[AArch64][GlobalISel] Fix lowering of i64->f32 itofp.

This is a GISel equivalent of #130665, preventing a double-rounding issue in
sitofp/uitofp by scalarizing i64->f32 converts. Most of the changes are made in
the ActionDefinitionsBuilder for G_SITOFP/G_UITOFP. Because it is legal to
convert i64->f16 itofp without double-rounding, but not a fpround i64->f16,
that variant is lowered to build the two extends.

Author: davemgreen

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

@@ -7620,6 +7620,18 @@ LegalizerHelper::LegalizeResult LegalizerHelper::lowerUITOFP(MachineInstr &MI) {
     return Legalized;
   }
 
+  // i64->fp16 itofp can be lowered to i64->f64,f64->f32,f32->f16. We cannot
+  // just convert fpround f64->f16 without double-rounding, so we manually
+  // perform the lowering here where we know it is valid.
+  if (DstTy.getScalarSizeInBits() == 16 && SrcTy.getScalarSizeInBits() == 64) {
+    auto M1 = MIRBuilder.buildUITOFP(SrcTy, Src);
+    LLT S32Ty = SrcTy.changeElementSize(32);
+    auto M2 = MIRBuilder.buildFPTrunc(S32Ty, M1);
+    MIRBuilder.buildFPTrunc(Dst, M2);
+    MI.eraseFromParent();
+    return Legalized;
+  }
+
   if (SrcTy != LLT::scalar(64))
     return UnableToLegalize;
 
@@ -7651,6 +7663,18 @@ LegalizerHelper::LegalizeResult LegalizerHelper::lowerSITOFP(MachineInstr &MI) {
     return Legalized;
   }
 
+  // i64->fp16 itofp can be lowered to i64->f64,f64->f32,f32->f16. We cannot
+  // just convert fpround f64->f16 without double-rounding, so we manually
+  // perform the lowering here where we know it is valid.
+  if (DstTy.getScalarSizeInBits() == 16 && SrcTy.getScalarSizeInBits() == 64) {
+    auto M1 = MIRBuilder.buildSITOFP(SrcTy, Src);
+    LLT S32Ty = SrcTy.changeElementSize(32);
+    auto M2 = MIRBuilder.buildFPTrunc(S32Ty, M1);
+    MIRBuilder.buildFPTrunc(Dst, M2);
+    MI.eraseFromParent();
+    return Legalized;
+  }
+
   if (SrcTy != S64)
     return UnableToLegalize;
 

llvm/lib/Target/AArch64/GISel/AArch64LegalizerInfo.cpp

@@ -917,16 +917,28 @@ AArch64LegalizerInfo::AArch64LegalizerInfo(const AArch64Subtarget &ST)
       .moreElementsToNextPow2(1)
       .widenScalarOrEltToNextPow2OrMinSize(1)
       .minScalar(1, s32)
+      .lowerIf([](const LegalityQuery &Query) {
+        return Query.Types[1].isVector() &&
+               Query.Types[1].getScalarSizeInBits() == 64 &&
+               Query.Types[0].getScalarSizeInBits() == 16;
+      })
       .widenScalarOrEltToNextPow2OrMinSize(0, /*MinSize=*/HasFP16 ? 16 : 32)
+      .scalarizeIf(
+          // v2i64->v2f32 needs to scalarize to avoid double-rounding issues.
+          [](const LegalityQuery &Query) {
+            return Query.Types[0].getScalarSizeInBits() == 32 &&
+                   Query.Types[1].getScalarSizeInBits() == 64;
+          },
+          0)
       .widenScalarIf(
-          [=](const LegalityQuery &Query) {
+          [](const LegalityQuery &Query) {
             return Query.Types[1].getScalarSizeInBits() <= 64 &&
                    Query.Types[0].getScalarSizeInBits() <
                        Query.Types[1].getScalarSizeInBits();
           },
           LegalizeMutations::changeElementSizeTo(0, 1))
       .widenScalarIf(
-          [=](const LegalityQuery &Query) {
+          [](const LegalityQuery &Query) {
             return Query.Types[0].getScalarSizeInBits() <= 64 &&
                    Query.Types[0].getScalarSizeInBits() >
                        Query.Types[1].getScalarSizeInBits();