| Issue |
177541
|
| Summary |
Miscompilation in float select transformation
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
ariskeirio
|
I found a miscompilation in an LLVM IR transformation that rewrites chained `fcmp` + `select` operations. The target IR produces **poison values** in cases where the source code produces well-defined results.
**Example:**
Source IR:
```SQL
define float @src(float %x) {
%a = fcmp nnan ninf ult float %x, 255.0
%b = select i1 %a, float 255.0, float %x
%c = fcmp nnan ugt float %x, 512.0
%r = select i1 %c, float 512.0, float %b
ret float %r
}
```
Target IR(optimized from source IR):
https://godbolt.org/z/4YsbM9Y65
```SQL
define float @tgt(float %x) {
%a.inv = fcmp nnan ninf oge float %x, 2.550000e+02
%b = select nnan ninf i1 %a.inv, float %x, float 2.550000e+02
%.inv = fcmp nnan ole float %b, 5.120000e+02
%r1 = select nnan i1 %.inv, float %b, float 5.120000e+02
ret float %r1
}
```
Here is the alive2-tv verification
https://alive2.llvm.org/ce/z/CceAMX
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
