[llvm] [X86] Try Folding icmp of v8i32 -> fcmp of v8f32 on AVX (PR #82290)

[Simon Pilgrim via llvm-commits]

================
@@ -23299,6 +23299,110 @@ static SDValue LowerVSETCC(SDValue Op, const X86Subtarget &Subtarget,
     }
   }
 
+  // We get bad codegen for v8i32 compares on avx targets (without avx2) so if
+  // possible convert to a v8f32 compare.
+  if (VTOp0.getVectorElementType() == MVT::i32 && VTOp0 == MVT::v8i32 &&
+      Subtarget.hasAVX() && !Subtarget.hasAVX2()) {
+    std::optional<KnownBits> KnownOps[2];
+    // Check if an op is known to be in a certain range.
+    auto OpInRange = [&DAG, Op, &KnownOps](unsigned OpNo, bool CmpLT,
+                                           const APInt Bound) {
+      if (!KnownOps[OpNo].has_value())
+        KnownOps[OpNo] = DAG.computeKnownBits(Op.getOperand(OpNo));
+
+      if (KnownOps[OpNo]->isUnknown())
+        return false;
+
+      std::optional<bool> Res;
+      if (CmpLT)
+        Res = KnownBits::ult(*KnownOps[OpNo], KnownBits::makeConstant(Bound));
+      else
+        Res = KnownBits::ugt(*KnownOps[OpNo], KnownBits::makeConstant(Bound));
+      return Res.has_value() && *Res;
+    };
+
+    bool OkayCvt = false;
+    bool OkayBitcast = false;
+
+    // For cvt up to 1 << (Significand Precision)
----------------
RKSimon wrote:

Would it make sense to use the APFloat semantics instead of hard coded numbers here?

https://github.com/llvm/llvm-project/pull/82290