| Issue |
58170
|
| Summary |
Loop Interchange bug due to loop guards not being interchanged
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
alexgatea
|
Right now loop guards are not interchanged which leads to incorrect control flow after the interchange.
For example, consider the following code before the interchange.
```
if(n > 0)
for(int i=0; i != n; i++)
if(m > 0)
for(int j=0; j != m; j++)
A[j][i] += B[j][i] + 1;
```
IR: [before.ll](https://github.com/llvm/llvm-project/files/9717558/before.ll.txt)
After interchange:
```
if(n > 0)
for(int j=0; j != m; j++)
for(int i=0; i != n; i++) {
if(m > 0)
A[j][i] += B[j][i] + 1;
else break;
}
```
IR: [after.ll.txt](https://github.com/llvm/llvm-project/files/9717613/after.ll.txt)
If m is -1 at run-time then the original code only executes the outer for-i loop whereas the resulting code is an infinite loop (since `j != m` always evaluates to false).
Possible solutions under consideration:
- See if running simple loop unswitch or LICM prior to loop interchange fixes the issue. I tried this and it doesn't seem to help; loop unswitch does nothing and LICM hoists the computation of the condition m > 0 out of the for-i loop but the branch remains in place (inside the for-i loop). Still, this might be able to be tweaked to work, I'm not sure.
- View non-user generated loop guards (checked via `isGuarded()` for example) as part of anatomy of a loop.
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
