| Issue |
55468
|
| Summary |
Missed optimization: Bitwise shifts don't backpropagate knowledge about their operands
|
| Labels |
|
| Assignees |
|
| Reporter |
josephcsible
|
Consider this C code:
```c
extern int x;
int f(int y) {
x = y % 32;
return 1 << y;
}
```
At `-O3`, it compiles to this assembly:
```asm
f: # @f
movl %edi, %ecx
leal 31(%rcx), %eax
testl %edi, %edi
cmovnsl %edi, %eax
andl $-32, %eax
movl %edi, %edx
subl %eax, %edx
movl %edx, x(%rip)
movl $1, %eax
shll %cl, %eax
retq
```
But this would have been more efficient:
```asm
f: # @f
movl %edi, %ecx
movl %edi, x(%rip)
movl $1, %eax
shll %cl, %eax
retq
```
If I manually put `__builtin_assume(y >= 0 && y < 32);` right before `return 1 << y;`, then I do get that more efficient assembly. But we should automatically be assuming that, since `y` gets used as the right operand of a bitwise shift operator.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
