cabi: preserve debug variable homes

[cpunion]

Problem

LLGo disables optimized C ABI lowering whenever DWARF is enabled. The lowering replaces a source parameter stack home with ABI storage, but its selective operand rewrite does not update LLVM debug records. Enabling it as-is can leave dbg.declare attached to stale storage.

Changes

• run the same C ABI transformer with and without DWARF;
• replace uses of the old parameter home so LLVM updates debug records and executable instructions together;
• restore only setup instructions before the parameter store to the original alloca;
• verify the setup operand and rewritten dbg.declare/load operands;
• run the existing large-aggregate probe with DWARF-enabled abi2.

The change stays inside the existing C ABI transformer and adds no debug-specific ABI or target exception.

Dependency update

#2143 and #2215 are merged. This branch is rebased directly onto current main and contains only its C ABI commit.

Verification

• go test ./internal/cabi ./internal/build;
• go test ./test/go -run 'TestLargeArrayReturn(AllABIModes|DWARF)$' -count=1;
• go test -vet=off ./...;
• git diff --check.

The unscoped go test ./... still hits the same Go toolchain printf analyzer panic in test/go on the merged #2143 baseline; disabling vet verifies the test suite itself.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[github-actions]

LLGo baseline benchmarks

1f7a6e985682 | workflow run | long-term charts

Program measurements

Platform	Workload	File size	vs base	Build	vs base	Run	vs base
Linux	cprintf	18456 B	+0.0%	320.780 ms	+10.2% (worse)	1.282 ms	-11.0% (better)
Linux	fmtprintf	1829968 B	+0.0%	3.023 s	+0.2% (worse)	3.139 ms	+0.1% (worse)
Linux	println	68008 B	+0.0%	293.418 ms	-3.4% (better)	1.575 ms	-16.1% (better)
macOS	cprintf	84672 B	+0.0%	358.095 ms	-6.1% (better)	2.466 ms	-14.8% (better)
macOS	fmtprintf	1869328 B	+0.0%	2.851 s	-15.6% (better)	13.664 ms	+12.2% (worse)
macOS	println	121200 B	+0.0%	330.233 ms	-0.4% (better)	3.301 ms	-14.9% (better)

Core language and compiler benchmarks

Platform	Benchmark	ns/op	vs base
Linux	BenchmarkLookupPCRandom	13.360 ns/op	+0.6% (worse)
Linux	BenchmarkMergeCompilerFlags	151 ns/op	+0.5% (worse)
Linux	BenchmarkMergeLinkerFlags	96.980 ns/op	+2.7% (worse)
Linux	BenchmarkChannelBuffered	33.670 ns/op	+0.1% (worse)
Linux	BenchmarkChannelHandoff	28217 ns/op	-3.2% (better)
Linux	BenchmarkDefer	48.590 ns/op	-0.2% (better)
Linux	BenchmarkDirectCall	1.557 ns/op	-0.1% (better)
Linux	BenchmarkGlobalRead	1.559 ns/op	+0.2% (worse)
Linux	BenchmarkGlobalWrite	2.478 ns/op	-0.0% (better)
Linux	BenchmarkGoroutine	33457 ns/op	-0.9% (better)
Linux	BenchmarkInterfaceCall	8.101 ns/op	+0.0% (worse)
Linux	BenchmarkRuntimeGetG	2.181 ns/op	-0.0% (better)
macOS	BenchmarkLookupPCRandom	14.140 ns/op	-6.5% (better)
macOS	BenchmarkMergeCompilerFlags	128.300 ns/op	-19.9% (better)
macOS	BenchmarkMergeLinkerFlags	118.900 ns/op	+47.6% (worse)
macOS	BenchmarkChannelBuffered	27.530 ns/op	+8.0% (worse)
macOS	BenchmarkChannelHandoff	10283 ns/op	+69.0% (worse)
macOS	BenchmarkDefer	33.390 ns/op	+5.4% (worse)
macOS	BenchmarkDirectCall	1.227 ns/op	+15.0% (worse)
macOS	BenchmarkGlobalRead	1.330 ns/op	+3.0% (worse)
macOS	BenchmarkGlobalWrite	1.335 ns/op	+6.0% (worse)
macOS	BenchmarkGoroutine	45387 ns/op	+25.8% (worse)
macOS	BenchmarkInterfaceCall	5.143 ns/op	+7.2% (worse)
macOS	BenchmarkRuntimeGetG	2.467 ns/op	+11.6% (worse)

Compared with 6670dae3884d measured in the same runner job.