| Issue |
128134
|
| Summary |
Missed optimization: AggressiveInstCombine misses combining loads from consecutive addresses
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
Cancelll
|
Example: https://godbolt.org/z/14cEnPoco
Alive2: https://alive2.llvm.org/ce/z/nQ5CMj
```llvm
define i32 @src(ptr %0) {
%2 = load i16, ptr %0, align 2
%3 = getelementptr inbounds nuw i8, ptr %0, i64 2
%4 = load i16, ptr %3, align 1
%5 = zext i16 %4 to i32
%6 = shl nuw i32 %5, 16
%7 = zext i16 %2 to i32
%8 = or disjoint i32 %6, %7
ret i32 %8
}
define i32 @tgt(ptr %0) {
%2 = load i32, ptr %0, align 2
ret i32 %2
}
```
AgressiveInstCombine should be able to do this optimization.
https://godbolt.org/z/hY4Gfcd66
I found this pattern in function DES_xcbc_encrypt in openssl/crypto/des/xcbc_enc.c
![]()
_______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
