| Issue |
203414
|
| Summary |
[AArch64] bool-packing loops at -O2 compile to a movemask wrapped in a dead stack frame and redundant masking
|
| Labels |
|
| Assignees |
|
| Reporter |
joseph-isaacs
|
# What happens
The following code (https://godbolt.org/z/7nvcor895) results in a unused `sp` operation.
```cpp
uint8_t above_threshold(const float* v, float t) {
uint8_t m = 0;
for (int i = 0; i < 8; i++)
m |= (uint32_t)(v[i] > t) << i;
return m;
}
```
`clang -O2 --target=aarch64-linux-gnu` compiles this to:
```llvm
define i8 @above_threshold(ptr %v, float %t) {
%0 = load <8 x float>, ptr %v, align 4
%1 = insertelement <8 x float> poison, float %t, i64 0
%2 = shufflevector <8 x float> %1, <8 x float> poison, <8 x i32> zeroinitializer
%3 = fcmp ogt <8 x float> %0, %2
%4 = bitcast <8 x i1> %3 to i8
ret i8 %4
}
```
```asm
above_threshold:
sub sp, sp, #16 ; <- nothing below ever touches memory
ldp q1, q2, [x0]
adrp x8, .LCPI0_0
dup v0.4s, v0.s[0]
fcmgt v2.4s, v2.4s, v0.4s
fcmgt v0.4s, v1.4s, v0.4s
ldr q1, [x8, :lo12:.LCPI0_0]
uzp1 v0.8h, v0.8h, v2.8h
and v0.16b, v0.16b, v1.16b
addv h0, v0.8h
fmov w8, s0
and w0, w8, #0xff
add sp, sp, #16 ; <-
ret
```
# How it can be fixed
There are likely two fix complementary fixes:
1. Have `vectorToScalarBitmask` lower this `bitcast <8 x i1> %3 to i8` via a `8 x i16` register.
2. Find the place where the lowering via the stack is replaced and fix up the `FrameIndex`.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
