| Issue |
87758
|
| Summary |
Clang differences when using commandline options or pragmas to turn on fast-math functionallity
|
| Labels |
|
| Assignees |
|
| Reporter |
karka228
|
When looking at the output from clang -cc1 for the program below
```
float powf(float x, float y);
float sqrtf(float x);
float f2(float x, int y) {
x = powf(x,y);
x = sqrtf(x);
return x;
}
```
Compiled with:
`$ clang -cc1 foo.c -triple x86_64-linux-gnu -emit-llvm -o - -ffast-math -ffp-contract=fast`
The result of the `-ffast-math` and ` -ffp-contract=fast` options is that both calls to pow and sqrt are annotated with fast math flag `fast` (see LangRef):
```
%2 = call fast float @llvm.pow.f32(float %0, float %conv)
%4 = call fast float @llvm.sqrt.f32(float %3)
```
When instead turning on fast math with `#pragma float_control(precise,off)` like this:
```
float powf(float x, float y);
float sqrtf(float x);
#pragma float_control(push)
#pragma float_control(precise,off)
float f2(float x, int y) {
x = powf(x,y);
x = sqrtf(x);
return x;
}
#pragma float_control(pop)
```
Compiled with:
`$ clang -cc1 foo.c -triple x86_64-linux-gnu -emit-llvm -o -`
The result of the pragma is that only the sqrt call get to be annotated with the fast math flag `fast`:
```
%2 = call float @llvm.pow.f32(float %0, float %conv)
%4 = call fast float @llvm.sqrt.f32(float %3)
```
Why are not `llvm.pow.f32` not annotated with fast math flag `fast` when using `#pragma float_control(precise,off)`?
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
