| Message ID | 20120423230312.GO16117@tyan-ft48-01.lab.bos.redhat.com |
|---|---|
| State | New |
| Headers | show |
On 04/23/12 16:03, Jakub Jelinek wrote: > PR middle-end/53084 > * varasm.c (compute_reloc_for_constant): Handle ADDR_EXPR > of MEM_REF. > (output_addressed_constants): Likewise. > > * gcc.c-torture/execute/pr53084.c: New test. Ok. r~
On Tue, Apr 24, 2012 at 1:03 AM, Jakub Jelinek <jakub@redhat.com> wrote: > Hi! > > output_addressed_constants isn't able to output constants addressed > with &MEM_REF<&something, X> and similarly compute_reloc_for_constant > doesn't handle it. Fixed thusly, bootstrapped/regtested on x86_64-linux and > i686-linux, ok for trunk/4.7? Isn't this latent in 4.6? Thanks, Richard. > 2012-04-23 Jakub Jelinek <jakub@redhat.com> > > PR middle-end/53084 > * varasm.c (compute_reloc_for_constant): Handle ADDR_EXPR > of MEM_REF. > (output_addressed_constants): Likewise. > > * gcc.c-torture/execute/pr53084.c: New test. > > --- gcc/varasm.c.jj 2012-04-23 18:55:46.000000000 +0200 > +++ gcc/varasm.c 2012-04-23 20:26:15.930660033 +0200 > @@ -3948,6 +3948,13 @@ compute_reloc_for_constant (tree exp) > tem = TREE_OPERAND (tem, 0)) > ; > > + if (TREE_CODE (tem) == MEM_REF > + && TREE_CODE (TREE_OPERAND (tem, 0)) == ADDR_EXPR) > + { > + reloc = compute_reloc_for_constant (TREE_OPERAND (tem, 0)); > + break; > + } > + > if (TREE_PUBLIC (tem)) > reloc |= 2; > else > @@ -4016,6 +4023,9 @@ output_addressed_constants (tree exp) > > if (CONSTANT_CLASS_P (tem) || TREE_CODE (tem) == CONSTRUCTOR) > output_constant_def (tem, 0); > + > + if (TREE_CODE (tem) == MEM_REF) > + output_addressed_constants (TREE_OPERAND (tem, 0)); > break; > > case PLUS_EXPR: > --- gcc/testsuite/gcc.c-torture/execute/pr53084.c 2012-04-16 20:35:50.603036390 +0200 > +++ gcc/testsuite/gcc.c-torture/execute/pr53084.c 2012-04-23 20:30:57.193896339 +0200 > @@ -0,0 +1,18 @@ > +/* PR middle-end/53084 */ > + > +extern void abort (void); > + > +__attribute__((noinline, noclone)) void > +bar (const char *p) > +{ > + if (p[0] != 'o' || p[1] != 'o' || p[2]) > + abort (); > +} > + > +int > +main () > +{ > + static const char *const foo[] = {"foo" + 1}; > + bar (foo[0]); > + return 0; > +} > > Jakub
On Tue, Apr 24, 2012 at 07:56:11AM +0200, Richard Guenther wrote: > On Tue, Apr 24, 2012 at 1:03 AM, Jakub Jelinek <jakub@redhat.com> wrote: > > Hi! > > > > output_addressed_constants isn't able to output constants addressed > > with &MEM_REF<&something, X> and similarly compute_reloc_for_constant > > doesn't handle it. Fixed thusly, bootstrapped/regtested on x86_64-linux and > > i686-linux, ok for trunk/4.7? > > Isn't this latent in 4.6? Possibly. The bug started being reproduceable on this testcase with http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=178312 and stopped being reproducible with http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=186687 Jakub
--- gcc/varasm.c.jj 2012-04-23 18:55:46.000000000 +0200 +++ gcc/varasm.c 2012-04-23 20:26:15.930660033 +0200 @@ -3948,6 +3948,13 @@ compute_reloc_for_constant (tree exp) tem = TREE_OPERAND (tem, 0)) ; + if (TREE_CODE (tem) == MEM_REF + && TREE_CODE (TREE_OPERAND (tem, 0)) == ADDR_EXPR) + { + reloc = compute_reloc_for_constant (TREE_OPERAND (tem, 0)); + break; + } + if (TREE_PUBLIC (tem)) reloc |= 2; else @@ -4016,6 +4023,9 @@ output_addressed_constants (tree exp) if (CONSTANT_CLASS_P (tem) || TREE_CODE (tem) == CONSTRUCTOR) output_constant_def (tem, 0); + + if (TREE_CODE (tem) == MEM_REF) + output_addressed_constants (TREE_OPERAND (tem, 0)); break; case PLUS_EXPR: --- gcc/testsuite/gcc.c-torture/execute/pr53084.c 2012-04-16 20:35:50.603036390 +0200 +++ gcc/testsuite/gcc.c-torture/execute/pr53084.c 2012-04-23 20:30:57.193896339 +0200 @@ -0,0 +1,18 @@ +/* PR middle-end/53084 */ + +extern void abort (void); + +__attribute__((noinline, noclone)) void +bar (const char *p) +{ + if (p[0] != 'o' || p[1] != 'o' || p[2]) + abort (); +} + +int +main () +{ + static const char *const foo[] = {"foo" + 1}; + bar (foo[0]); + return 0; +}