Re: Fix twolf -funroll-loops -O3 miscompilation (a semi-latent web.c bug)

[Paolo Bonzini <bonzini@gnu.org>]

Il 15/10/2012 10:37, Steven Bosscher ha scritto:
>> I prefer to declare the notes invalid and drop the notes.
> Then, afaic, our only option is to drop them all in web, as per attached patch.
>
> I strongly disagree with this approach though. It destroys information
> that is correct, that we had before DF_RD_PRUNE_DEAD_DEFS, that we can
> update, and that helps with optimization. With renaming these notes
> are valid, and do not refer to dead regs

I agree it is bad.  But I do not understand the last sentence: I
suppose you mean that _without_ renaming these notes are valid, on the
other hand it is normal that some of the notes will be dropped if you
shorten live ranges.

Without removing all of the notes you can do something like this:

- drop the deferred rescanning from web.c.  Instead, make replace_ref
return a bool and call df_insn_rescan manually from web_main.

- attribute new registers to webs in a separate pass that happens
before rewriting, and compute a special version of LR_IN/LR_OUT that
uses the rewritten registers.

- process instructions in reverse order; before starting the visit of
a basic block, initialize the local LR bitmap with the rewritten
LR_OUT of the previous step

- after rewriting and scanning each statement, simulate liveness using
the new defs and uses.

- after rewriting each statement, look for EQ_USES referring to
registers that are dead just before the statement, and delete
REG_EQUAL notes if this is the case

Paolo

> This whole discussion about notes being dead has gone in completely
> the wrong direction. With renaming these notes are valid, and do not
> refer to dead regs. Perhaps you could be convinced if you look at
> Honza's test case with the patch of r192413 reverted.
> The test case still fails with --param max-unroll-times=3, that makes
> visualizing the problem easier.