[PATCH, rtl]: Mode-switching: Mark block as nontransparent if its exiting mode != no_mode

[PATCH, rtl]: Mode-switching: Mark block as nontransparent if its exiting mode != no_mode

[Uros Bizjak]

[-- Attachment #1: Type: text/plain, Size: 624 bytes --]

Hello!

Attached patch fixes an oversight in mode-switching. For blocks
without ANY mode requirements, we have to consider instructions with
MODE_AFTER mode changes. If the exiting mode from the block is
different that no_mode (the mode we start), we have to mark the block
as nontransparent.

2013-11-08  Uros Bizjak  <ubizjak@gmail.com>

    * mode-switching.c (optimize_mode_switching): Mark block as
    nontransparent, if last_mode at block exit is different from no_mode.

Bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32}
core-avx-i configured bootstrap.

OK for mainline and branches?

Thanks,
Uros.

[-- Attachment #2: p.diff.txt --]
[-- Type: text/plain, Size: 481 bytes --]

Index: mode-switching.c
===================================================================
--- mode-switching.c	(revision 204496)
+++ mode-switching.c	(working copy)
@@ -577,6 +577,8 @@ optimize_mode_switching (void)
 	    {
 	      ptr = new_seginfo (no_mode, BB_END (bb), bb->index, live_now);
 	      add_seginfo (info + bb->index, ptr);
+	      if (last_mode != no_mode)
+		bitmap_clear_bit (transp[bb->index], j);
 	    }
 	}
 #if defined (MODE_ENTRY) && defined (MODE_EXIT)

Re: [PATCH, rtl]: Mode-switching: Mark block as nontransparent if its exiting mode != no_mode

[Eric Botcazou]

> Attached patch fixes an oversight in mode-switching. For blocks
> without ANY mode requirements, we have to consider instructions with
> MODE_AFTER mode changes. If the exiting mode from the block is
> different that no_mode (the mode we start), we have to mark the block
> as nontransparent.
> 
> 2013-11-08  Uros Bizjak  <ubizjak@gmail.com>
> 
>     * mode-switching.c (optimize_mode_switching): Mark block as
>     nontransparent, if last_mode at block exit is different from no_mode.
> 
> Bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32}
> core-avx-i configured bootstrap.
> 
> OK for mainline and branches?

OK for mainline and 4.8 branch, if you complete the comment just above, for 
example "from no_mode, in which case we need to..."

-- 
Eric Botcazou

Re: [PATCH, rtl]: Mode-switching: Mark block as nontransparent if its exiting mode != no_mode

[Uros Bizjak]

On Sun, Nov 10, 2013 at 5:45 PM, Eric Botcazou <ebotcazou@adacore.com> wrote:
>> Attached patch fixes an oversight in mode-switching. For blocks
>> without ANY mode requirements, we have to consider instructions with
>> MODE_AFTER mode changes. If the exiting mode from the block is
>> different that no_mode (the mode we start), we have to mark the block
>> as nontransparent.
>>
>> 2013-11-08  Uros Bizjak  <ubizjak@gmail.com>
>>
>>     * mode-switching.c (optimize_mode_switching): Mark block as
>>     nontransparent, if last_mode at block exit is different from no_mode.
>>
>> Bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32}
>> core-avx-i configured bootstrap.
>>
>> OK for mainline and branches?
>
> OK for mainline and 4.8 branch, if you complete the comment just above, for
> example "from no_mode, in which case we need to..."

Thanks, I have committed attached patch. I will wait a couple of days
to see if the patch causes any problems with other mode-switching
targets before backporting it to 4.8.

Uros.

Re: [PATCH, rtl]: Mode-switching: Mark block as nontransparent if its exiting mode != no_mode

[Uros Bizjak]

[-- Attachment #1: Type: text/plain, Size: 1162 bytes --]

On Sun, Nov 10, 2013 at 7:40 PM, Uros Bizjak <ubizjak@gmail.com> wrote:
> On Sun, Nov 10, 2013 at 5:45 PM, Eric Botcazou <ebotcazou@adacore.com> wrote:
>>> Attached patch fixes an oversight in mode-switching. For blocks
>>> without ANY mode requirements, we have to consider instructions with
>>> MODE_AFTER mode changes. If the exiting mode from the block is
>>> different that no_mode (the mode we start), we have to mark the block
>>> as nontransparent.
>>>
>>> 2013-11-08  Uros Bizjak  <ubizjak@gmail.com>
>>>
>>>     * mode-switching.c (optimize_mode_switching): Mark block as
>>>     nontransparent, if last_mode at block exit is different from no_mode.
>>>
>>> Bootstrapped and regression tested on x86_64-pc-linux-gnu {,-m32}
>>> core-avx-i configured bootstrap.
>>>
>>> OK for mainline and branches?
>>
>> OK for mainline and 4.8 branch, if you complete the comment just above, for
>> example "from no_mode, in which case we need to..."
>
> Thanks, I have committed attached patch. I will wait a couple of days
> to see if the patch causes any problems with other mode-switching
> targets before backporting it to 4.8.

Well, now with the patch.

Uros.

[-- Attachment #2: p.diff.txt --]
[-- Type: text/plain, Size: 6456 bytes --]

Index: ChangeLog
===================================================================
--- ChangeLog	(revision 204648)
+++ ChangeLog	(working copy)
@@ -1,3 +1,8 @@
+2013-11-10  Uros Bizjak  <ubizjak@gmail.com>
+
+	* mode-switching.c (optimize_mode_switching): Mark block as
+	nontransparent, if last_mode at block exit is different from no_mode.
+
 2013-11-09  Jan-Benedict Glaw  <jbglaw@lug-owl.de>
 
 	* function.c (NAME__MAIN): Move to...
@@ -90,8 +95,7 @@
 2013-11-08  Richard Biener  <rguenther@suse.de>
 
 	PR tree-optimization/59047
-	* tree-predcom.c (ref_at_iteration): Handle bitfield accesses
-	properly.
+	* tree-predcom.c (ref_at_iteration): Handle bitfield accesses properly.
 
 2013-11-08  Ilya Enkovich  <ilya.enkovich@intel.com>
 
@@ -124,8 +128,7 @@
 	(TYPE_QUALS, TYPE_QUALS_NO_ADDR_SPACE): Add TYPE_QUAL_ATOMIC.
 	(TYPE_QUALS_NO_ADDR_SPACE_NO_ATOMIC): New macro.
 	(atomicQI_type_node, atomicHI_type_node, atomicSI_type_node)
-	(atomicDI_type_node, atomicTI_type_node): New macros for type
-	nodes.
+	(atomicDI_type_node, atomicTI_type_node): New macros for type nodes.
 	* tree.c (set_type_quals): Set TYPE_ATOMIC.
 	(find_atomic_core_type): New function.
 	(build_qualified_type): Adjust alignment for qualified types.
@@ -134,23 +137,20 @@
 	atomicHI_type_node, atomicSI_type_node, atomicDI_type_node and
 	atomicTI_type_node.
 	* print-tree.c (print_node): Print atomic qualifier.
-	* tree-pretty-print.c (dump_generic_node): Print atomic type
-	attribute.
+	* tree-pretty-print.c (dump_generic_node): Print atomic type attribute.
 	* target.def (atomic_assign_expand_fenv): New hook.
 	* doc/tm.texi.in (TARGET_ATOMIC_ASSIGN_EXPAND_FENV): New @hook.
 	* doc/tm.texi: Regenerate.
 	* targhooks.c (default_atomic_assign_expand_fenv): New function.
 	* targhooks.h (default_atomic_assign_expand_fenv): Declare.
-	* sync-builtins.def (__atomic_feraiseexcept): New built-in
-	function.
+	* sync-builtins.def (__atomic_feraiseexcept): New built-in function.
 	* config/i386/i386-builtin-types.def (VOID_FTYPE_PUSHORT): New
 	function type.
 	* config/i386/i386.c (enum ix86_builtins): Add
 	IX86_BUILTIN_FNSTENV, IX86_BUILTIN_FLDENV, IX86_BUILTIN_FNSTSW and
 	IX86_BUILTIN_FNCLEX.
 	(bdesc_special_args): Add __builtin_ia32_fnstenv,
-	__builtin_ia32_fldenv, __builtin_ia32_fnstsw and
-	__builtin_ia32_fnclex.
+	__builtin_ia32_fldenv, __builtin_ia32_fnstsw and __builtin_ia32_fnclex.
 	(ix86_expand_builtin): Handle the new built-in functions.
 	(ix86_atomic_assign_expand_fenv): New function.
 	(TARGET_ATOMIC_ASSIGN_EXPAND_FENV): New macro.
@@ -183,8 +183,7 @@
 	* tree-ssa-loop-im.c
 	(move_computations_dom_walker::before_dom_children): Same.
 	* tree-ssa-loop-manip.c (rewrite_phi_with_iv): Same.
-	* tree-ssa-pre.c (eliminate_dom_walker::before_dom_children):
-	Same.
+	* tree-ssa-pre.c (eliminate_dom_walker::before_dom_children): Same.
 	* tree-ssa-propagate.c (substitute_and_fold): Same.
 	* tree-vect-loop.c (vect_finalize_reduction): Same.
 	* tree-vect-stmts.c (vectorizable_call): Same.
@@ -218,15 +217,13 @@
 	* doc/invoke.texi (-fisolate-erroneous-paths): Document.
 
 	* gimple-ssa-isolate-paths.c (gate_isolate_erroneous_paths):
-	No longer check if we have __builtin_trap, assume it's
-	available.
+	No longer check if we have __builtin_trap, assume it's available.
 
 2013-11-07  Diego Novillo  <dnovillo@google.com>
 
 	* attribs.c (lookup_scoped_attribute_spec): Make static.
 	(get_attribute_namespace): Likewise.
-	* builtins.c (more_const_call_expr_args_p): Move from
-	tree.h.
+	* builtins.c (more_const_call_expr_args_p): Move from tree.h.
 	(validate_arglist): Move earlier in the file. Make static.
 	(expand_stack_restore): Move from stmt.c
 	(expand_stack_save): Move from stmt.c
@@ -303,7 +300,7 @@
 	(in_array_bounds_p): Move to tree-eh.c.
 	(range_in_array_bounds_p): Move to tree-eh.c.
 	(truth_type_for): Move to gimple-fold.c.
-	(list_equal_p): remove.
+	(list_equal_p): Remove.
 	* tree.h (decl_assembler_name_equal): Remove.
 	(decl_assembler_name_hash): Remove.
 	(truth_type_for): Remove.
@@ -349,13 +346,12 @@
 	* doc/sourcebuild.texi (Top Level) <lto-plugin>: GNU ld can use
 	linker plugins, too.
 
-	* config/arc/arc.h (LINK_COMMAND_SPEC): For
-	-ftree-parallelize-loops=*, link to libgomp and its dependencies.
+	* config/arc/arc.h (LINK_COMMAND_SPEC): For -ftree-parallelize-loops=*,
+	link to libgomp and its dependencies.
 	* config/ia64/hpux.h (LIB_SPEC): Likewise.
 	* config/pa/pa-hpux11.h (LIB_SPEC): Likewise.
 	* config/pa/pa64-hpux.h (LIB_SPEC): Likewise.
-	* gcc.c (GOMP_SELF_SPECS): Update comment about libgomp's
-	dependencies.
+	* gcc.c (GOMP_SELF_SPECS): Update comment about libgomp's dependencies.
 
 2013-11-07  Jakub Jelinek  <jakub@redhat.com>
 
@@ -398,8 +394,7 @@
 	(canonicalize_value): ... this.  Also handle stripping of
 	TREE_OVERFLOW.
 	(get_value, set_lattice_value, get_value_for_expr): Adjust.
-	* gimple-fold.c (canonicalize_constructor_val): Strip
-	TREE_OVERFLOW.
+	* gimple-fold.c (canonicalize_constructor_val): Strip TREE_OVERFLOW.
 	* tree-ssa-threadedge.c (set_ssa_name_value): Likewise.
 
 2013-11-07  Richard Biener  <rguenther@suse.de>
@@ -458,8 +453,7 @@
 	(finish_live_range_shrinkage): Ditto.
 	* sched-deps.c (create_insn_reg_set): Make void return value.
 	* passes.def: Add pass_live_range_shrinkage.
-	* ira.c (update_equiv_regs): Don't move if
-	flag_live_range_shrinkage.
+	* ira.c (update_equiv_regs): Don't move if flag_live_range_shrinkage.
 	* haifa-sched.c (live_range_shrinkage_p): New.
 	(initialize_live_range_shrinkage, finish_live_range_shrinkage):
 	New functions.
Index: mode-switching.c
===================================================================
--- mode-switching.c	(revision 204638)
+++ mode-switching.c	(working copy)
@@ -571,12 +571,15 @@ optimize_mode_switching (void)
 
 	  info[bb->index].computing = last_mode;
 	  /* Check for blocks without ANY mode requirements.
-	     N.B. because of MODE_AFTER, last_mode might still be different
-	     from no_mode.  */
+	     N.B. because of MODE_AFTER, last_mode might still
+	     be different from no_mode, in which case we need to
+	     mark the block as nontransparent.  */
 	  if (!any_set_required)
 	    {
 	      ptr = new_seginfo (no_mode, BB_END (bb), bb->index, live_now);
 	      add_seginfo (info + bb->index, ptr);
+	      if (last_mode != no_mode)
+		bitmap_clear_bit (transp[bb->index], j);
 	    }
 	}
 #if defined (MODE_ENTRY) && defined (MODE_EXIT)