This is a nasty bug that points to some memory corruption issue or something. time ~/llvm/Release/bin/opt -predsimplify -disable-output bugpoint-reduced-simplified.bc 0.055u 0.009s 0:00.06 83.3% 0+0k 0+1io 0pf+0w time ~/llvm/Debug/bin/opt -predsimplify -disable-output bugpoint-reduced-simplified.bc <forever> Note that the .bc *does* have a large nasty cfg, but it is strange that this issue should only affect a debug opt and not a release opt. This points to some memory clobbering, or deleting something, but not removing it from a map or something (i.e. it's address is still the key or value of some map). I'm disabling predsimplify in llvm-gcc until this is resolved, this breaks debug bootstraps on both ppc and x86. -Chris
Created attachment 422 [details] bc file
Could it be a wonky assert? Did you try Release+Asserts build? Reid.
entirely possible. No I didn't try.
I tried the testcase under valgrind and it didn't show any memory errors. That said, I haven't been able to reproduce this yet on x86/linux.
Given that predsimplify is being rewritten, the problem doesn't reproduce on Linux and this hasn't shown up recently, can we just close this WONTFIX ?
Perhaps when predsimplify is rewritten. Until then, we can't verify it. -Chris
Created attachment 556 [details] updated to current bytecode format I was never able to reproduce the original bug, but this does seem to expose poor performance, possibly due to debugging code. I'm seeing large numbers of nodes with a lot of interconnections, probably constants. This isn't surprising, as the constant code is full of fixme's at the moment.
This is done.