[TargetLowering] Use Correct VT for Multi-out Asm (#116024)

This was overlooked in 7d94043 - when
inline assembly has multiple outputs, they are returned as members of a
struct, and the `getAsmOperandType` needs to be called for each member
of struct. The difference between this and the single-output case is
that in the latter, there isn't a struct wrapping the outputs.

I noticed this when trying to use the same mechanism in the RISC-V
backend.

Committing two tests:
- One that shows a crash before this change, which is fixed by this
change.
- One (commented out) that shows a different crash with tied
inputs/outputs. This is commented as it is not fixed by this change and
needs more work in target-independent inline asm handling code.

Author: lenary

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

@@ -5753,7 +5753,8 @@ TargetLowering::ParseConstraints(const DataLayout &DL,
       assert(!Call.getType()->isVoidTy() && "Bad inline asm!");
       if (auto *STy = dyn_cast<StructType>(Call.getType())) {
         OpInfo.ConstraintVT =
-            getSimpleValueType(DL, STy->getElementType(ResNo));
+            getAsmOperandValueType(DL, STy->getElementType(ResNo))
+                .getSimpleVT();
       } else {
         assert(ResNo == 0 && "Asm only has one result!");
         OpInfo.ConstraintVT =

llvm/test/CodeGen/AArch64/ls64-inline-asm.ll

@@ -1,8 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=aarch64 -mattr=+ls64 -verify-machineinstrs -o - %s | FileCheck %s
 
-%struct.foo = type { [8 x i64] }
-
 define void @load(ptr %output, ptr %addr) {
 ; CHECK-LABEL: load:
 ; CHECK:       // %bb.0: // %entry
@@ -103,3 +101,37 @@ entry:
   call void asm sideeffect "st64b $0,[$1]", "r,r,~{memory}"(i512 %s.sroa.0.0.insert.insert, ptr %addr)
   ret void
 }
+
+define void @multi_output(ptr %addr) {
+; CHECK-LABEL: multi_output:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    //APP
+; CHECK-NEXT:    ld64b x0, [x0]
+; CHECK-NEXT:    mov x8, x0
+; CHECK-NEXT:    //NO_APP
+; CHECK-NEXT:    stp x6, x7, [x8, #48]
+; CHECK-NEXT:    stp x4, x5, [x8, #32]
+; CHECK-NEXT:    stp x2, x3, [x8, #16]
+; CHECK-NEXT:    stp x0, x1, [x8]
+; CHECK-NEXT:    ret
+entry:
+  %val = call { i512, ptr } asm sideeffect "ld64b $0, [$2]; mov $1, $2", "=r,=r,r,~{memory}"(ptr %addr)
+  %val0 = extractvalue { i512, ptr } %val, 0
+  %val1 = extractvalue { i512, ptr } %val, 1
+  store i512 %val0, ptr %val1, align 8
+  ret void
+}
+
+; FIXME: This case still crashes in RegsForValue::AddInlineAsmOperands without
+; additional changes. I believe this is a bug in target-independent code, that
+; is worked around in the RISC-V and SystemZ backends, but should almost
+; certainly be fixed instead.
+; define void @tied_constraints(ptr %addr) {
+; entry:
+;   %in = load i512, ptr %addr, align 8
+;   %val = call { i512, ptr } asm sideeffect "nop", "=r,=r,0,1,~{memory}"(i512 %in, ptr %addr)
+;   %val0 = extractvalue { i512, ptr } %val, 0
+;   %val1 = extractvalue { i512, ptr } %val, 1
+;   store i512 %val0, ptr %val1, align 8
+;   ret void
+; }