| Message ID | 20100902214518.GX1269@tyan-ft48-01.lab.bos.redhat.com |
|---|---|
| State | New |
| Headers | show |
On Thu, Sep 2, 2010 at 11:45 PM, Jakub Jelinek <jakub@redhat.com> wrote: > Hi! > > TYPE_MODE of a generic vector isn't necessarily BLKmode, it can be > some integer mode too (vector_type_mode > /* For integers, try mapping it to a same-sized scalar mode. */ > hunk), but expand_expr handles VECTOR_CST with non-VECTOR_MODE_P > mode always the same, but forcing the initializer into memory, which > rtl_for_decl_init obviously doesn't want and asserts it doesn't happen. > > Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for > trunk/4.5? Ok. Thanks, Richard. > 2010-09-02 Jakub Jelinek <jakub@redhat.com> > > PR debug/45500 > * dwarf2out.c (rtl_for_decl_init): Give up for all generic vectors, > not just generic vectors with BLKmode. > > * gcc.target/i386/pr45500.c: New test. > > --- gcc/dwarf2out.c.jj 2010-08-31 15:39:16.000000000 +0200 > +++ gcc/dwarf2out.c 2010-09-02 19:20:15.213521284 +0200 > @@ -16256,7 +16256,8 @@ rtl_for_decl_init (tree init, tree type) > ; > /* Vectors only work if their mode is supported by the target. > FIXME: generic vectors ought to work too. */ > - else if (TREE_CODE (type) == VECTOR_TYPE && TYPE_MODE (type) == BLKmode) > + else if (TREE_CODE (type) == VECTOR_TYPE > + && !VECTOR_MODE_P (TYPE_MODE (type))) > ; > /* If the initializer is something that we know will expand into an > immediate RTL constant, expand it now. We must be careful not to > --- gcc/testsuite/gcc.target/i386/pr45500.c.jj 2010-09-02 19:14:47.746364347 +0200 > +++ gcc/testsuite/gcc.target/i386/pr45500.c 2010-09-02 19:14:06.000000000 +0200 > @@ -0,0 +1,6 @@ > +/* PR debug/45500 */ > +/* { dg-do compile } */ > +/* { dg-options "-g -msse" } */ > + > +typedef char V __attribute__ ((__vector_size__ (16))); > +static const V s = { '\n', '\r', '?', '\\' }; > > Jakub >
--- gcc/dwarf2out.c.jj 2010-08-31 15:39:16.000000000 +0200 +++ gcc/dwarf2out.c 2010-09-02 19:20:15.213521284 +0200 @@ -16256,7 +16256,8 @@ rtl_for_decl_init (tree init, tree type) ; /* Vectors only work if their mode is supported by the target. FIXME: generic vectors ought to work too. */ - else if (TREE_CODE (type) == VECTOR_TYPE && TYPE_MODE (type) == BLKmode) + else if (TREE_CODE (type) == VECTOR_TYPE + && !VECTOR_MODE_P (TYPE_MODE (type))) ; /* If the initializer is something that we know will expand into an immediate RTL constant, expand it now. We must be careful not to --- gcc/testsuite/gcc.target/i386/pr45500.c.jj 2010-09-02 19:14:47.746364347 +0200 +++ gcc/testsuite/gcc.target/i386/pr45500.c 2010-09-02 19:14:06.000000000 +0200 @@ -0,0 +1,6 @@ +/* PR debug/45500 */ +/* { dg-do compile } */ +/* { dg-options "-g -msse" } */ + +typedef char V __attribute__ ((__vector_size__ (16))); +static const V s = { '\n', '\r', '?', '\\' };
Hi! TYPE_MODE of a generic vector isn't necessarily BLKmode, it can be some integer mode too (vector_type_mode /* For integers, try mapping it to a same-sized scalar mode. */ hunk), but expand_expr handles VECTOR_CST with non-VECTOR_MODE_P mode always the same, but forcing the initializer into memory, which rtl_for_decl_init obviously doesn't want and asserts it doesn't happen. Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk/4.5? 2010-09-02 Jakub Jelinek <jakub@redhat.com> PR debug/45500 * dwarf2out.c (rtl_for_decl_init): Give up for all generic vectors, not just generic vectors with BLKmode. * gcc.target/i386/pr45500.c: New test. Jakub