[clang] [llvm] [LLVMABI] Implement the System V X86-64 ABI (PR #194718)
via llvm-commits
llvm-commits at lists.llvm.org
Mon May 11 14:51:56 PDT 2026
https://github.com/adams381 commented:
Posted a consumer-side validation: compiled 10 SysV-x86_64 ABI test files from `clang/test/CodeGen/X86/` through both the new path (`-fexperimental-abi-lowering`, with the createPairType suggestion + a second fix described inline below) and the legacy path (classic codegen), then diffed function signatures.
**Result: 139 / 139 functions match across all 10 files.**
| Test file | Match |
|---|---|
| `x86_64-arguments.c` | 73 / 73 |
| `x86_64-floatvectors.c` | 7 / 7 |
| `x86_64-longdouble.c` | 11 / 11 |
| `va-arg-sse.c` | 1 / 1 |
| `bfloat-abi.c` | 10 / 10 |
| `bfloat-half-abi.c` | 10 / 10 |
| `fp128-abi.c` | 4 / 4 |
| `avx512fp16-abi.c` | 18 / 18 |
| `avx512fp16-complex-abi.c` | 1 / 1 |
| `x86_64-atomic-128.c` | 4 / 4 |
The validation found one additional bug beyond the createPairType issue — the `Complex_X87` return path has the same `[N x i8]` padding pattern. A one-click suggestion is inline. With both fixes applied, it's 139/139.
Methodology: signatures normalized via `grep '^define' | sed 's/dso_local //;s/ #[0-9]*//;...'`. Tests selected: SysV-x86_64 only (skipped Win32, Darwin, x86_32 variants where the new path shouldn't activate). All compiled with `-triple x86_64-unknown-linux-gnu` plus each file's normal target features.
https://github.com/llvm/llvm-project/pull/194718
More information about the llvm-commits
mailing list