On 11/2/18 7:40 AM, Richard Biener wrote: > > The following fixes PR87852, a latent bug in fwprop which when verifying > whether it may propagate a use from its definition site has a shortcut > > /* Check if the reg in USE has only one definition. We already > know that this definition reaches use, or we wouldn't be here. > However, this is invalid for hard registers because if they are > live at the beginning of the function it does not mean that we > have an uninitialized access. */ > regno = DF_REF_REGNO (use); > def = DF_REG_DEF_CHAIN (regno); > if (def > && DF_REF_NEXT_REG (def) == NULL > && regno >= FIRST_PSEUDO_REGISTER) > return false; > > not considering the case of a loop where the def might not dominate > the use. In fact earlier code in the very same function does > handle this case but only for the case where we'd try propagating > a later def into an earlier use. > > Bootstrapped on x86_64-unknown-linux-gnu, testing in progress. > > OK for trunk? > > Thanks, > Richard. > > 2018-11-02 Richard Biener <rguent...@suse.de> > > PR rtl-optimization/87852 > * fwprop.c (use_killed_between): Only consider single-defs of the > use in the definition stmt that dominate it. You may have just saved me major headaches. I'm in the middle of trying to debug an unreported ARM codegen bug. I just walked through the CSE dump and everything looks OK, and it's mucked up in fwprop1. The thing that caught my eye was a pseudo where the def does not dominate a use in a loop. ie, on the first iteration of the loop the value is undefined (but we'll never read it at runtime due to other checks)....
/me hopes this is the same thing and ultimately explains the mis-compilation of python I'm seeing. Jeff
