c-rhodes added inline comments.
================
Comment at: llvm/test/CodeGen/AArch64/sve-intrinsics-perm-select.ll:809
+define <vscale x 8 x bfloat> @rev_bf16(<vscale x 8 x bfloat> %a) {
+; CHECK-LABEL: rev_bf16
----------------
sdesmalen wrote:
> Does this test not need the `+bf16` attribute to work? (which implies the
> patterns are missing the right predicate)
It should do, sorry I missed that. I've tried:
```diff --git a/llvm/lib/Target/AArch64/SVEInstrFormats.td
b/llvm/lib/Target/AArch64/SVEInstrFormats.td
index 46cca2a..5ab2502 100644
--- a/llvm/lib/Target/AArch64/SVEInstrFormats.td
+++ b/llvm/lib/Target/AArch64/SVEInstrFormats.td
@@ -1124,10 +1124,13 @@ multiclass sve_int_perm_reverse_z<string asm,
SDPatternOperator op> {
def : SVE_1_Op_Pat<nxv4i32, op, nxv4i32, !cast<Instruction>(NAME # _S)>;
def : SVE_1_Op_Pat<nxv2i64, op, nxv2i64, !cast<Instruction>(NAME # _D)>;
- def : SVE_1_Op_Pat<nxv8bf16, op, nxv8bf16, !cast<Instruction>(NAME # _H)>;
def : SVE_1_Op_Pat<nxv8f16, op, nxv8f16, !cast<Instruction>(NAME # _H)>;
def : SVE_1_Op_Pat<nxv4f32, op, nxv4f32, !cast<Instruction>(NAME # _S)>;
def : SVE_1_Op_Pat<nxv2f64, op, nxv2f64, !cast<Instruction>(NAME # _D)>;
+
+ let Predicates = [HasBF16] in {
+ def : SVE_1_Op_Pat<nxv8bf16, op, nxv8bf16, !cast<Instruction>(NAME # _H)>;
+ }
}```
but this still works without `+bf16`. I noticed in your patch D82187 you check
`Subtarget->hasBF16()` for `MVT::nxv8bf16` at select phase of ISEL, I guess
it's different here with patterns. I also noticed we add the register class for
`MVT::nxv8bf16` in AArch64ISelLowering without checking `Subtarget->hasBF16()`
which I suspect is a bug. This test requires `+bf16` with that fixed but I
wonder why the predicate isn't being recognised.
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D82182/new/
https://reviews.llvm.org/D82182
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
