| Issue |
126892
|
| Summary |
[X86] Unnecessary promotion for fneg / fabs half calls
|
| Labels |
good first issue,
backend:X86,
missed-optimization,
llvm:SelectionDAG
|
| Assignees |
|
| Reporter |
RKSimon
|
https://gcc.godbolt.org/z/PGshMcsTY
```cpp
_Float16 fabs_f16(_Float16 x) {
return static_cast<_Float16>(__builtin_elementwise_abs(x));
}
_Float16 fneg_f16(_Float16 x) {
return static_cast<_Float16>(-(x));
}
```
These shouldn't require vcvtph2ps/vcvtps2ph or __extendhfsf2/__truncsfhf2 calls - we should be able to use the bitwise ops directly on the v8f16 type.
```asm
.LCPI0_0:
.long 0x7fffffff # float NaN
.long 0x7fffffff # float NaN
.long 0x7fffffff # float NaN
.long 0x7fffffff # float NaN
fabs_f16(_Float16): # @fabs_f16(_Float16)
# %bb.0: # %entry
vcvtph2ps %xmm0, %xmm0
vandps .LCPI0_0(%rip), %xmm0, %xmm0
vcvtps2ph $4, %xmm0, %xmm0
retq
# -- End function
.LCPI1_0:
.long 0x80000000 # float -0
.long 0x80000000 # float -0
.long 0x80000000 # float -0
.long 0x80000000 # float -0
fneg_f16(_Float16): # @fneg_f16(_Float16)
# %bb.0: # %entry
vcvtph2ps %xmm0, %xmm0
vxorps .LCPI1_0(%rip), %xmm0, %xmm0
vcvtps2ph $4, %xmm0, %xmm0
retq
# -- End function
```
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
