.74 us / -1.7% (better)
Linux	cprintf\_lto	19120 B	0 B / +0.0%	335.892 ms	+528.8 us / +0.2% (worse)	1.255 ms	-143 us / -10.2% (better)
Linux	fmtprintf	1877440 B	+17288 B / +0.9% (worse)	2.822 s	+61.22 ms / +2.2% (worse)	3.192 ms	+5.129 us / +0.2% (worse)
Linux	fmtprintf\_lto	1797016 B	+19912 B / +1.1% (worse)	15.104 s	+276.7 ms / +1.9% (worse)	3.234 ms	+106.7 us / +3.4% (worse)
Linux	println	68776 B	0 B / +0.0%	339.929 ms	-5.743 ms / -1.7% (better)	1.551 ms	-34.02 us / -2.1% (better)
Linux	println\_lto	62464 B	0 B / +0.0%	547.561 ms	-2.946 ms / -0.5% (better)	1.607 ms	-52.74 us / -3.2% (better)
macOS	cprintf	84672 B	0 B / +0.0%	336.565 ms	-61.34 ms / -15.4% (better)	2.339 ms	-399.5 us / -14.6% (better)
macOS	cprintf\_lto	100912 B	0 B / +0.0%	344.192 ms	-49.66 ms / -12.6% (better)	2.452 ms	-546.4 us / -18.2% (better)
macOS	fmtprintf	1909104 B	+16384 B / +0.9% (worse)	2.235 s	-958 ms / -30.0% (better)	11.395 ms	-2.666 ms / -19.0% (better)
macOS	fmtprintf\_lto	1642928 B	+16528 B / +1.0% (worse)	16.783 s	-11.99 s / -41.7% (better)	6.280 ms	-3.157 ms / -33.5% (better)
macOS	println	121360 B	0 B / +0.0%	405.199 ms	-48.88 ms / -10.8% (better)	4.130 ms	-173.6 us / -4.0% (better)
macOS	println\_lto	128528 B	0 B / +0.0%	478.510 ms	-179.4 ms / -27.3% (better)	3.593 ms	-2.498 ms / -41.0% (better)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs base
Linux	BenchmarkLookupPCRandom	13.270 ns/op	+0.05 ns/op / +0.4% (worse)
Linux	BenchmarkMergeCompilerFlags	150.400 ns/op	-0.3 ns/op / -0.2% (better)
Linux	BenchmarkMergeLinkerFlags	94.720 ns/op	+0.57 ns/op / +0.6% (worse)
Linux	BenchmarkChannelBuffered	35.530 ns/op	+1.54 ns/op / +4.5% (worse)
Linux	BenchmarkChannelHandoff	28419 ns/op	-1402 ns/op / -4.7% (better)
Linux	BenchmarkDefer	43.860 ns/op	-1.41 ns/op / -3.1% (better)
Linux	BenchmarkDirectCall	1.557 ns/op	0 ns/op / +0.0%
Linux	BenchmarkGlobalRead	1.557 ns/op	+0.001 ns/op / +0.1% (worse)
Linux	BenchmarkGlobalWrite	2.487 ns/op	+0.007 ns/op / +0.3% (worse)
Linux	BenchmarkGoroutine	31358 ns/op	-329 ns/op / -1.0% (better)
Linux	BenchmarkInterfaceCall	7.783 ns/op	-0.314 ns/op / -3.9% (better)
Linux	BenchmarkRuntimeGetG	2.179 ns/op	-0.001 ns/op / -0.04587% (better)
macOS	BenchmarkLookupPCRandom	11.300 ns/op	-1.7 ns/op / -13.1% (better)
macOS	BenchmarkMergeCompilerFlags	98.300 ns/op	-31.9 ns/op / -24.5% (better)
macOS	BenchmarkMergeLinkerFlags	63.110 ns/op	-14.46 ns/op / -18.6% (better)
macOS	BenchmarkChannelBuffered	21.620 ns/op	-2.78 ns/op / -11.4% (better)
macOS	BenchmarkChannelHandoff	6744 ns/op	-1018 ns/op / -13.1% (better)
macOS	BenchmarkDefer	28.020 ns/op	-1.1 ns/op / -3.8% (better)
macOS	BenchmarkDirectCall	1.005 ns/op	-0.057 ns/op / -5.4% (better)
macOS	BenchmarkGlobalRead	0.995 ns/op	-0.0366 ns/op / -3.5% (better)
macOS	BenchmarkGlobalWrite	0.997 ns/op	-0.0505 ns/op / -4.8% (better)
macOS	BenchmarkGoroutine	31818 ns/op	+1790 ns/op / +6.0% (worse)
macOS	BenchmarkInterfaceCall	5.734 ns/op	+1.322 ns/op / +30.0% (worse)
macOS	BenchmarkRuntimeGetG	1.995 ns/op	-0.028 ns/op / -1.4% (better)

Compared with 0314faeba69d measured in the same runner job.

[codecov]

Codecov Report

❌ Patch coverage is 70.58824% with 35 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
cl/static_init.go	69.56%	22 Missing and 13 partials ⚠️

📢 Thoughts on this report? Let us know!

[cpunion]

All review feedback addressed in the latest commit:

• Slice-Aliasing Hazard: Switched from direct append(basePath, subPath...) to appendStaticInitPath with explicit slice copy to ensure complete path isolation across siblings.
• UnOp Single-Consumer Validation: Added strict validation in collectStaticGlobalInits, collectAllocStores, and handleStoreVal to guarantee *ssa.UnOp(MUL) is only consumed by the folded store before instruction suppression.
• Package Loader Cleanup: Removed unused pkg *Package parameter from targetCompilerAndArch() and updated doc comments on LoadEx.
• Test Coverage: Added TestStaticGlobalPointerIndirectionLiteralInit and TestCollectAllocStoresFromSSA in cl/rewrite_internal_test.go.