[GlobalISel] Fix bitcast fewerElements with scalar narrow types. (#153364)

For a <8 x i32> -> <2 x i128> bitcast, that under aarch64 is split into
two halfs, the scalar i128 remainder was causing problems, causing a
crash with invalid vector types. This makes sure they are handled
correctly in fewerElementsBitcast.

Author: davemgreen

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

@@ -5574,12 +5574,19 @@ LegalizerHelper::fewerElementsBitcast(MachineInstr &MI, unsigned int TypeIdx,
 
   unsigned NewElemCount =
       NarrowTy.getSizeInBits() / SrcTy.getScalarSizeInBits();
-  LLT SrcNarrowTy = LLT::fixed_vector(NewElemCount, SrcTy.getElementType());
-
-  // Split the Src and Dst Reg into smaller registers
   SmallVector<Register> SrcVRegs, BitcastVRegs;
-  if (extractGCDType(SrcVRegs, DstTy, SrcNarrowTy, SrcReg) != SrcNarrowTy)
-    return UnableToLegalize;
+  if (NewElemCount == 1) {
+    LLT SrcNarrowTy = SrcTy.getElementType();
+
+    auto Unmerge = MIRBuilder.buildUnmerge(SrcNarrowTy, SrcReg);
+    getUnmergeResults(SrcVRegs, *Unmerge);
+  } else {
+    LLT SrcNarrowTy = LLT::fixed_vector(NewElemCount, SrcTy.getElementType());
+
+    // Split the Src and Dst Reg into smaller registers
+    if (extractGCDType(SrcVRegs, DstTy, SrcNarrowTy, SrcReg) != SrcNarrowTy)
+      return UnableToLegalize;
+  }
 
   // Build new smaller bitcast instructions
   // Not supporting Leftover types for now but will have to

llvm/test/CodeGen/AArch64/bitcast.ll

@@ -617,6 +617,31 @@ define <8 x i64> @bitcast_v16i32_v8i64(<16 x i32> %a, <16 x i32> %b){
   ret <8 x i64> %d
 }
 
+define <8 x i32> @scalar_i128(<2 x i128> %a) {
+; CHECK-SD-LABEL: scalar_i128:
+; CHECK-SD:       // %bb.0:
+; CHECK-SD-NEXT:    fmov d1, x2
+; CHECK-SD-NEXT:    fmov d0, x0
+; CHECK-SD-NEXT:    mov v1.d[1], x3
+; CHECK-SD-NEXT:    mov v0.d[1], x1
+; CHECK-SD-NEXT:    add v0.4s, v0.4s, v0.4s
+; CHECK-SD-NEXT:    add v1.4s, v1.4s, v1.4s
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: scalar_i128:
+; CHECK-GI:       // %bb.0:
+; CHECK-GI-NEXT:    mov v0.d[0], x0
+; CHECK-GI-NEXT:    mov v1.d[0], x2
+; CHECK-GI-NEXT:    mov v0.d[1], x1
+; CHECK-GI-NEXT:    mov v1.d[1], x3
+; CHECK-GI-NEXT:    add v0.4s, v0.4s, v0.4s
+; CHECK-GI-NEXT:    add v1.4s, v1.4s, v1.4s
+; CHECK-GI-NEXT:    ret
+  %c = bitcast <2 x i128> %a to <8 x i32>
+  %d = add <8 x i32> %c, %c
+  ret <8 x i32> %d
+}
+
 ; ===== Vectors with Non-Pow 2 Widths =====
 
 define <6 x i16> @bitcast_v3i32_v6i16(<3 x i32> %a, <3 x i32> %b){