[hajimehoshi]

What issue is this addressing?

Closes #458

What type of issue is this addressing?

bug

What this PR does | solves

The i386 System V ABI requires the stack pointer to be 16-byte aligned at a call, and compilers rely on it for aligned SSE stack accesses such as movaps, which fault with sigcode=128 addr=0x0 otherwise. Two linux/386 trampolines broke this:

• syscallX (sys_386.s) is entered on an aligned stack from runtime.asmcgocall, but its four PUSHLs and SUBL $160 left ESP at 12 mod 16 at CALL AX. linux/386 uses this path with Cgo enabled too, since syscall_cgo_linux.go excludes 386.
• call5 (internal/fakecgo/trampolines_386.s), used for pthread_create, malloc, setenv and others, did not realign at all, so its alignment depended on the calling Go frame.

Besides the musl/libstdc++ crash in the issue, this crashes Ebitengine's GOARCH=386 CI on Ubuntu 26.04 (Mesa 26.0.8) in glXQueryExtensionsString.

Both trampolines now save SP in the callee-saved BP and align SP with ANDL $~15, SP before placing the arguments, as call5 in trampolines_amd64.s already does.

TestRegisterFunc_StackAlignment calls a naked C function that returns the caller's stack pointer modulo 16 at the call. It runs on linux/386 and on amd64, where the alignment in syscallX depends on its frame size.

The 386 setg_trampoline and call5 FP operands are also renamed to match their Go declarations (G, ret), which fixes GOARCH=386 go vet's asmdecl reports. The generated code is byte-identical.

Verification:

• In an Ubuntu 26.04 container (gcc 15.2, qemu-i386), the new test fails before the fix with "got 12, want 0" (Cgo on and off) and passes after. The root package passes on linux/386 and linux/amd64 with Cgo on and off.
• On darwin/amd64 (Rosetta), the test passes, and it fails with "got 8, want 0" when syscallX's frame is shifted by 8 bytes.
• Ebitengine's internal/graphicscommand 386 tests crash with the CI signature before the fix and pass after.
• The Alpine/musl reproduction in the issue was not rerun.

Filed by Claude (Claude Code), on behalf of @hajimehoshi

🤖 Generated with Claude Code

[hajimehoshi]

@TotallyGamerJet PTAL

[Copilot]

🟢 Approval recommended

The Linux/386 C-call stack-alignment fix has no active review issues.

No open issues 🎉

What changed in this PR

Fixes Linux/386 C-call stack alignment in PureGo’s trampolines, addressing the ABI violation behind the crashes reported in #458.

Changes:

• Aligns the stack to 16 bytes before C calls in syscallX and fakecgo’s call5, preserving the original stack pointer for restoration.
• Adds a regression test for call-site stack alignment on Linux/386 and amd64.
• Renames 386 assembly operands to match their Go declarations and resolve go vet reports.

File	Description
testdata/​abitest/​abi_test.c	Adds a C helper that measures call-site stack alignment.
sys_386.s	Aligns the C-call stack and restores the saved stack pointer.
internal/​fakecgo/​trampolines_386.s	Aligns call5’s stack and corrects assembly operand names.
func_test.go	Adds the stack-alignment regression test.

_{Review effort: Balanced}

[TotallyGamerJet]

LGTM