| Message ID | 20110316233202.GQ30899@tyan-ft48-01.lab.bos.redhat.com |
|---|---|
| State | New |
| Headers | show |
On Thu, Mar 17, 2011 at 12:32 AM, Jakub Jelinek <jakub@redhat.com> wrote: > Hi! > > Currently expand_debug_expr doesn't always guarantee canonical > MEM_EXPRs, if MEM_EXPR isn't canonical, sometimes aliasing code ICEs on it > (since recent Richard's change fortunately only if --enable-checking=yes). > > The following patch canonicalizes those using fold and if canonicalization > isn't successful, just clears the MEM_EXPR. > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk and > 4.6.1? > > 2011-03-17 Jakub Jelinek <jakub@redhat.com> > > PR debug/48134 > * cfgexpand.c (expand_debug_expr) <component_ref>: Use fold to > canonicalize MEM_EXPR, if it still isn't canonical afterwards, just > clear MEM_EXPR. > > * gcc.dg/pr48134.c: New test. > > --- gcc/cfgexpand.c.jj 2011-03-16 18:30:12.000000000 +0100 > +++ gcc/cfgexpand.c 2011-03-16 22:05:13.000000000 +0100 > @@ -2712,6 +2712,8 @@ expand_debug_expr (tree exp) > > if (MEM_P (op0)) > { > + tree mem_exp; > + > if (mode1 == VOIDmode) > /* Bitfield. */ > mode1 = smallest_mode_for_size (bitsize, MODE_INT); > @@ -2736,6 +2738,12 @@ expand_debug_expr (tree exp) > if (op0 == orig_op0) > op0 = shallow_copy_rtx (op0); > set_mem_attributes (op0, exp, 0); > + mem_exp = fold (exp); > + if (!SSA_VAR_P (mem_exp) > + && TREE_CODE (mem_exp) != MEM_REF > + && TREE_CODE (mem_exp) != TARGET_MEM_REF) > + mem_exp = NULL_TREE; > + set_mem_expr (op0, mem_exp); That will, for the testcase drop the MEM[&t.s].w completely, right? I think you eventually want to use maybe_fold_reference which does canonicalize mem-refs after propagation (but also watch for bogus debug information due to constant folding of initializers I ran into because of setting DECL_INITIAL to error_mark_node for unused globals - I'll try to dig up the patch I had to fix that). Richard. > } > > if (bitpos == 0 && mode == GET_MODE (op0)) > --- gcc/testsuite/gcc.dg/pr48134.c.jj 2011-03-16 22:04:35.000000000 +0100 > +++ gcc/testsuite/gcc.dg/pr48134.c 2011-03-16 22:02:13.000000000 +0100 > @@ -0,0 +1,25 @@ > +/* PR debug/48134 */ > +/* { dg-do compile } */ > +/* { dg-options "-O2 -fstack-check=specific -fno-tree-dse -fno-tree-fre -fno-tree-loop-optimize -g" } */ > + > +struct S { int w, z; }; > +struct T { struct S s; }; > +int i; > + > +static inline struct S > +bar (struct S x) > +{ > + i++; > + return x; > +} > + > +int > +foo (struct T t, struct S s) > +{ > + struct S *c = &s; > + if (i) > + c = &t.s; > + t.s.w = 3; > + s = bar (*c); > + return t.s.w; > +} > > Jakub >
On Thu, Mar 17, 2011 at 9:41 AM, Richard Guenther <richard.guenther@gmail.com> wrote: > On Thu, Mar 17, 2011 at 12:32 AM, Jakub Jelinek <jakub@redhat.com> wrote: >> Hi! >> >> Currently expand_debug_expr doesn't always guarantee canonical >> MEM_EXPRs, if MEM_EXPR isn't canonical, sometimes aliasing code ICEs on it >> (since recent Richard's change fortunately only if --enable-checking=yes). >> >> The following patch canonicalizes those using fold and if canonicalization >> isn't successful, just clears the MEM_EXPR. >> >> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk and >> 4.6.1? >> >> 2011-03-17 Jakub Jelinek <jakub@redhat.com> >> >> PR debug/48134 >> * cfgexpand.c (expand_debug_expr) <component_ref>: Use fold to >> canonicalize MEM_EXPR, if it still isn't canonical afterwards, just >> clear MEM_EXPR. >> >> * gcc.dg/pr48134.c: New test. >> >> --- gcc/cfgexpand.c.jj 2011-03-16 18:30:12.000000000 +0100 >> +++ gcc/cfgexpand.c 2011-03-16 22:05:13.000000000 +0100 >> @@ -2712,6 +2712,8 @@ expand_debug_expr (tree exp) >> >> if (MEM_P (op0)) >> { >> + tree mem_exp; >> + >> if (mode1 == VOIDmode) >> /* Bitfield. */ >> mode1 = smallest_mode_for_size (bitsize, MODE_INT); >> @@ -2736,6 +2738,12 @@ expand_debug_expr (tree exp) >> if (op0 == orig_op0) >> op0 = shallow_copy_rtx (op0); >> set_mem_attributes (op0, exp, 0); >> + mem_exp = fold (exp); >> + if (!SSA_VAR_P (mem_exp) >> + && TREE_CODE (mem_exp) != MEM_REF >> + && TREE_CODE (mem_exp) != TARGET_MEM_REF) >> + mem_exp = NULL_TREE; >> + set_mem_expr (op0, mem_exp); > > That will, for the testcase drop the MEM[&t.s].w completely, right? > I think you eventually want to use maybe_fold_reference which does > canonicalize mem-refs after propagation (but also watch for bogus > debug information due to constant folding of initializers I > ran into because of setting DECL_INITIAL to error_mark_node for > unused globals - I'll try to dig up the patch I had to fix that). I am testing the following. Richard.
--- gcc/cfgexpand.c.jj 2011-03-16 18:30:12.000000000 +0100 +++ gcc/cfgexpand.c 2011-03-16 22:05:13.000000000 +0100 @@ -2712,6 +2712,8 @@ expand_debug_expr (tree exp) if (MEM_P (op0)) { + tree mem_exp; + if (mode1 == VOIDmode) /* Bitfield. */ mode1 = smallest_mode_for_size (bitsize, MODE_INT); @@ -2736,6 +2738,12 @@ expand_debug_expr (tree exp) if (op0 == orig_op0) op0 = shallow_copy_rtx (op0); set_mem_attributes (op0, exp, 0); + mem_exp = fold (exp); + if (!SSA_VAR_P (mem_exp) + && TREE_CODE (mem_exp) != MEM_REF + && TREE_CODE (mem_exp) != TARGET_MEM_REF) + mem_exp = NULL_TREE; + set_mem_expr (op0, mem_exp); } if (bitpos == 0 && mode == GET_MODE (op0)) --- gcc/testsuite/gcc.dg/pr48134.c.jj 2011-03-16 22:04:35.000000000 +0100 +++ gcc/testsuite/gcc.dg/pr48134.c 2011-03-16 22:02:13.000000000 +0100 @@ -0,0 +1,25 @@ +/* PR debug/48134 */ +/* { dg-do compile } */ +/* { dg-options "-O2 -fstack-check=specific -fno-tree-dse -fno-tree-fre -fno-tree-loop-optimize -g" } */ + +struct S { int w, z; }; +struct T { struct S s; }; +int i; + +static inline struct S +bar (struct S x) +{ + i++; + return x; +} + +int +foo (struct T t, struct S s) +{ + struct S *c = &s; + if (i) + c = &t.s; + t.s.w = 3; + s = bar (*c); + return t.s.w; +}