[X86] Fold (f16 bitcast extract_vectorelt(v,0)) to (extract_vectorelt (v8f16 bitcast(v,0))) (#125877)

Also handles possible truncations from i32 to i16.

Cleans up some of the poor codegen identified in #98630

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -45160,6 +45160,19 @@ static SDValue combineBitcast(SDNode *N, SelectionDAG &DAG,
     }
   }
 
+  // Attempt to peek through f16 bitcasted extractions hidden by truncation.
+  if (VT == MVT::f16 && SrcVT == MVT::i16) {
+    SDValue Src = peekThroughTruncates(N0);
+    if (Src.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
+        Src.getOperand(0).getValueSizeInBits() == 128 &&
+        isNullConstant(Src.getOperand(1))) {
+      SDLoc DL(N);
+      return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
+                         DAG.getBitcast(MVT::v8f16, Src.getOperand(0)),
+                         DAG.getVectorIdxConstant(0, DL));
+    }
+  }
+
   // Since MMX types are special and don't usually play with other vector types,
   // it's better to handle them early to be sure we emit efficient code by
   // avoiding store-load conversions.

llvm/test/CodeGen/X86/bfloat.ll

@@ -82,8 +82,6 @@ define bfloat @add2(bfloat %a, bfloat %b) nounwind {
 ; X86-NEXT:    vmovd %eax, %xmm1
 ; X86-NEXT:    vaddss %xmm0, %xmm1, %xmm0
 ; X86-NEXT:    vcvtneps2bf16 %xmm0, %xmm0
-; X86-NEXT:    vmovw %xmm0, %eax
-; X86-NEXT:    vmovw %eax, %xmm0
 ; X86-NEXT:    retl
 ;
 ; SSE2-LABEL: add2:
@@ -110,8 +108,6 @@ define bfloat @add2(bfloat %a, bfloat %b) nounwind {
 ; FP16-NEXT:    vmovd %eax, %xmm1
 ; FP16-NEXT:    vaddss %xmm0, %xmm1, %xmm0
 ; FP16-NEXT:    vcvtneps2bf16 %xmm0, %xmm0
-; FP16-NEXT:    vmovw %xmm0, %eax
-; FP16-NEXT:    vmovw %eax, %xmm0
 ; FP16-NEXT:    retq
 ;
 ; AVXNC-LABEL: add2:
@@ -124,8 +120,6 @@ define bfloat @add2(bfloat %a, bfloat %b) nounwind {
 ; AVXNC-NEXT:    vmovd %eax, %xmm1
 ; AVXNC-NEXT:    vaddss %xmm0, %xmm1, %xmm0
 ; AVXNC-NEXT:    {vex} vcvtneps2bf16 %xmm0, %xmm0
-; AVXNC-NEXT:    vmovd %xmm0, %eax
-; AVXNC-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
 ; AVXNC-NEXT:    retq
   %add = fadd bfloat %a, %b
   ret bfloat %add
@@ -432,8 +426,6 @@ define bfloat @add_constant2(bfloat %a) nounwind {
 ; X86-NEXT:    vmovd %eax, %xmm0
 ; X86-NEXT:    vaddss {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm0
 ; X86-NEXT:    vcvtneps2bf16 %xmm0, %xmm0
-; X86-NEXT:    vmovw %xmm0, %eax
-; X86-NEXT:    vmovw %eax, %xmm0
 ; X86-NEXT:    retl
 ;
 ; SSE2-LABEL: add_constant2:
@@ -454,8 +446,6 @@ define bfloat @add_constant2(bfloat %a) nounwind {
 ; FP16-NEXT:    vmovd %eax, %xmm0
 ; FP16-NEXT:    vaddss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
 ; FP16-NEXT:    vcvtneps2bf16 %xmm0, %xmm0
-; FP16-NEXT:    vmovw %xmm0, %eax
-; FP16-NEXT:    vmovw %eax, %xmm0
 ; FP16-NEXT:    retq
 ;
 ; AVXNC-LABEL: add_constant2:
@@ -465,8 +455,6 @@ define bfloat @add_constant2(bfloat %a) nounwind {
 ; AVXNC-NEXT:    vmovd %eax, %xmm0
 ; AVXNC-NEXT:    vaddss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
 ; AVXNC-NEXT:    {vex} vcvtneps2bf16 %xmm0, %xmm0
-; AVXNC-NEXT:    vmovd %xmm0, %eax
-; AVXNC-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
 ; AVXNC-NEXT:    retq
   %add = fadd bfloat %a, 1.0
   ret bfloat %add

llvm/test/CodeGen/X86/canonicalize-vars-f16-type.ll

@@ -154,8 +154,6 @@ define half @complex_canonicalize_fmul_half(half %a, half %b) nounwind {
 ; AVX512-NEXT:    vcvtph2ps %xmm0, %xmm0
 ; AVX512-NEXT:    vsubss %xmm1, %xmm0, %xmm0
 ; AVX512-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
-; AVX512-NEXT:    vmovd %xmm0, %eax
-; AVX512-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
 ; AVX512-NEXT:    retq
 entry:
 
@@ -239,15 +237,11 @@ define void @v_test_canonicalize_v2half(<2 x half> addrspace(1)* %out) nounwind
 ; AVX512-NEXT:    vxorps %xmm3, %xmm3, %xmm3
 ; AVX512-NEXT:    vblendps {{.*#+}} xmm2 = xmm2[0],xmm3[1,2,3]
 ; AVX512-NEXT:    vcvtps2ph $4, %xmm2, %xmm2
-; AVX512-NEXT:    vmovd %xmm2, %eax
-; AVX512-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm2
 ; AVX512-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
 ; AVX512-NEXT:    vcvtph2ps %xmm0, %xmm0
 ; AVX512-NEXT:    vmulss %xmm1, %xmm0, %xmm0
 ; AVX512-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm3[1,2,3]
 ; AVX512-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
-; AVX512-NEXT:    vmovd %xmm0, %eax
-; AVX512-NEXT:    vpinsrw $0, %eax, %xmm0, %xmm0
 ; AVX512-NEXT:    vpunpcklwd {{.*#+}} xmm0 = xmm0[0],xmm2[0],xmm0[1],xmm2[1],xmm0[2],xmm2[2],xmm0[3],xmm2[3]
 ; AVX512-NEXT:    vmovd %xmm0, (%rdi)
 ; AVX512-NEXT:    retq