[X86] Remove restriction to pre-type legalization for combine in scalarizeExtEltFP

[abhishek-kaushik22]

Remove the restriction to pre-type legalization from combine since it was already not being enforced for AVX512 where v8i1 is a legal type but i1 itself is not.

Closes #117684

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: None (abhishek-kaushik22)

Changes

Remove the restriction to pre-type legalization from combine since it was already not being enforced for AVX512 where v8i1 is a legal type but i1 itself is not.

---

Full diff: https://github.com/llvm/llvm-project/pull/117682.diff

2 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+21-13)
• (modified) llvm/test/CodeGen/X86/extractelement-fp.ll (+12-4)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index e4533570f75086..0a0c54492c2995 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -45842,7 +45842,8 @@ static SDValue combineExtractWithShuffle(SDNode *N, SelectionDAG &DAG,
 /// Extracting a scalar FP value from vector element 0 is free, so extract each
 /// operand first, then perform the math as a scalar op.
 static SDValue scalarizeExtEltFP(SDNode *ExtElt, SelectionDAG &DAG,
-                                 const X86Subtarget &Subtarget) {
+                                 const X86Subtarget &Subtarget,
+                                 TargetLowering::DAGCombinerInfo &DCI) {
   assert(ExtElt->getOpcode() == ISD::EXTRACT_VECTOR_ELT && "Expected extract");
   SDValue Vec = ExtElt->getOperand(0);
   SDValue Index = ExtElt->getOperand(1);
@@ -45877,23 +45878,30 @@ static SDValue scalarizeExtEltFP(SDNode *ExtElt, SelectionDAG &DAG,
   // Vector FP selects don't fit the pattern of FP math ops (because the
   // condition has a different type and we have to change the opcode), so deal
   // with those here.
-  // FIXME: This is restricted to pre type legalization by ensuring the setcc
-  // has i1 elements. If we loosen this we need to convert vector bool to a
-  // scalar bool.
   if (Vec.getOpcode() == ISD::VSELECT &&
       Vec.getOperand(0).getOpcode() == ISD::SETCC &&
       Vec.getOperand(0).getValueType().getScalarType() == MVT::i1 &&
       Vec.getOperand(0).getOperand(0).getValueType() == VecVT) {
     // ext (sel Cond, X, Y), 0 --> sel (ext Cond, 0), (ext X, 0), (ext Y, 0)
     SDLoc DL(ExtElt);
-    SDValue Ext0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
-                               Vec.getOperand(0).getValueType().getScalarType(),
-                               Vec.getOperand(0), Index);
-    SDValue Ext1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
-                               Vec.getOperand(1), Index);
-    SDValue Ext2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
-                               Vec.getOperand(2), Index);
-    return DAG.getNode(ISD::SELECT, DL, VT, Ext0, Ext1, Ext2);
+    bool AfterLegalize = DCI.getDAGCombineLevel() >= llvm::AfterLegalizeTypes;
+    EVT VecOperandType = Vec.getOperand(0).getValueType();
+    EVT LegalType = DAG.getTargetLoweringInfo().getTypeToTransformTo(
+        *DAG.getContext(), VecOperandType.getScalarType());
+    // The second condition checks that we don't create an invalid extract e.g
+    // 32 bit extract from a v*i64. This will cause a crash on 32-bit machines.
+    if (!AfterLegalize ||
+        LegalType.getSizeInBits() >= VecOperandType.getScalarSizeInBits()) {
+      SDValue Ext0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL,
+                                 AfterLegalize ? LegalType
+                                               : VecOperandType.getScalarType(),
+                                 Vec.getOperand(0), Index);
+      SDValue Ext1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
+                                 Vec.getOperand(1), Index);
+      SDValue Ext2 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
+                                 Vec.getOperand(2), Index);
+      return DAG.getNode(ISD::SELECT, DL, VT, Ext0, Ext1, Ext2);
+    }
   }
 
   // TODO: This switch could include FNEG and the x86-specific FP logic ops
@@ -46242,7 +46250,7 @@ static SDValue combineExtractVectorElt(SDNode *N, SelectionDAG &DAG,
   if (SDValue V = combineArithReduction(N, DAG, Subtarget))
     return V;
 
-  if (SDValue V = scalarizeExtEltFP(N, DAG, Subtarget))
+  if (SDValue V = scalarizeExtEltFP(N, DAG, Subtarget, DCI))
     return V;
 
   if (CIdx)
diff --git a/llvm/test/CodeGen/X86/extractelement-fp.ll b/llvm/test/CodeGen/X86/extractelement-fp.ll
index 944f6bbfd0bfbe..967493d3ac608b 100644
--- a/llvm/test/CodeGen/X86/extractelement-fp.ll
+++ b/llvm/test/CodeGen/X86/extractelement-fp.ll
@@ -320,10 +320,18 @@ 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:    vmovsd {{.*#+}} xmm2 = [1.0E+0,0.0E+0]
-; X64-NEXT:    vandpd %xmm2, %xmm1, %xmm1
-; X64-NEXT:    vinsertf128 $1, %xmm0, %ymm1, %ymm0
-; X64-NEXT:    vpermpd {{.*#+}} ymm0 = ymm0[0,2,3,3]
+; X64-NEXT:    vmovq   %xmm1, %rax
+; X64-NEXT:    testq   %rax, %rax
+; X64-NEXT:    jne     .{{\.?LBB[0-9]+_[0-9]+}}
+; X64-NEXT:  # %bb.2:
+; X64-NEXT:    vxorpd  %xmm1, %xmm1, %xmm1
+; X64-NEXT:    jmp     .{{\.?LBB[0-9]+_[0-9]+}}
+; X64-NEXT:  .{{\.?LBB[0-9]+_[0-9]+}}:
+; X64-NEXT:    vmovsd  {{.*#+}} xmm1 = [1.0E+0,0.0E+0]
+; X64-NEXT:  .{{\.?LBB[0-9]+_[0-9]+}}:
+; X64-NEXT:    vshufpd $1, %xmm0, %xmm0, %xmm2 # xmm2 = xmm0[1,0]
+; X64-NEXT:    vunpcklpd       %xmm0, %xmm1, %xmm0 # xmm0 = xmm1[0],xmm0[0]
+; X64-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
 ; X64-NEXT:    retq
 ;
 ; X86-LABEL: extvselectsetcc_crash:

[phoebewang]

This looks regression to me.

[RKSimon]

Please can you regenerate extractelement-fp.ll as there looks to be other changes

[RKSimon]

(sorry I had missed that this had been replaced with #117681)