https://github.com/madhur13490 requested changes to this pull request.
The type mapping looks right, and I like that the IR spellings (tuple → struct
of N, svcount → aarch64.svcount) match what CodeGenTypes already produces.
Single SVE vectors, svbool_t, and __SVCount_t all classify identically to the
classic path, so those are correct.
The change that maps SVE types now routes them into
AArch64TargetInfo::classifyArgumentType, whose vector path is still just
reportNYI(...) + getDirect(). That warning doesn't gate anything, and
shouldUseLLVMABILowering is on for every AArch64 function, so the getDirect()
gets used. For an SVE tuple that AAPCS64 passes indirectly once the Z/P
registers are full (§5.10 + rule C.8; Clang does this in
`coerceAndExpandPureScalableAggregate`), we now disagree with the classic path.
I applied the patch on a debug build and it aborts:
`void f(svint32x4_t a, svint32x2_t b, svint32x4_t c) {} // 4+2+4 = 10 > 8 Z
regs`
```
warning: Vector argument type handling is not yet implemented ... (x3)
For argument 2 of type __clang_svint32x4_t: Kind mismatch (expected: Indirect,
given: Direct)
<aborts>
```
Release would silently miscompile the call instead. Since the fix lives in
llvm/lib/ABI/Targets/AArch64.cpp, I think this PR either needs to grow to cover
it — gate SVE out in `shouldUseLLVMABILowering`, or implement the
indirect-on-exhaustion path — or the description should be explicit that
`-fexperimental-abi-lowering` isn't safe for SVE signatures yet.
Two more things that belong here because they're in an untouched file
(llvm/lib/ABI/Targets/AArch64.cpp):
The getDirect() fallback for vectors in
`classifyArgumentType/classifyReturnType` is the actual place the fix goes.
The comment in `passAsAggregateType` — "For now, they don't get through the
type mapper" — is no longer true after this change; please update it here or in
a follow-up.
CI: the red "Build and Test Linux AArch64" check is a runner out-of-disk error,
I think not a real failure.
https://github.com/llvm/llvm-project/pull/221375
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
