| Issue |
172097
|
| Summary |
[missed-opt] [selectiondag] Bad codegen on shift of a 128-bit number by an assumed-small amount
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
purplesyringa
|
[Godbolt](https://godbolt.org/z/arEhKn1P1)
This is minimized from real code, so it looks odd. The `assume` is intended to show that the shift cannot overflow, and the high half of `s` can be computed with something like `shld` on x86-64.
In reality, if `assume(k < 64)` is added, the generated code still checks for `k & 64` (`with_assume`). Using `k & 63` as the index generates good code and lets LLVM realize this is a rotate (`with_masking`). Interestingly, mixing `k & 63` and `assume` still produces bad code (`with_assume_and_masking`).
```cpp
#include <stdint.h>
uint64_t with_assume(uint64_t x, uint64_t k) {
__builtin_assume(k < 64);
__uint128_t s = (__uint128_t)x << k;
return (uint64_t)s + (uint64_t)(s >> 64);
}
uint64_t with_assume_and_masking(uint64_t x, uint64_t k) {
__builtin_assume(k < 64);
__uint128_t s = (__uint128_t)x << (k & 63);
return (uint64_t)s + (uint64_t)(s >> 64);
}
uint64_t with_masking(uint64_t x, uint64_t k) {
__uint128_t s = (__uint128_t)x << (k & 63);
return (uint64_t)s + (uint64_t)(s >> 64);
}
```
```asm
with_assume(unsigned long, unsigned long):
mov rcx, rsi
xor edx, edx
shld rdx, rdi, cl
shl rdi, cl
xor eax, eax
test cl, 64
cmovne rdx, rdi
cmove rax, rdi
add rax, rdx
ret
with_assume_and_masking(unsigned long, unsigned long):
mov rcx, rsi
xor edx, edx
shld rdx, rdi, cl
shl rdi, cl
xor eax, eax
test cl, 64
cmovne rdx, rdi
cmove rax, rdi
add rax, rdx
ret
with_masking(unsigned long, unsigned long):
mov rcx, rsi
mov rax, rdi
rol rax, cl
ret
```
As far as I can tell, the reason for bad codegen in `with_assume_and_masking` is that `assume` causes `& 63` to be optimized out during constprop before isel, and then the `assume` is removed during codegenprepare and backend has no idea the shift amount is bounded. For `with_assume` alone, the removal of `assume` similarly means that backend is not aware of the bounds.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
