[llvm] [X86][GlobalISel] Enable scalar versions of G_UITOFP and G_FPTOUI (PR #100079)

[Matt Arsenault via llvm-commits]

================
@@ -7108,6 +7108,77 @@ LegalizerHelper::lowerU64ToF32BitOps(MachineInstr &MI) {
   return Legalized;
 }
 
+// Expand s32 = G_UITOFP s64 to an IEEE float representation using bit
+// operations and G_SITOFP
+LegalizerHelper::LegalizeResult
+LegalizerHelper::lowerU64ToF32WithSITOFP(MachineInstr &MI) {
+  auto [Dst, Src] = MI.getFirst2Regs();
+  const LLT S64 = LLT::scalar(64);
+  const LLT S32 = LLT::scalar(32);
+  const LLT S1 = LLT::scalar(1);
+
+  assert(MRI.getType(Src) == S64 && MRI.getType(Dst) == S32);
+
+  // For i64 < INT_MAX we simply reuse SITOFP.
+  // Otherwise, divide i64 by 2, round result by ORing with the lowest bit
+  // saved before division, convert to float by SITOFP, multiply the result
+  // by 2.
+  auto One = MIRBuilder.buildConstant(S64, 1);
+  auto Zero = MIRBuilder.buildConstant(S64, 0);
+  // Result if Src < INT_MAX
+  auto SmallResult = MIRBuilder.buildSITOFP(S32, Src);
+  // Result if Src >= INT_MAX
+  auto Halved = MIRBuilder.buildLShr(S64, Src, One);
+  auto LowerBit = MIRBuilder.buildAnd(S64, Src, One);
+  auto RoundedHalved = MIRBuilder.buildOr(S64, Halved, LowerBit);
+  auto HalvedFP = MIRBuilder.buildSITOFP(S32, RoundedHalved);
+  auto LargeResult = MIRBuilder.buildFAdd(S32, HalvedFP, HalvedFP);
+  // Choose
+  auto IsLarge =
+      MIRBuilder.buildICmp(CmpInst::Predicate::ICMP_SLT, S1, Src, Zero);
----------------
arsenm wrote:

IsLarge = slt x, 0? 

https://github.com/llvm/llvm-project/pull/100079