| Issue |
242595
|
| Summary |
[AMDGPU] Debug uses change i1 lane-mask lowering
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
yqtian-se
|
Compiling this kernel with `-g` changes how lane masks are initialized and
allocated. In particular, one mask is initialized from `exec_lo` instead of
`-1`.
```opencl
__kernel void phi(__global int *out, __global int *in, int n) {
int i = 0;
int tid = get_local_id(0);
bool value = true;
bool previous;
do {
previous = value;
if (value)
value = in[i + tid] == 0;
i++;
} while (i != n);
bool last = previous;
out[tid] = value;
}
```
```sh
clang --target=amdgcn-amd-amdhsa -mcpu=gfx1010 -x cl -cl-std=CL2.0 \
-nogpulib -ffreestanding -O1 -g -S input.cl -o with-g.s
clang --target=amdgcn-amd-amdhsa -mcpu=gfx1010 -x cl -cl-std=CL2.0 \
-nogpulib -ffreestanding -O1 -g0 -S input.cl -o without-g.s
diff -u without-g.s with-g.s
```
The relevant difference is:
```diff
- s_mov_b32 s5, -1
+ s_mov_b32 s5, exec_lo
```
The first difference appears in `si-i1-copies`. `mergeIncomingLaneMasks`
collects blocks from `use_instructions(DstReg)`, including blocks containing
only debug uses. A `DBG_VALUE` outside the loop therefore changes the computed
post-dominator bound and selects different lane-mask lowering.
Debug-only uses should not change this decision. This was reproduced with
clang `7517bcdb0b87`; the responsible code remains unchanged on LLVM main
`2451a1eadfa1`.
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs