[prev in list] [next in list] [prev in thread] [next in thread] 

List:       gcc
Subject:    Re: [tree-ssa] Kills by VDEFs in mark_def_sites
From:       law () redhat ! com
Date:       2004-04-30 21:31:33
Message-ID: 200404302131.i3ULVXXo016140 () speedy ! slc ! redhat ! com
[Download RAW message or body]

In message <20040414213604.GA32480@atrey.karlin.mff.cuni.cz>, Zdenek Dvorak wri
tes:
 >Hello,
 >
 >a comment in tree-into-ssa.c:mark_def_sites states that
 >
 >/* Note that virtual definitions are irrelevant for computing KILLS
 >   because a VDEF does not constitute a killing definition of the
 >   variable.  */
 >
 >and indeed, we do not set a bit in "kills" bitmap for VDEFs.
 >
 >There are two things that make this part of code a bit weird:
 >
 >1) For operand of VDEF and for VUSE we test whether the bit is set
 >   in the "kills" bitmap; however this is clearly useless, as there
 >   is no way how there could be anything set in this bitmap for virtual
 >   operands.
 >
 >2) The comment above is moreorless useless as well.  Even if we set the
 >   bit in the "kills" bitmap, everything would work exactly the same
 >   way, since VDEF has an operand that is processed before the VDEF's
 >   result, so the basic block processed would be marked in the livein
 >   bitmap as well.
 >
 >Are these observations correct, or am I missing something?
I believe your observations are correct.  I believe this code only made
sense in the world where we still mixed real and virtual operands in
strange and not so wonderful ways.

I'll give things a spin with the kills bitmap gunk removed for virtuals.

jeff


[prev in list] [next in list] [prev in thread] [next in thread]