[llvm] [X86] Fold concat(PCMP*(), PCMP*()) -> CMPPS(concat, concat) on AVX1 targets (PR #95915)

[via llvm-commits]

================
@@ -56111,11 +56133,50 @@ static SDValue combineConcatVectorOps(const SDLoc &DL, MVT VT,
       break;
     case X86ISD::PCMPEQ:
     case X86ISD::PCMPGT:
-      if (!IsSplat && VT.is256BitVector() && Subtarget.hasInt256() &&
+      if (!IsSplat && VT.is256BitVector() &&
+          (Subtarget.hasInt256() || VT == MVT::v8i32) &&
           (IsConcatFree(VT, Ops, 0) || IsConcatFree(VT, Ops, 1))) {
-        return DAG.getNode(Op0.getOpcode(), DL, VT,
-                           ConcatSubOperand(VT, Ops, 0),
-                           ConcatSubOperand(VT, Ops, 1));
+        if (Subtarget.hasInt256())
+          return DAG.getNode(Op0.getOpcode(), DL, VT,
+                             ConcatSubOperand(VT, Ops, 0),
+                             ConcatSubOperand(VT, Ops, 1));
+
+        // Without AVX2, see if we can cast the values to v8f32 and use fcmp.
+        // TODO: Handle v4f64 as well?
+        unsigned MaxSigBitsLHS = 0, MaxSigBitsRHS = 0;
+        for (unsigned I = 0; I != NumOps; ++I) {
+          MaxSigBitsLHS =
+              std::max(MaxSigBitsLHS,
+                       DAG.ComputeMaxSignificantBits(Ops[I].getOperand(0)));
+          MaxSigBitsRHS =
+              std::max(MaxSigBitsRHS,
+                       DAG.ComputeMaxSignificantBits(Ops[I].getOperand(1)));
+          if (MaxSigBitsLHS == EltSizeInBits && MaxSigBitsRHS == EltSizeInBits)
----------------
goldsteinn wrote:

Ah yes :)

https://github.com/llvm/llvm-project/pull/95915