[PATCH] Fix PR61564 - optimize attribute/pragma accepting any option

[PATCH] Fix PR61564 - optimize attribute/pragma accepting any option

[Richard Biener]

This fixes PR61564 by diagnosing (and ignoring) options not marked with
'Optimization' being applied to #pragma GCC optimize or via the
optimize attribute.

The reason is that while we save/restore option state for 'Optimize'
marked options we don't do that for other options.  Thus while such
options do not end up in the per-function optimize state applying them
still clobbers the global state.

Bootstrap and regtest running on x86_64-unknown-linux-gnu.

Ok for trunk?

Note this also points to the unfortunate ways the FEs apply
the pragma to functions which seems to be via adding optimize
attributes literally rather than copying a tree optimization
node constructable by parse_optimize_options.  That results
in excessive diagnostics also reproducable without the patch
if you use sth like #pragma GCC optimize ("-Xhost").

Thanks,
Richard.

2016-06-07  Richard Biener  <rguenther@suse.de>

	PR c/61564
	* c-common.c (parse_optimize_options): Only apply CL_OPTIMIZATION
	options and warn about others.

	* gcc.dg/Wpragmas-1.c: New testcase.
	* gcc.dg/Wattributes-4.c: Likewise.

Index: gcc/c-family/c-common.c
===================================================================
*** gcc/c-family/c-common.c	(revision 237167)
--- gcc/c-family/c-common.c	(working copy)
*************** parse_optimize_options (tree args, bool
*** 9580,9585 ****
--- 9580,9608 ----
    decode_cmdline_options_to_array_default_mask (opt_argc, opt_argv,
  						&decoded_options,
  						&decoded_options_count);
+   /* Drop non-Optimization options.  */
+   unsigned j = 1;
+   for (i = 1; i < decoded_options_count; ++i)
+     {
+       if (! (cl_options[decoded_options[i].opt_index].flags & CL_OPTIMIZATION))
+ 	{
+ 	  ret = false;
+ 	  if (attr_p)
+ 	    warning (OPT_Wattributes,
+ 		     "bad option %s to optimize attribute",
+ 		     decoded_options[i].orig_option_with_args_text);
+ 	  else
+ 	    warning (OPT_Wpragmas,
+ 		     "bad option %s to pragma attribute",
+ 		     decoded_options[i].orig_option_with_args_text);
+ 	  continue;
+ 	}
+       if (i != j)
+ 	decoded_options[j] = decoded_options[i];
+       j++;
+     }
+   decoded_options_count = j;
+   /* And apply them.  */
    decode_options (&global_options, &global_options_set,
  		  decoded_options, decoded_options_count,
  		  input_location, global_dc);
Index: gcc/testsuite/gcc.dg/Wpragmas-1.c
===================================================================
*** gcc/testsuite/gcc.dg/Wpragmas-1.c	(revision 0)
--- gcc/testsuite/gcc.dg/Wpragmas-1.c	(working copy)
***************
*** 0 ****
--- 1,11 ----
+ /* { dg-do compile } */
+ 
+ #pragma GCC push_options
+ #pragma GCC optimize ("-fno-lto") /* { dg-warning "bad option" } */
+ int main(void){return 0;}
+ #pragma GCC pop_options
+ 
+ /* ???  The FEs still apply the pragma string as optimize attribute to
+    all functions thus the diagnostic will be repeated for each function
+    affected.  */
+ /* { dg-message "bad option" "" { target *-*-* } 0 } */
Index: gcc/testsuite/gcc.dg/Wattributes-4.c
===================================================================
*** gcc/testsuite/gcc.dg/Wattributes-4.c	(revision 0)
--- gcc/testsuite/gcc.dg/Wattributes-4.c	(working copy)
***************
*** 0 ****
--- 1,3 ----
+ /* { dg-do compile } */
+ 
+ int __attribute__((optimize("no-lto"))) main(void){return 0;} /* { dg-warning "bad option" } */

Re: [PATCH] Fix PR61564 - optimize attribute/pragma accepting any option

[Jakub Jelinek]

On Tue, Jun 07, 2016 at 10:15:39AM +0200, Richard Biener wrote:
> 
> This fixes PR61564 by diagnosing (and ignoring) options not marked with
> 'Optimization' being applied to #pragma GCC optimize or via the
> optimize attribute.
> 
> The reason is that while we save/restore option state for 'Optimize'
> marked options we don't do that for other options.  Thus while such
> options do not end up in the per-function optimize state applying them
> still clobbers the global state.
> 
> Bootstrap and regtest running on x86_64-unknown-linux-gnu.
> 
> Ok for trunk?

Ok (though it surprises me we haven't done that from the beginning).

> 2016-06-07  Richard Biener  <rguenther@suse.de>
> 
> 	PR c/61564
> 	* c-common.c (parse_optimize_options): Only apply CL_OPTIMIZATION
> 	options and warn about others.
> 
> 	* gcc.dg/Wpragmas-1.c: New testcase.
> 	* gcc.dg/Wattributes-4.c: Likewise.

	Jakub

Re: [PATCH] Fix PR61564 - optimize attribute/pragma accepting any option

[Richard Biener]

On Tue, 7 Jun 2016, Jakub Jelinek wrote:

> On Tue, Jun 07, 2016 at 10:15:39AM +0200, Richard Biener wrote:
> > 
> > This fixes PR61564 by diagnosing (and ignoring) options not marked with
> > 'Optimization' being applied to #pragma GCC optimize or via the
> > optimize attribute.
> > 
> > The reason is that while we save/restore option state for 'Optimize'
> > marked options we don't do that for other options.  Thus while such
> > options do not end up in the per-function optimize state applying them
> > still clobbers the global state.
> > 
> > Bootstrap and regtest running on x86_64-unknown-linux-gnu.
> > 
> > Ok for trunk?
> 
> Ok (though it surprises me we haven't done that from the beginning).

Yes, I was surprised by that as well...

Richard.

> > 2016-06-07  Richard Biener  <rguenther@suse.de>
> > 
> > 	PR c/61564
> > 	* c-common.c (parse_optimize_options): Only apply CL_OPTIMIZATION
> > 	options and warn about others.
> > 
> > 	* gcc.dg/Wpragmas-1.c: New testcase.
> > 	* gcc.dg/Wattributes-4.c: Likewise.

Re: [PATCH] Fix PR61564 - optimize attribute/pragma accepting any option

[Richard Biener]

On Tue, 7 Jun 2016, Jakub Jelinek wrote:

> On Tue, Jun 07, 2016 at 10:15:39AM +0200, Richard Biener wrote:
> > 
> > This fixes PR61564 by diagnosing (and ignoring) options not marked with
> > 'Optimization' being applied to #pragma GCC optimize or via the
> > optimize attribute.
> > 
> > The reason is that while we save/restore option state for 'Optimize'
> > marked options we don't do that for other options.  Thus while such
> > options do not end up in the per-function optimize state applying them
> > still clobbers the global state.
> > 
> > Bootstrap and regtest running on x86_64-unknown-linux-gnu.
> > 
> > Ok for trunk?
> 
> Ok (though it surprises me we haven't done that from the beginning).

It exposed that gcc.dg/ipa/pr70646.c uses a non-Optimization option #pragma
(fixed) and that -ffast-math is not marked Optimization (but its
"components" are, likewise -funsafe-math-optimizations is).

Fixed both, bootstrapped / tested on x86_64-unknown-linux-gnu and applied.

Richard.

2016-06-07  Richard Biener  <rguenther@suse.de>

	PR c/61564
	* c-common.c (parse_optimize_options): Only apply CL_OPTIMIZATION
	options and warn about others.
	* common.opt (ffast-math): Make Optimization.

	* gcc.dg/Wpragmas-1.c: New testcase.
	* gcc.dg/Wattributes-4.c: Likewise.
	* gcc.dg/ipa/pr70646.c: Drop optimize pragma in favor of dg-option
	entry.

Index: gcc/c-family/c-common.c
===================================================================
*** gcc/c-family/c-common.c	(revision 237167)
--- gcc/c-family/c-common.c	(working copy)
*************** parse_optimize_options (tree args, bool
*** 9580,9585 ****
--- 9580,9608 ----
    decode_cmdline_options_to_array_default_mask (opt_argc, opt_argv,
  						&decoded_options,
  						&decoded_options_count);
+   /* Drop non-Optimization options.  */
+   unsigned j = 1;
+   for (i = 1; i < decoded_options_count; ++i)
+     {
+       if (! (cl_options[decoded_options[i].opt_index].flags & CL_OPTIMIZATION))
+ 	{
+ 	  ret = false;
+ 	  if (attr_p)
+ 	    warning (OPT_Wattributes,
+ 		     "bad option %s to optimize attribute",
+ 		     decoded_options[i].orig_option_with_args_text);
+ 	  else
+ 	    warning (OPT_Wpragmas,
+ 		     "bad option %s to pragma attribute",
+ 		     decoded_options[i].orig_option_with_args_text);
+ 	  continue;
+ 	}
+       if (i != j)
+ 	decoded_options[j] = decoded_options[i];
+       j++;
+     }
+   decoded_options_count = j;
+   /* And apply them.  */
    decode_options (&global_options, &global_options_set,
  		  decoded_options, decoded_options_count,
  		  input_location, global_dc);
Index: gcc/testsuite/gcc.dg/Wpragmas-1.c
===================================================================
*** gcc/testsuite/gcc.dg/Wpragmas-1.c	(revision 0)
--- gcc/testsuite/gcc.dg/Wpragmas-1.c	(working copy)
***************
*** 0 ****
--- 1,11 ----
+ /* { dg-do compile } */
+ 
+ #pragma GCC push_options
+ #pragma GCC optimize ("-fno-lto") /* { dg-warning "bad option" } */
+ int main(void){return 0;}
+ #pragma GCC pop_options
+ 
+ /* ???  The FEs still apply the pragma string as optimize attribute to
+    all functions thus the diagnostic will be repeated for each function
+    affected.  */
+ /* { dg-message "bad option" "" { target *-*-* } 0 } */
Index: gcc/testsuite/gcc.dg/Wattributes-4.c
===================================================================
*** gcc/testsuite/gcc.dg/Wattributes-4.c	(revision 0)
--- gcc/testsuite/gcc.dg/Wattributes-4.c	(working copy)
***************
*** 0 ****
--- 1,3 ----
+ /* { dg-do compile } */
+ 
+ int __attribute__((optimize("no-lto"))) main(void){return 0;} /* { dg-warning "bad option" } */
Index: gcc/testsuite/gcc.dg/ipa/pr70646.c
===================================================================
*** gcc/testsuite/gcc.dg/ipa/pr70646.c	(revision 237165)
--- gcc/testsuite/gcc.dg/ipa/pr70646.c	(working copy)
***************
*** 1,7 ****
  /* { dg-do run } */
! /* { dg-options "-O2" } */
! 
! #pragma GCC optimize("no-unit-at-a-time")
  
  typedef unsigned char u8;
  typedef unsigned long long u64;
--- 1,5 ----
  /* { dg-do run } */
! /* { dg-options "-O2 -fno-unit-at-a-time" } */
  
  typedef unsigned char u8;
  typedef unsigned long long u64;
Index: gcc/common.opt
===================================================================
*** gcc/common.opt	(revision 237165)
--- gcc/common.opt	(working copy)
*************** EnumValue
*** 1287,1293 ****
  Enum(excess_precision) String(standard) Value(EXCESS_PRECISION_STANDARD)
  
  ffast-math
! Common
  
  ffat-lto-objects
  Common Var(flag_fat_lto_objects)
--- 1287,1293 ----
  Enum(excess_precision) String(standard) Value(EXCESS_PRECISION_STANDARD)
  
  ffast-math
! Common Optimization
  
  ffat-lto-objects
  Common Var(flag_fat_lto_objects)