[X86] SimplifyDemandedVectorEltsForTargetNode - replace packed fcmp node with scalar fcmp node if only element0 is demanded (#140563)

These unnecessary vectorisation can appear due to fplogic opcodes only being available for 128-bit types - which can prevent folds that only work on the scalar source types and also lead to fcmp of garbage data in the upper elements.

Fixes #140534

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -43450,6 +43450,28 @@ bool X86TargetLowering::SimplifyDemandedVectorEltsForTargetNode(
     KnownZero = LHSZero;
     break;
   }
+  case X86ISD::CMPM:
+  case X86ISD::CMPP: {
+    // Scalarize packed fp comparison if we only require element 0.
+    if (DemandedElts == 1) {
+      SDLoc dl(Op);
+      MVT VT = Op.getSimpleValueType();
+      MVT OpSVT = Op.getOperand(0).getSimpleValueType().getScalarType();
+      SDValue LHS = TLO.DAG.getExtractVectorElt(dl, OpSVT, Op.getOperand(0), 0);
+      SDValue RHS = TLO.DAG.getExtractVectorElt(dl, OpSVT, Op.getOperand(1), 0);
+      SDValue CC = Op.getOperand(2);
+      if (Opc == X86ISD::CMPM) {
+        SDValue Cmp =
+            TLO.DAG.getNode(X86ISD::FSETCCM, dl, MVT::v1i1, LHS, RHS, CC);
+        return TLO.CombineTo(
+            Op, TLO.DAG.getInsertSubvector(dl, TLO.DAG.getUNDEF(VT), Cmp, 0));
+      }
+      SDValue Cmp = TLO.DAG.getNode(X86ISD::FSETCC, dl, OpSVT, LHS, RHS, CC);
+      return TLO.CombineTo(Op,
+                           TLO.DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Cmp));
+    }
+    break;
+  }
   case X86ISD::PCMPEQ:
   case X86ISD::PCMPGT: {
     APInt LHSUndef, LHSZero;

llvm/test/CodeGen/X86/and-or-setcc.ll

@@ -17,8 +17,8 @@ define i1 @and_ord(float %a, float %b) {
 ; X64-LABEL: and_ord:
 ; X64:       # %bb.0:
 ; X64-NEXT:    xorps %xmm2, %xmm2
-; X64-NEXT:    cmpordps %xmm2, %xmm1
-; X64-NEXT:    cmpordps %xmm2, %xmm0
+; X64-NEXT:    cmpordss %xmm2, %xmm1
+; X64-NEXT:    cmpordss %xmm2, %xmm0
 ; X64-NEXT:    andps %xmm1, %xmm0
 ; X64-NEXT:    movd %xmm0, %eax
 ; X64-NEXT:    # kill: def $al killed $al killed $eax
@@ -44,8 +44,8 @@ define i1 @or_uno(float %a, float %b) {
 ; X64-LABEL: or_uno:
 ; X64:       # %bb.0:
 ; X64-NEXT:    xorps %xmm2, %xmm2
-; X64-NEXT:    cmpunordps %xmm2, %xmm1
-; X64-NEXT:    cmpunordps %xmm2, %xmm0
+; X64-NEXT:    cmpunordss %xmm2, %xmm1
+; X64-NEXT:    cmpunordss %xmm2, %xmm0
 ; X64-NEXT:    orps %xmm1, %xmm0
 ; X64-NEXT:    movd %xmm0, %eax
 ; X64-NEXT:    # kill: def $al killed $al killed $eax

llvm/test/CodeGen/X86/extract-vselect-setcc.ll

@@ -5,10 +5,9 @@ define void @PR117684(i1 %cond, <8 x float> %vec, ptr %ptr1, ptr %ptr2) #0 {
 ; CHECK-LABEL: PR117684:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vxorps %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vcmpnltps %ymm1, %ymm0, %k1
+; CHECK-NEXT:    vcmpnltss %xmm1, %xmm0, %k1
 ; CHECK-NEXT:    vbroadcastss {{.*#+}} xmm0 = [NaN,NaN,NaN,NaN]
 ; CHECK-NEXT:    vinsertf32x4 $0, %xmm0, %ymm0, %ymm0 {%k1} {z}
-; CHECK-NEXT:    vxorps %xmm1, %xmm1, %xmm1
 ; CHECK-NEXT:    vmulss %xmm1, %xmm0, %xmm0
 ; CHECK-NEXT:    vmulss {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm2
 ; CHECK-NEXT:    vbroadcastss %xmm2, %ymm2

llvm/test/CodeGen/X86/extractelement-fp.ll

@@ -319,7 +319,7 @@ define void @extsetcc(<4 x float> %x) {
 define <3 x double> @extvselectsetcc_crash(<2 x double> %x) {
 ; X64-LABEL: extvselectsetcc_crash:
 ; X64:       # %bb.0:
-; X64-NEXT:    vcmpeqpd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1
+; X64-NEXT:    vcmpeqsd {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm1
 ; X64-NEXT:    vmovsd {{.*#+}} xmm2 = [1.0E+0,0.0E+0]
 ; X64-NEXT:    vandpd %xmm2, %xmm1, %xmm1
 ; X64-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
@@ -328,7 +328,7 @@ define <3 x double> @extvselectsetcc_crash(<2 x double> %x) {
 ;
 ; X86-LABEL: extvselectsetcc_crash:
 ; X86:       # %bb.0:
-; X86-NEXT:    vcmpeqpd {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm1
+; X86-NEXT:    vcmpeqsd {{\.?LCPI[0-9]+_[0-9]+}}, %xmm0, %xmm1
 ; X86-NEXT:    vmovsd {{.*#+}} xmm2 = [1.0E+0,0.0E+0]
 ; X86-NEXT:    vandpd %xmm2, %xmm1, %xmm1
 ; X86-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0