| Issue |
86785
|
| Summary |
[LoopVectorize] LoopVectorize produces redundant instructions due to IV widening in IndVarSimplify
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
LittleMeepo
|
```
void fun1(int* restrict A ,int* restrict B){
int i;
for(i = 1; i <= 1000; i++){
if (i % 2 == 0) {
A[i] = B[i] * B[i];
} else {
A[i] = B[i] + B[i];
}
}
}
```
Difference between GCC13.2 and LLVM:
https://godbolt.org/z/4v993aq9j
If set `-mllvm --indvars-widen-indvars=false` , the merge operation of p register will be simplified. The reason is that the type of IV is converted from i32 to i64 in IndVarSimplify.
GCC uses `.s` format insts when `i` is `int`, uses `.d` format insts when `i` is `long long`.
Can LLVM achieve similar operations? Adding options manually doesn't look very smart after all.
```
.LBB0_1: // =>This Inner Loop Header: Depth=1
mov z3.d, z1.d
add z4.d, z0.d, z2.d
and z0.d, z0.d, #0x1
add z1.d, z1.d, z2.d
and z3.d, z3.d, #0x1
cmpeq p2.d, p0/z, z0.d, #0 // <-- redundant inst
ld1w { z0.s }, p1/z, [x11, x8, lsl #2]
cmpeq p3.d, p0/z, z3.d, #0 // <-- redundant inst
mul z3.s, z0.s, z0.s
lsl z0.s, z0.s, #1
uzp1 p2.s, p2.s, p3.s // <-- redundant inst
mov z0.s, p2/m, z3.s
st1w { z0.s }, p1, [x12, x8, lsl #2]
mov z0.d, z4.d
incw x8
cmp x10, x8
b.ne .LBB0_1
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
