Re: PR 23551: why should we coalesce inlined variables?

[Jan Hubicka <hubicka at ucw dot cz>]

> On Jun 28, 2007, Andrew MacLeod <amacleod@redhat.com> wrote:
> 
> > I fooled around with some simple cases, and didn't find anything that
> > made much difference.  Whats a testcase you have which shows this
> > helping?
> 
> Err...  I'm sure I had something when I posted the patch, but the
> build tree in which I came up with the testcase is long gone, and I no
> longer remember the details.  On a more recent build tree, I couldn't
> really find any case in which the patch really made a difference as to
> debug info :-(

Hi,
this patch seems to be responsible for 90% memory consumption increase
at rtl-optimization/28071 -O3:


comparing PR rtl-optimization/28071 testcase compilation at -O3 -fno-tree-pre -fno-tree-fre level:
  Overall memory allocated via mmap and sbrk increased from 1031749k to 1965705k, overall 90.52%
  Peak amount of GGC memory allocated before garbage collecting increased from 184687k to 190538k, overall 3.17%
  Peak amount of GGC memory still allocated after garbage collecting increased from 172203k to 178054k, overall 3.40%
  Amount of produced GGC garbage increased from 348399k to 389440k, overall 11.78%
    Overall memory needed: 1031749k -> 1965705k
    Peak memory use before GGC: 184687k -> 190538k
    Peak memory use after GGC: 172203k -> 178054k
    Maximum of released memory in single GGC run: 80995k -> 81025k
    Garbage: 348399k -> 389440k
    Leak: 47832k -> 47091k
    Overhead: 45175k -> 48668k
    GGC runs: 66 -> 65

Head of the ChangeLog is:

--- /usr/src/SpecTests/sandbox-britten-memory/x86_64/mem-result/ChangeLog	2007-07-06 05:24:13.000000000 +0000
+++ /usr/src/SpecTests/sandbox-britten-memory/gcc/gcc/ChangeLog	2007-07-06 21:46:50.000000000 +0000
@@ -1,3 +1,107 @@
+2007-07-06  Ian Lance Taylor  <iant@google.com>
+            Zack Weinberg  <zackw@panix.com>
+
+	PR middle-end/32441
+	* builtins.c (std_expand_builtin_va_start): Don't use make_tree.
+
+2007-07-06  Richard Sandiford  <richard@codesourcery.com>
+
+	* config/sh/sh.md (*prefetch_i4): Disable for TARGET_VXWORKS_RTP.
+	(prefetch): Likewise if "pref" would be used.
+
+2007-07-06  Josh Conner  <jconner@apple.com>
+
+	PR middle-end/32602
+	PR middle-end/32603
+	* calls.c (store_one_arg): Handle arguments which are partially
+	on the stack when detecting argument overlap.
+
+2007-07-06  Bernd Schmidt  <bernd.schmidt@analog.com>
+
+	* reload1.c (choose_reload_regs): Set reload_spill_index for regs
+	chosen during find_reloads.
+
+2007-07-06  Richard Guenther  <rguenther@suse.de>
+
+	* gimplify.c (gimplify_call_expr): Prefer DECL_ARGUMENTS over
+	TYPE_ARG_TYPES for verification of argument types.  Use
+	DECL_ARG_TYPE instead of the PARM_DECL type.  Take excess
+	parameters as variable arguments.
+
+2007-07-06  Andreas Krebbel  <krebbel1@de.ibm.com>
+
+	* libgcc2.h (word_type): Type definition removed.
+	(cmp_return_type, shift_count_type): Type definitions added.
+	(__lshrdi3, __ashldi3, __ashrdi3): word_type of second parameter
+	replaced with shift_count_type.
+	(__cmpdi2, __ucmpdi2): word_type of return type replaced with
+	cmp_return_type.
+	* libgcc2.c (__udivmoddi4, __moddi3): Type of local variable c
+	changed from word_type to Wtype.
+	(__lshrdi3, __ashldi3, __ashrdi3): word_type of second parameter
+	replaced with shift_count_type.
+	(__cmpdi2, __ucmpdi2): word_type of return type replaced with
+	cmp_return_type.
+	* c-common.c (handle_mode_attribute): Handling for libgcc_cmp_return and
+	libgcc_shift_count attribute added.
+	* target-def.h (TARGET_LIBGCC_CMP_RETURN_MODE,
+	TARGET_LIBGCC_SHIFT_COUNT_MODE): New target hooks defined.
+	(TARGET_INITIALIZER): New target hooks added.
+	* targhooks.c (default_libgcc_cmp_return_mode,
+	default_libgcc_shift_count_mode): Default implementations for the new
+	target hooks added.
+	* targhooks.h (default_libgcc_cmp_return_mode,
+	default_libgcc_shift_count_mode): Function prototypes added.
+	* target.h (struct gcc_target): Fields for the new target hooks added.
+	* optabs.c (expand_binop): Use shift_count_mode when expanding shift
+	as library call.
+	(prepare_cmp_insn): Use cmp_return_mode when expanding comparison as
+	library call.
+
+	* doc/tm.texi (TARGET_LIBGCC_CMP_RETURN_MODE,
+	TARGET_LIBGCC_SHIFT_COUNT_MODE): Documentation added.
+
+	* config/s390/s390.c (s390_libgcc_cmp_return_mode,
+	s390_libgcc_shift_count_mode): Functions added.
+	(TARGET_LIBGCC_CMP_RETURN_MODE,	TARGET_LIBGCC_SHIFT_COUNT_MODE): Target
+	hooks defined.
+
+2007-07-06  Richard Sandiford  <richard@codesourcery.com>
+
+	* config/mips/mips.c (compute_frame_size): Restore the original
+	gp_sp_offset for !GENERATE_MIPS16E_SAVE_RESTORE and remove the
+	fp_size term from the GENERATE_MIPS16E_SAVE_RESTORE calculation.
+	Document why the difference is needed.
+
+2007-07-06  Richard Guenther  <rguenther@suse.de>
+
+	* c-common.c (boolean_increment): Use correctly typed
+	constant.
+
+2007-07-06  Richard Sandiford  <richard@codesourcery.com>
+
+	* config/mips/mips.c (mips16e_save_restore_pattern_p): Check that
+	the topmost argument register is not also included in the save mask.
+	(mips16e_collect_argument_save_p): Take a pointer to the argument
+	register, rather than a pointer to the number of arguments.
+	(mips16e_collect_argument_saves): Only include argument saves
+	that aren't in the register mask.
+
+2007-07-06  Uros Bizjak  <ubizjak@gmail.com>
+
+	PR rtl_optimization/32450
+	* function.c (thread_prologue_and_epilogue_insns): Emit blockage insn
+	to ensure that instructions are not moved into the prologue when
+	profiling is on.  Remove unused prologue_end variable.
+	(expand_function_end): Emit blockage insn instead of ASM_INPUT rtx
+	as a scheduling barrier.
+
+2007-07-06  Alexandre Oliva  <aoliva@redhat.com>
+
+	PR debug/23551
+	* tree-ssa-copyrename.c (copy_rename_partition_coalesce):
+	Disregard DECL_FROM_INLINE.
+
 2007-07-05  Adam Nemet  <anemet@caviumnetworks.com>
 
 	* rtlanal.c (num_sign_bit_copies1): Improve cases of ANDing or
--- /usr/src/SpecTests/sandbox-britten-memory/x86_64/mem-result/ChangeLog.cp	2007-07-06 05:24:12.000000000 +0000
+++ /usr/src/SpecTests/sandbox-britten-memory/gcc/gcc/cp/ChangeLog	2007-07-06 21:46:49.000000000 +0000
@@ -1,3 +1,8 @@
+2007-07-06  Richard Guenther  <rguenther@suse.de>
+
+	* init.c (build_new_1): Use the correct pointer type.
+	* typeck2.c (build_m_component_ref): Likewise.
+
 2007-07-05  Mark Mitchell  <mark@codesourcery.com>
 
 	PR c++/32245


The results can be reproduced by building a compiler with

--enable-gather-detailed-mem-stats targetting x86-64

and compiling preprocessed combine.c or testcase from PR8632 with:

-fmem-report --param=ggc-min-heapsize=1024 --param=ggc-min-expand=1 -Ox -Q

The memory consumption summary appears in the dump after detailed listing
of the places they are allocated in.  Peak memory consumption is actually
computed by looking for maximal value in {GC XXXX -> YYYY} report.

Your testing script.