llvm-bolt –reg-reassin core dump when the input binary(mysqld) is PGO optimized

[llongint]

It's been described here before.

test.s.txt
test.fdata.txt

Run the following command to reproduce the problem:

mv test.s.txt test.s
gcc -O0 test.s -Wl,-q -o a.out
./a.out
llvm-bolt ./a.out -o ./a.opt -data=./test.fdata.txt --reg-reassign --print-regreassign -debug-only=regreassign
./a.opt

./a.opt will fall into an infinite loop.

[llongint]

@aaupov

[llvmbot]

@llvm/issue-subscribers-bolt

[aaupov]

@llongint - sorry about the delay. I will get back to this issue next week. In the meantime you can disable this optimization as layout optimizations generally provide much higher value.

[llongint]

Thanks. That's okay, but it looks like rafaelauler is already looking at this issue. See here

[aaupov]

@llongint – the problem is that BOLT doesn't perform cross-function reg-reassign. It would be better to disable function splitting in the compiler so that BOLT performs it using the accurate profile. That may in turn enable reg-reassign for this particular case.

[llongint]

That should work, but I haven't found an GCC option to turn off.

Also, I noticed that BOLT gives this hint, if this issue cannot be solved, maybe we can exit this optimization when this is detected?

BOLT-WARNING: debug info will be stripped from the binary. Use -update-debug-sections to keep it.
BOLT-WARNING: disabling lite mode (-lite) when split functions are present

[aaupov]

The flag to disable GCC function splitting is -fno-reorder-blocks-and-partition.
Split function support is by best effort, and reg-reassign is an experimental optimization. The two weren't extensively tested together.

[llongint]

Can we simply fix it like this D141931