================
@@ -26,3 +26,42 @@ __m512bh test_mm512_undefined_pbh(void) {
// OGCG: ret <32 x bfloat> zeroinitializer
return _mm512_undefined_pbh();
}
+
+__mmask32 test_mm512_mask_fpclass_pbh_mask(__mmask32 __U, __m512bh __A) {
+ // CIR-LABEL: _mm512_mask_fpclass_pbh_mask
+ // CIR: %[[A:.*]] = cir.call_llvm_intrinsic "x86.avx10.fpclass.bf16.512"
+ // CIR: %[[B:.*]] = cir.cast bitcast {{.*}} : !u32i -> !cir.vector<32 x
!cir.bool>
+ // CIR: %[[C:.*]] = cir.binop(and, %[[A]], %[[B]]) : !cir.vector<32 x
!cir.bool>
+ // CIR: cir.cast bitcast %[[C]] : !cir.vector<32 x !cir.bool> -> !u32i
+
+ // LLVM-LABEL: test_mm512_mask_fpclass_pbh_mask
+ // LLVM: %[[A:.*]] = call <32 x i1> @llvm.x86.avx10.fpclass.bf16.512
+ // LLVM: %[[B:.*]] = bitcast i32 {{.*}} to <32 x i1>
+ // LLVM: %[[C:.*]] = and <32 x i1> %[[A]], %[[B]]
+ // LLVM: bitcast <32 x i1> %[[C]] to i32
+
+ // OGCG-LABEL: test_mm512_mask_fpclass_pbh_mask
+ // OGCG: %[[A:.*]] = call <32 x i1> @llvm.x86.avx10.fpclass.bf16.512
+ // OGCG: %[[B:.*]] = bitcast i32 {{.*}} to <32 x i1>
+ // OGCG: %[[C:.*]] = and <32 x i1> %[[A]], %[[B]]
+ // OGCG: bitcast <32 x i1> %[[C]] to i32
+ return _mm512_mask_fpclass_pbh_mask(__U, __A, 4);
+}
+
+__mmask32 test_mm512_fpclass_pbh_mask(__m512bh __A) {
+ // CIR-LABEL: _mm512_fpclass_pbh_mask
+ // CIR: %[[A:.*]] = cir.call_llvm_intrinsic "x86.avx10.fpclass.bf16.512"
+ // CIR: %[[B:.*]] = cir.cast bitcast {{.*}} : !u32i -> !cir.vector<32 x
!cir.bool>
+ // CIR: %[[C:.*]] = cir.binop(and, %[[A]], %[[B]]) : !cir.vector<32 x
!cir.bool>
+ // CIR: cir.cast bitcast %[[C]] : !cir.vector<32 x !cir.bool> -> !u32i
+
+ // LLVM-LABEL: test_mm512_fpclass_pbh_mask
+ // LLVM: %[[A:.*]] = call <32 x i1> @llvm.x86.avx10.fpclass.bf16.512
+ // LLVM: %[[B:.*]] = and <32 x i1> %[[A]], splat (i1 true)
----------------
andykaylor wrote:
It wasn't clear to me at first where the `-1` was coming from, but I see that
it is from this definition in the header file:
```
#define _mm512_fpclass_pbh_mask(__A, imm) \
((__mmask32)__builtin_ia32_vfpclassbf16512_mask( \
(__v32bf)(__m512bh)(__A), (int)(imm), (__mmask32) - 1))
```
We should fold the -1. I'm looking into a couple of possible approaches to
doing that.
https://github.com/llvm/llvm-project/pull/172813
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
