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

[Simon Pilgrim via llvm-commits]

================
@@ -497,6 +497,53 @@ X86LegalizerInfo::X86LegalizerInfo(const X86Subtarget &STI,
       .clampScalar(0, s32, sMaxScalar)
       .widenScalarToNextPow2(1);
 
+  // For G_UITOFP and G_FPTOUI without AVX512, we have to custom legalize s16
+  // manually. Otherwise, in custom handler there is no way to understand
+  // whether s32 is an original type and we need to promote it to s64 or s32 is
+  // obtained after widening s16 and we shouldn't widen it to s64.
+  //
+  // For AVX512 we simply widen types as there is direct mapping from opcodes
+  // to asm instructions.
+  getActionDefinitionsBuilder(G_UITOFP)
+      .legalIf([=](const LegalityQuery &Query) {
+        return HasAVX512 && typeInSet(0, {s32, s64})(Query) &&
+               typeInSet(1, {s32, s64})(Query);
+      })
+      .customIf([=](const LegalityQuery &Query) {
+        if (HasAVX512)
+          return false;
+        return (HasSSE1 &&
+                (typePairInSet(0, 1, {{s32, s32}, {s32, s16}})(Query) ||
+                 (Is64Bit && typePairInSet(0, 1, {{s32, s64}})(Query)))) ||
----------------
RKSimon wrote:

For uitofp, SSE has to cheat and zero-extend to i16->i32 / i32->i64 and use sitofp - so we need Is64Bit for i32->f32 as well. We only get real uitofp with AVX512

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