[X86][FP16] Add the missing legal action for EXTRACT_SUBVECTOR

Fixes #57340

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D132563

(cherry picked from commit 12b203e)

Author: phoebewang

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -1521,7 +1521,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     // Extract subvector is special because the value type
     // (result) is 128-bit but the source is 256-bit wide.
     for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32, MVT::v2i64,
-                     MVT::v4f32, MVT::v2f64 }) {
+                     MVT::v8f16, MVT::v4f32, MVT::v2f64 }) {
       setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
     }
 
@@ -1861,7 +1861,7 @@ X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
     // (result) is 256-bit but the source is 512-bit wide.
     // 128-bit was made Legal under AVX1.
     for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64,
-                     MVT::v8f32, MVT::v4f64 })
+                     MVT::v16f16, MVT::v8f32, MVT::v4f64 })
       setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
 
     for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32, MVT::v8i64,

llvm/test/CodeGen/X86/avx512-f16c-v16f16-fadd.ll

@@ -4,23 +4,17 @@
 define <16 x half> @foo(<16 x half> %a, <16 x half> %b) nounwind {
 ; CHECK-LABEL: foo:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    pushq %rbp
-; CHECK-NEXT:    movq %rsp, %rbp
-; CHECK-NEXT:    andq $-32, %rsp
-; CHECK-NEXT:    subq $96, %rsp
-; CHECK-NEXT:    vmovaps %ymm1, {{[0-9]+}}(%rsp)
-; CHECK-NEXT:    vmovaps %ymm0, (%rsp)
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %ymm0
-; CHECK-NEXT:    vcvtph2ps (%rsp), %ymm1
-; CHECK-NEXT:    vaddps %ymm0, %ymm1, %ymm0
+; CHECK-NEXT:    vcvtph2ps %xmm1, %ymm2
+; CHECK-NEXT:    vcvtph2ps %xmm0, %ymm3
+; CHECK-NEXT:    vaddps %ymm2, %ymm3, %ymm2
+; CHECK-NEXT:    vcvtps2ph $4, %ymm2, %xmm2
+; CHECK-NEXT:    vextractf128 $1, %ymm1, %xmm1
+; CHECK-NEXT:    vcvtph2ps %xmm1, %ymm1
+; CHECK-NEXT:    vextractf128 $1, %ymm0, %xmm0
+; CHECK-NEXT:    vcvtph2ps %xmm0, %ymm0
+; CHECK-NEXT:    vaddps %ymm1, %ymm0, %ymm0
 ; CHECK-NEXT:    vcvtps2ph $4, %ymm0, %xmm0
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %ymm1
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %ymm2
-; CHECK-NEXT:    vaddps %ymm1, %ymm2, %ymm1
-; CHECK-NEXT:    vcvtps2ph $4, %ymm1, %xmm1
-; CHECK-NEXT:    vinsertf128 $1, %xmm1, %ymm0, %ymm0
-; CHECK-NEXT:    movq %rbp, %rsp
-; CHECK-NEXT:    popq %rbp
+; CHECK-NEXT:    vinsertf128 $1, %xmm0, %ymm2, %ymm0
 ; CHECK-NEXT:    retq
   %1 = fadd <16 x half> %a, %b
   ret <16 x half> %1

llvm/test/CodeGen/X86/avx512-skx-v32f16-fadd.ll

@@ -4,23 +4,17 @@
 define <32 x half> @foo(<32 x half> %a, <32 x half> %b) nounwind {
 ; CHECK-LABEL: foo:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    pushq %rbp
-; CHECK-NEXT:    movq %rsp, %rbp
-; CHECK-NEXT:    andq $-64, %rsp
-; CHECK-NEXT:    subq $192, %rsp
-; CHECK-NEXT:    vmovaps %zmm1, {{[0-9]+}}(%rsp)
-; CHECK-NEXT:    vmovaps %zmm0, (%rsp)
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %zmm0
-; CHECK-NEXT:    vcvtph2ps (%rsp), %zmm1
-; CHECK-NEXT:    vaddps %zmm0, %zmm1, %zmm0
+; CHECK-NEXT:    vcvtph2ps %ymm1, %zmm2
+; CHECK-NEXT:    vcvtph2ps %ymm0, %zmm3
+; CHECK-NEXT:    vaddps %zmm2, %zmm3, %zmm2
+; CHECK-NEXT:    vcvtps2ph $4, %zmm2, %ymm2
+; CHECK-NEXT:    vextractf64x4 $1, %zmm1, %ymm1
+; CHECK-NEXT:    vcvtph2ps %ymm1, %zmm1
+; CHECK-NEXT:    vextractf64x4 $1, %zmm0, %ymm0
+; CHECK-NEXT:    vcvtph2ps %ymm0, %zmm0
+; CHECK-NEXT:    vaddps %zmm1, %zmm0, %zmm0
 ; CHECK-NEXT:    vcvtps2ph $4, %zmm0, %ymm0
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %zmm1
-; CHECK-NEXT:    vcvtph2ps {{[0-9]+}}(%rsp), %zmm2
-; CHECK-NEXT:    vaddps %zmm1, %zmm2, %zmm1
-; CHECK-NEXT:    vcvtps2ph $4, %zmm1, %ymm1
-; CHECK-NEXT:    vinsertf64x4 $1, %ymm1, %zmm0, %zmm0
-; CHECK-NEXT:    movq %rbp, %rsp
-; CHECK-NEXT:    popq %rbp
+; CHECK-NEXT:    vinsertf64x4 $1, %ymm0, %zmm2, %zmm0
 ; CHECK-NEXT:    retq
   %1 = fadd <32 x half> %a, %b
   ret <32 x half> %1