This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [PATCH] Fix PR59822
- From: Richard Biener <rguenther at suse dot de>
- To: Jakub Jelinek <jakub at redhat dot com>
- Cc: gcc-patches at gcc dot gnu dot org
- Date: Thu, 16 Jan 2014 09:40:24 +0100 (CET)
- Subject: Re: [PATCH] Fix PR59822
- Authentication-results: sourceware.org; auth=none
- References: <alpine dot LSU dot 2 dot 11 dot 1401151405270 dot 4623 at zhemvz dot fhfr dot qr> <20140115173414 dot GD892 at tucnak dot redhat dot com>
On Wed, 15 Jan 2014, Jakub Jelinek wrote:
> On Wed, Jan 15, 2014 at 02:06:24PM +0100, Richard Biener wrote:
> >
> > This should fix PR59822, us creating in valid SSA form when
> > hoisting a[i_2] out of a loop where i_2 is defined inside the loop.
> >
> > Bootstrap / regtest running on x86_64-unknown-linux-gnu.
>
> Unfortunately with my patch on top of that (with obvious adjustment
> that it applies against the changed vectorizable_load code)
> still fails on lots of testcases, e.g.
> /usr/src/gcc/obj877/gcc/xgcc -B/usr/src/gcc/obj877/gcc/ /usr/src/gcc/gcc/testsuite/gcc.dg/vect/pr31343.c -fno-diagnostics-show-caret -fdiagnostics-color=never -msse2 -ftree-vectorize -fno-vect-cost-model -fno-common -O2 -fdump-tree-vect-details -S -o pr31343.s
> /usr/src/gcc/gcc/testsuite/gcc.dg/vect/pr31343.c: In function 'main1':
> /usr/src/gcc/gcc/testsuite/gcc.dg/vect/pr31343.c:12:1: internal compiler error: in op_iter_init, at ssa-iterators.h:602
> 0x89aa2fe op_iter_init
> ../../gcc/ssa-iterators.h:600
> 0x81232a9 op_iter_init
> ../../gcc/ssa-iterators.h:600
> 0x89bd936 op_iter_init_tree
> ../../gcc/ssa-iterators.h:662
> 0x89bd936 hoist_defs_of_uses
> ../../gcc/tree-vect-stmts.c:5507
> 0x89bd936 vectorizable_load
> ../../gcc/tree-vect-stmts.c:6442
> I think that has to do with nested loop vectorization and
> inv_p being true in the inner loop for:
> # j_54 = PHI <j_45(5), 0(4)>
> ...
> _40 = ub[j_54];
> and thus def_stmt is GIMPLE_PHI and
> FOR_EACH_SSA_TREE_OPERAND isn't allowed on those.
> Perhaps we want to avoid doing the hoisting for nested loop vectorization?
> And perhaps hoist_defs_of_uses should return false if
> gimple_code (def_stmt) == GIMPLE_PHI?
That should be enough I think (we can't hoist PHIs anyway).
Richard.