[llvm] [ARM] Fix VMOVRRD combine with non-canonical inserts. (PR #109639)

[via llvm-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-arm

Author: David Green (davemgreen)

<details>
<summary>Changes</summary>

In some situations, in the test case here with the multiple calls being late legalized, we can see inserts of the form:
```
     b = insert a, x, 0
    c = insert b, y, 1
   d = insert c, z, 0
  bc = bitcast d
 e = extract bc, 0
r = vmovrrd e
```
The redundent insert will usually be removed, but in some cases are not prior to PerformVMOVRRDCombine. The code was finding the last insert from each lane (x and y), as opposed to the last (z and y).

---
Full diff: https://github.com/llvm/llvm-project/pull/109639.diff


2 Files Affected:

- (modified) llvm/lib/Target/ARM/ARMISelLowering.cpp (+2-2) 
- (modified) llvm/test/CodeGen/Thumb2/mve-soft-float-abi.ll (+66) 


``````````diff
diff --git a/llvm/lib/Target/ARM/ARMISelLowering.cpp b/llvm/lib/Target/ARM/ARMISelLowering.cpp
index db564d7e841aa9..a03928b618df03 100644
--- a/llvm/lib/Target/ARM/ARMISelLowering.cpp
+++ b/llvm/lib/Target/ARM/ARMISelLowering.cpp
@@ -15131,9 +15131,9 @@ static SDValue PerformVMOVRRDCombine(SDNode *N,
     SDValue Op0, Op1;
     while (BV.getOpcode() == ISD::INSERT_VECTOR_ELT) {
       if (isa<ConstantSDNode>(BV.getOperand(2))) {
-        if (BV.getConstantOperandVal(2) == Offset)
+        if (BV.getConstantOperandVal(2) == Offset && !Op0)
           Op0 = BV.getOperand(1);
-        if (BV.getConstantOperandVal(2) == Offset + 1)
+        if (BV.getConstantOperandVal(2) == Offset + 1 && !Op1)
           Op1 = BV.getOperand(1);
       }
       BV = BV.getOperand(0);
diff --git a/llvm/test/CodeGen/Thumb2/mve-soft-float-abi.ll b/llvm/test/CodeGen/Thumb2/mve-soft-float-abi.ll
index 4b76906034057f..41d2c02a73cd0a 100644
--- a/llvm/test/CodeGen/Thumb2/mve-soft-float-abi.ll
+++ b/llvm/test/CodeGen/Thumb2/mve-soft-float-abi.ll
@@ -543,3 +543,69 @@ define <4 x i32> @insertextract(i32 %x, i32 %y) {
   %4 = insertelement <4 x i32> %3, i32 %y, i32 3
   ret <4 x i32> %4
 }
+
+declare void @print_uint32x4_t(<4 x i32> %val)
+define i32 @main(i64 %x, i64 %y) {
+; CHECK-LE-LABEL: main:
+; CHECK-LE:       @ %bb.0: @ %entry
+; CHECK-LE-NEXT:    .save {r4, lr}
+; CHECK-LE-NEXT:    push {r4, lr}
+; CHECK-LE-NEXT:    .vsave {d8, d9}
+; CHECK-LE-NEXT:    vpush {d8, d9}
+; CHECK-LE-NEXT:    .pad #8
+; CHECK-LE-NEXT:    sub sp, #8
+; CHECK-LE-NEXT:    vmov.32 q4[2], r2
+; CHECK-LE-NEXT:    mov r4, r1
+; CHECK-LE-NEXT:    mov r1, r0
+; CHECK-LE-NEXT:    vmov.32 q4[3], r3
+; CHECK-LE-NEXT:    movs r0, #0
+; CHECK-LE-NEXT:    mov r2, r1
+; CHECK-LE-NEXT:    mov r3, r4
+; CHECK-LE-NEXT:    vstr d9, [sp]
+; CHECK-LE-NEXT:    bl print_uint32x4_t
+; CHECK-LE-NEXT:    movs r0, #0
+; CHECK-LE-NEXT:    movs r2, #1
+; CHECK-LE-NEXT:    mov r3, r4
+; CHECK-LE-NEXT:    vstr d9, [sp]
+; CHECK-LE-NEXT:    bl print_uint32x4_t
+; CHECK-LE-NEXT:    movs r0, #0
+; CHECK-LE-NEXT:    add sp, #8
+; CHECK-LE-NEXT:    vpop {d8, d9}
+; CHECK-LE-NEXT:    pop {r4, pc}
+;
+; CHECK-BE-LABEL: main:
+; CHECK-BE:       @ %bb.0: @ %entry
+; CHECK-BE-NEXT:    .save {r4, lr}
+; CHECK-BE-NEXT:    push {r4, lr}
+; CHECK-BE-NEXT:    .vsave {d8, d9}
+; CHECK-BE-NEXT:    vpush {d8, d9}
+; CHECK-BE-NEXT:    .pad #8
+; CHECK-BE-NEXT:    sub sp, #8
+; CHECK-BE-NEXT:    vmov.32 q0[2], r2
+; CHECK-BE-NEXT:    mov r4, r1
+; CHECK-BE-NEXT:    mov r1, r0
+; CHECK-BE-NEXT:    vmov.32 q0[3], r3
+; CHECK-BE-NEXT:    vrev64.32 q4, q0
+; CHECK-BE-NEXT:    movs r0, #0
+; CHECK-BE-NEXT:    mov r2, r1
+; CHECK-BE-NEXT:    mov r3, r4
+; CHECK-BE-NEXT:    vstr d9, [sp]
+; CHECK-BE-NEXT:    bl print_uint32x4_t
+; CHECK-BE-NEXT:    movs r0, #0
+; CHECK-BE-NEXT:    movs r2, #1
+; CHECK-BE-NEXT:    mov r3, r4
+; CHECK-BE-NEXT:    vstr d9, [sp]
+; CHECK-BE-NEXT:    bl print_uint32x4_t
+; CHECK-BE-NEXT:    movs r0, #0
+; CHECK-BE-NEXT:    add sp, #8
+; CHECK-BE-NEXT:    vpop {d8, d9}
+; CHECK-BE-NEXT:    pop {r4, pc}
+entry:
+  %a = insertelement <2 x i64> poison, i64 %x, i64 0
+  %b = insertelement <2 x i64> %a, i64 %y, i64 1
+  %c = bitcast <2 x i64> %b to <4 x i32>
+  %i = insertelement <4 x i32> %c, i32 1, i64 0
+  tail call void @print_uint32x4_t(i32 0, <4 x i32> %c)
+  tail call void @print_uint32x4_t(i32 0, <4 x i32> %i)
+  ret i32 0
+}

``````````

</details>


https://github.com/llvm/llvm-project/pull/109639