[SelectionDAG][X86] Support inline assembly returning an mmx register into a type with fewer than 64 bits.

[craig.topper]

Summary

It's possible to use the 'y' mmx constraint with a type narrower than 64-bits.

This patch supports this by bitcasting the mmx type to 64-bits and then
truncating to the desired type.

There are probably other missing type combinations we need to support, but this
is the case we have a bug report for.

Fixes PR41748.

[mgorny]

Thanks. I can confirm that it fixes my original program.

[RKSimon]

LGTM as a stopgap

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

329

extra whitespace but is it worth just using MVT::i64?

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

if (PartEVT.isFloatingPoint() && ValueVT.isFloatingPoint()) {
     if (ValueVT.bitsLT(Val.getValueType()))
       return DAG.getNode(
           ISD::FP_ROUND, DL, ValueVT, Val,
           DAG.getTargetConstant(1, DL, TLI.getPointerTy(DAG.getDataLayout())));
 
     return DAG.getNode(ISD::FP_EXTEND, DL, ValueVT, Val);
   }
 
+  // Handle MMX to a narrower integer type by bitcasting MMX to integer and
+  // then truncating.
+  if (PartEVT == MVT::x86mmx && ValueVT.isInteger() &&
+      ValueVT.bitsLT(PartEVT)) {
+    PartEVT = EVT::getIntegerVT(*DAG.getContext(),  PartEVT.getSizeInBits());
+    Val = DAG.getNode(ISD::BITCAST, DL, PartEVT, Val);
+    return DAG.getNode(ISD::TRUNCATE, DL, ValueVT, Val);
+  }
+
   report_fatal_error("Unknown mismatch in getCopyFromParts!");
 }
 
 static void diagnosePossiblyInvalidConstraint(LLVMContext &Ctx, const Value *V,
                                               const Twine &ErrMsg) {
   const Instruction *I = dyn_cast_or_null<Instruction>(V);
   if (!V)
     return Ctx.emitError(ErrMsg);

llvm/test/CodeGen/X86/pr41748.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-apple-macosx10.14.0 -mattr=mmx | FileCheck %s
+
+define i32 @foo(i32 %a) {
+; CHECK-LABEL: foo:
+; CHECK:       ## %bb.0: ## %entry
+; CHECK-NEXT:    ## InlineAsm Start
+; CHECK-NEXT:    movd %edi, %mm0
+; CHECK-NEXT:    ## InlineAsm End
+; CHECK-NEXT:    movd %mm0, %eax
+; CHECK-NEXT:    retq
+entry:
+  %0 = tail call i32 asm sideeffect "movd    $1, $0", "=y,r,~{dirflag},~{fpsr},~{flags}"(i32 %a)
+  ret i32 %0
+}