On Mon, 20 May 2019, Christophe Lyon wrote:
> On Mon, 20 May 2019 at 09:16, Richard Biener <rguent...@suse.de> wrote:
> >
> > On Mon, 20 May 2019, Christophe Lyon wrote:
> >
> > > On Fri, 17 May 2019 at 11:24, Richard Biener <rguent...@suse.de> wrote:
> > > >
> > > >
> > > > Bootstrap / testing in progress on x86_64-unknown-linux-gnu.
> > > >
> > > > Richard.
> > > >
> > > > 2019-05-17 Richard Biener <rguent...@suse.de>
> > > >
> > > > c/
> > > > * gimple-parser.c (c_parser_gimple_statement): Handle
> > > > __VEC_PERM.
> > > > (c_parser_gimple_unary_expression): Likewise.
> > > > (c_parser_gimple_parentized_ternary_expression): New function.
> > > >
> > > > * gimple-pretty-print.c (dump_ternary_rhs): Handle dumping
> > > > VEC_PERM_EXPR as __VEC_PERM with -gimple.
> > > >
> > > > * gcc.dg/gimplefe-41.c: New testcase.
> > > >
> > >
> > > Hi,
> > >
> > > This causes an ICE on arm:
> > > /gcc/testsuite/gcc.dg/gimplefe-41.c: In function '_mm_add_sd':
> > > /gcc/testsuite/gcc.dg/gimplefe-41.c:39:1: error: vectors with
> > > different element number found in vector permute expression
> > > __v2df
> > > __v2df
> > > vector(2) double
> > > __v2di
> > > z_6 = VEC_PERM_EXPR <x_4(D), _7, { 2, 1, 0, 0 }>;
> > > during IPA pass: *free_lang_data
> > > /gcc/testsuite/gcc.dg/gimplefe-41.c:39:1: internal compiler error:
> > > verify_gimple failed
> > > 0xc47110 verify_gimple_in_cfg(function*, bool)
> > > /gcc/tree-cfg.c:5396
> > > 0xb0b673 execute_function_todo
> > > /gcc/passes.c:1963
> > > 0xb0bf3d do_per_function
> > > /gcc/passes.c:1645
> > > 0xb0c009 execute_todo
> > > /gcc/passes.c:2017
> >
> > Hope this fixes it and the other issue.
> >
> > Tested on x86_64-unknown-linux-gnu, applied.
> >
> > Richard.
> >
> > 2019-05-20 Richard Biener <rguent...@suse.de>
> >
> > PR testsuite/90518
> > * gcc.dg/gimplefe-40.c: Restrict to targets with appropriate
> > vector support.
> > * gcc.dg/gimplefe-41.c: Likewise.
> >
> Yes this fixes the problem on arm. Well, it avoids the ICE.
It was invalid GIMPLE IL for arm which is obviously "easy" to
generate using the GIMPLE FE ... Thanks for testing.
Richard.
