Fix for df problems, and and an ifcvt improvement

[Bernd Schmidt <bernds at codesourcery dot com>]

I was working on a patch that tries to make use of
df_simulate_one_insn_forwards, and I noticed that things weren't working
quite as I was hoping.  Some debugging showed that after calling
df_simulate_one_insn_forwards on a CALL insn, all the call clobbered
registers were incorrectly marked as live.

The function calls df_simulate_find_defs to mark all set registers as
live, then clears anything for which it can find a REG_DEAD/REG_UNUSED
note.  df_simulate_find_defs is broken for this purpose: it should
ignore clobbers, but it ignores partial and conditional sets.  Confusion
about how to handle the latter two is a common problem with the df code;
in this case at least it's clearly wrong to ignore them.

While working on this, I discovered by accident that a function in
ifcvt.c which uses this infrastructure is more conservative than it
needs to be.  Consider case 2 from one of the comments:

   (2)
        if (test) goto E; // x not live    [TEST]
        x = big();
        goto L;
        E:
        x = b;        [ MERGE_BB ]
        goto M;

   becomes

        x = b;
        if (test) goto M;
        x = big();
        goto L;

Currently, it fails if (MERGE_SET & TEST_SET) != 0, but it really only
needs to make sure that real sets (i.e. not clobbers) in MERGE_BB aren't
overwritten in TEST.  I've added a new function
df_simulate_find_noclobber_defs, which is now used for this, and for
df_simulate_one_insn_forwards.

Bootstrapped and regression tested on i686-linux.  Tests on arm aren't
complete but look good.  Ok (maybe also 4.5)?


Bernd

Attachment: df-problems.diff
Description: Text document