| Issue |
108904
|
| Summary |
ppc64el miscompilation on signed integer argument passing
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
tpearson-ssc
|
On all Clang versions after Clang 16, llc miscompiles a simple C++ test program testing integer argument passing:
```
#include <stdio.h>
#include <math.h>
int main() {
volatile double result;
volatile int exp = -53;
volatile double base = 3635833917682146;
result = ldexp(base, exp);
printf("Result: %f\n", result);
return 0;
}
```
The program should output `0.403659`, but on Clang 17 and newer outputs `inf`.
On calling ldexp, `r4` is used to pass `exp`. Unfortunately, only the lower 32 bits of `exp` are passed in `r4`, leading to an incorrect result. This is due to incorrect backend code generation:
```
8bc: 60 00 3f c8 lfd f1,96(r31)
- 8c0: 6e 00 9f e8 lwa r4,108(r31) <-- CORRECT (sign extends values for signed 64-bit integer parameter passing in r4)
+ 8c0: 6c 00 9f 80 lwz r4,108(r31) <-- WRONG (zeroes upper 32 bits of parameter in r4)
8c4: 9d fd ff 4b bl 660 <0000001b.plt_call.ldexp@@GLIBC_2.17>
8c8: 18 00 41 e8 ld r2,24(r1)
8cc: 70 00 3f d8 stfd f1,112(r31)
```
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
