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

[Simon Pilgrim via llvm-commits]

================
@@ -23391,6 +23391,136 @@ 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 == 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.value_or(false);
+    };
+
+    bool OkayCvt = false;
+    bool OkayBitcast = false;
+
+    const fltSemantics &Sem = SelectionDAG::EVTToAPFloatSemantics(MVT::f32);
+
+    // For cvt up to 1 << (Significand Precision), (1 << 24 for ieee float)
+    const APInt MaxConvertableCvt =
+        APInt::getOneBitSet(32, APFloat::semanticsPrecision(Sem));
+    // For bitcast up to (and including) first inf representation (0x7f800000 +
+    // 1 for ieee float)
+    const APInt MaxConvertableBitcast =
+        APFloat::getInf(Sem).bitcastToAPInt() + 1;
+    // For bitcast we also exclude de-norm values. This is absolutely necessary
+    // for strict semantic correctness, but DAZ (de-norm as zero) will break if
+    // we don't have this check.
+    const APInt MinConvertableBitcast =
+        APFloat::getSmallestNormalized(Sem).bitcastToAPInt() - 1;
+
+    assert(
+        MaxConvertableBitcast.getBitWidth() == 32 &&
+        MaxConvertableCvt == (1U << 24) &&
+        MaxConvertableBitcast == 0x7f800001 &&
+        MinConvertableBitcast.isNonNegative() &&
+        MaxConvertableBitcast.sgt(MinConvertableBitcast) &&
+        "This transform has only been verified to IEEE Single Precision Float");
+
+    // For bitcast we need both lhs/op1 u< MaxConvertableBitcast
+    // NB: It might be worth it to enable to bitcast version for unsigned avx2
+    // comparisons as they typically require multiple instructions to lower
+    // (they don't fit `vpcmpeq`/`vpcmpgt` well).
+    if (OpInRange(1, /*CmpLT*/ true, MaxConvertableBitcast) &&
+        OpInRange(1, /*CmpLT*/ false, MinConvertableBitcast) &&
+        OpInRange(0, /*CmpLT*/ true, MaxConvertableBitcast) &&
+        OpInRange(0, /*CmpLT*/ false, MinConvertableBitcast)) {
+      OkayBitcast = true;
+    }
+    // We want to convert icmp -> fcmp using `sitofp` iff one of the converts
+    // will be constant folded.
+    else if ((DAG.isConstantValueOfAnyType(peekThroughBitcasts(Op1)) ||
+              DAG.isConstantValueOfAnyType(peekThroughBitcasts(Op0)))) {
+      if (isUnsignedIntSetCC(Cond)) {
+        // For cvt + unsigned compare we need both lhs/rhs >= 0 and either lhs
+        // or rhs < MaxConvertableCvt
+
+        if (OpInRange(1, /*CmpLT*/ true, APInt::getSignedMinValue(32)) &&
+            OpInRange(0, /*CmpLT*/ true, APInt::getSignedMinValue(32)) &&
----------------
RKSimon wrote:

@goldsteinn The comment says both lhs/rhs >= 0 but the comparisons are with MIN_INT ?

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