| Issue |
129947
|
| Summary |
Missed optimization: `umax(1 << x, 1 << y) => 1 << umax(x, y)`
|
| Labels |
llvm:optimizations,
llvm:instcombine,
missed-optimization
|
| Assignees |
|
| Reporter |
Kmeakin
|
# C code
https://godbolt.org/z/YnEj5KGrW
```c
#include <stdint.h>
uint32_t src1(uint32_t x, uint32_t y) {
uint32_t shl_x = 1 << x;
uint32_t shl_y = 1 << y;
return (shl_x > shl_y) ? shl_x : shl_y;
}
uint32_t tgt1(uint32_t x, uint32_t y) {
uint32_t shl_x = 1 << x;
uint32_t shl_y = 1 << y;
return (x > y) ? shl_x : shl_y;
}
uint32_t src2(uint32_t x, uint32_t y) {
uint32_t shl_x = 1 << x;
uint32_t shl_y = 1 << y;
return (shl_x < shl_y) ? shl_x : shl_y;
}
uint32_t tgt2(uint32_t x, uint32_t y) {
uint32_t shl_x = 1 << x;
uint32_t shl_y = 1 << y;
return (x < y) ? shl_x : shl_y;
}
```
# Alive proof
https://alive2.llvm.org/ce/z/bv698b
```llvm
----------------------------------------
define i32 @src1(i32 noundef %#0, i32 noundef %#1) nofree willreturn memory(none) {
#2:
%#3 = shl nuw i32 1, noundef %#0
%#4 = shl nuw i32 1, noundef %#1
%#5 = umax i32 %#3, %#4
%#range_0_%#5 = !range i32 %#5, i32 1, i32 2147483649
ret i32 %#range_0_%#5
}
=>
define i32 @tgt1(i32 noundef %#0, i32 noundef %#1) nofree willreturn memory(none) {
#2:
%#3 = umax i32 noundef %#0, noundef %#1
%#4 = shl nuw i32 1, %#3
%#range_0_%#4 = !range i32 %#4, i32 1, i32 2147483649
ret i32 %#range_0_%#4
}
Transformation seems to be correct!
----------------------------------------
define i32 @src2(i32 noundef %#0, i32 noundef %#1) nofree willreturn memory(none) {
#2:
%#3 = shl nuw i32 1, noundef %#0
%#4 = shl nuw i32 1, noundef %#1
%#5 = umin i32 %#3, %#4
%#range_0_%#5 = !range i32 %#5, i32 1, i32 2147483649
ret i32 %#range_0_%#5
}
=>
define i32 @tgt2(i32 noundef %#0, i32 noundef %#1) nofree willreturn memory(none) {
#2:
%#3 = umin i32 noundef %#0, noundef %#1
%#4 = shl nuw i32 1, %#3
%#range_0_%#4 = !range i32 %#4, i32 1, i32 2147483649
ret i32 %#range_0_%#4
}
Transformation seems to be correct!
Summary:
2 correct transformations
0 incorrect transformations
0 failed-to-prove transformations
0 Alive2 errors
```
# AArch64 assembly
```asm
src1:
mov w8, #1
lsl w9, w8, w0
lsl w8, w8, w1
cmp w9, w8
csel w0, w9, w8, hi
ret
tgt1:
cmp w0, w1
mov w8, #1
csel w9, w0, w1, hi
lsl w0, w8, w9
ret
src2:
mov w8, #1
lsl w9, w8, w0
lsl w8, w8, w1
cmp w9, w8
csel w0, w9, w8, lo
ret
tgt2:
cmp w0, w1
mov w8, #1
csel w9, w0, w1, lo
lsl w0, w8, w9
ret
```
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
