GCC 11 and 12 have pathological compile time on test_fuzz_expression_code_gen output

[inducer]

I was debugging this CI failure, and what I discovered is that gcc 11 and 12 (as used on the beers) have pathologically long compile times on the output of, say,

python test_expression.py 'test_fuzz_expression_code_gen(cl._csc, "complex", 4, lp.ExecutableCTarget)'

after the merge of inducer/pymbolic#110.

Sample generated code is in this gist: https://gist.github.com/inducer/802e227e86e3ff4fd95aa0fefae462fe

gcc -c post-110.c

won't finish in reasonable time for me, but pre-110.c will. Also, gcc 10 is nearly instantaneous on both.

I think the main thing that's affecting this inducer/pymbolic@12569d5.

Nb: Github actions is unaffected, as (I think) the gcc there is older, possibly 10.

[inducer]

The easy way out of this would be to just go back to flattening, but I'd kind of like to move away from that... floating point math is not associative after all.

[kaushikcfd]

Can reproduce this. Even happens with -O0 flags. perf top shows walk_tree_1 routine in GCC is where it takes forever. To be honest, the expression isn't that big at all, probably some regression in gcc-11+.

[inducer]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107127

[inducer]

cc @wence- (You guys might get bitten by this.)

[inducer]

#687 is probably as far as that will go.

[inducer]

I'll go ahead and close this. It's arguably not our bug.

[inducer]

https://gitlab.tiker.net/inducer/loopy/-/jobs/467837 #687 has CI run times back down to normal levels.

[wence-]

cc @wence- (You guys might get bitten by this.)

Thanks, cc @dham and @connorjward too