Use the Linux MADV_DONTNEED call to unmap free pages in the garbage
collector.Then keep the unmapped pages in the free list. This avoid
excessive memory fragmentation on large LTO bulds, which can lead
to gcc bumping into the Linux vm_max_map limit per process.
gcc/:
2011-10-08 Andi Kleen <ak@linux.intel.com>
PR other/50636
* config.in, configure: Regenerate.
* configure.ac (madvise): Add to AC_CHECK_FUNCS.
* ggc-page.c (USING_MADVISE): Add.
(page_entry): Add discarded field.
(alloc_page): Check for discarded pages.
(release_pages): Add USING_MADVISE branch.
From-SVN: r180093
libcpp/ChangeLog
* line-map.c (linemap_macro_map_loc_to_exp_point): Avoid setting a
variable without using it if ENABLE_CHECKING is not defined. Mark
the LOCATION parameter as being unused.
From-SVN: r180090
This patch basically arranges for the allocation size of line_map
buffers to be as close as possible to a power of two. This
*significantly* decreases peak memory consumption as (macro) maps are
numerous and stay live during all the compilation.
The patch adds a new ggc_round_alloc_size interface to the ggc
allocator. In each of the two main allocator implementations ('page'
and 'zone') the function has been extracted from the main allocation
function code and returns the actual size of the allocated memory
region, thus giving a chance to the caller to maximize the amount of
memory it actually uses from the allocated memory region. In the
'none' allocator implementation (that uses xmalloc) the
ggc_round_alloc_size just returns the requested allocation size.
Co-Authored-By: Dodji Seketeli <dodji@redhat.com>
From-SVN: r180086
This patch adds statistics about line maps' memory consumption and
macro expansion to the output of -fmem-report. It has been useful in
trying to reduce the memory consumption of the macro maps support.
Co-Authored-By: Dodji Seketeli <dodji@redhat.com>
From-SVN: r180085
This patch adds -fdebug-cpp option. When used with -E this dumps the
relevant macro map before every single token. This clutters the output
a lot but has proved to be invaluable in tracking some bugs during the
development of the virtual location support.
Co-Authored-By: Dodji Seketeli <dodji@redhat.com>
From-SVN: r180084
In this third instalment the diagnostic machinery -- when faced with
the virtual location of a token resulting from macro expansion -- uses
the new linemap APIs to unwind the stack of macro expansions that led
to that token and emits a [hopefully] more useful message than what we
have today.
diagnostic_report_current_module has been slightly changed to use the
location given by client code instead of the global input_location
variable. This results in more precise diagnostic locations in
general but then the patch adjusts some C++ tests which output changed
as a result of this.
Three new regression tests have been added.
The mandatory screenshot goes like this:
[dodji@adjoa gcc]$ cat -n test.c
1 #define OPERATE(OPRD1, OPRT, OPRD2) \
2 OPRD1 OPRT OPRD2;
3
4 #define SHIFTL(A,B) \
5 OPERATE (A,<<,B)
6
7 #define MULT(A) \
8 SHIFTL (A,1)
9
10 void
11 g ()
12 {
13 MULT (1.0);/* 1.0 << 1; <-- so this is an error. */
14 }
[dodji@adjoa gcc]$ ./cc1 -quiet -ftrack-macro-expansion test.c
test.c: In function 'g':
test.c:5:14: erreur: invalid operands to binary << (have 'double' and 'int')
test.c:2:9: note: in expansion of macro 'OPERATE'
test.c:5:3: note: expanded from here
test.c:5:14: note: in expansion of macro 'SHIFTL'
test.c:8:3: note: expanded from here
test.c:8:3: note: in expansion of macro 'MULT2'
test.c:13:3: note: expanded from here
Co-Authored-By: Dodji Seketeli <dodji@redhat.com>
From-SVN: r180083
This second instalment uses the infrastructure of the previous patch
to allocate a macro map for each macro expansion and assign a virtual
location to each token resulting from the expansion.
To date when cpp_get_token comes across a token that happens to be a
macro, the macro expander kicks in, expands the macro, pushes the
resulting tokens onto a "token context" and returns a dummy padding
token. The next call to cpp_get_token goes look into the token context
for the next token [which is going to result from the previous macro
expansion] and returns it. If the token is a macro, the macro expander
kicks in and you know the story.
This patch piggy-backs on that macro expansion process, so to speak.
First it modifies the macro expander to make it create a macro map for
each macro expansion. It then allocates a virtual location for each
resulting token. Virtual locations of tokens resulting from macro
expansions are then stored on a special kind of context called an
"expanded tokens context". In other words, in an expanded tokens
context, there are tokens resulting from macro expansion and their
associated virtual locations. cpp_get_token_with_location is modified
to return the virtual location of tokens resulting from macro
expansion. Note that once all tokens from an expanded token context have
been consumed and the context and is freed, the memory used to store the
virtual locations of the tokens held in that context is freed as well.
This helps reducing the overall peak memory consumption.
The client code that was getting macro expansion point location from
cpp_get_token_with_location now gets virtual location from it. Those
virtual locations can in turn be resolved into the different
interesting physical locations thanks to the linemap API exposed by
the previous patch.
Expensive progress. Possibly. So this whole virtual location
allocation business is switched off by default. So by default no
extended token is created. No extended token context is created
either. One has to use -ftrack-macro-expansion to switch this on. This
complicates the code but I believe it can be useful as some of our
friends found out at http://llvm.org/bugs/show_bug.cgi?id=5610
The patch tries to reduce the memory consumption by freeing some token
context memory that was being reused before. I didn't notice any
compilation slow down due to this immediate freeing on my GNU/Linux
system.
As no client code tries to resolve virtual locations to anything but
what was being done before, no new test case has been added.
Co-Authored-By: Dodji Seketeli <dodji@redhat.com>
From-SVN: r180082
This is the first instalment of a set which goal is to track locations
of tokens across macro expansions. Tom Tromey did the original work
and attached the patch to PR preprocessor/7263. This opus is a
derivative of that original work.
This patch modifies the linemap module of libcpp to add virtual
locations support.
A virtual location is a mapped location that can resolve to several
different physical locations. It can always resolve to the spelling
location of a token. For tokens resulting from macro expansion it can
resolve to:
- either the location of the expansion point of the macro.
- or the location of the token in the definition of the
macro
- or, if the token is an argument of a function-like macro,
the location of the use of the matching macro parameter in
the definition of the macro
The patch creates a new type of line map called a macro map. For every
single macro expansion, there is a macro map that generates a virtual
location for every single resulting token of the expansion.
The good old type of line map we all know is now called an ordinary
map. That one still encodes spelling locations as it has always had.
As a result linemap_lookup as been extended to return a macro map when
given a virtual location resulting from a macro expansion. The layout
of structs line_map has changed to support this new type of map. So
did the layout of struct line_maps. Accessor macros have been
introduced to avoid messing with the implementation details of these
datastructures directly. This helped already as we have been testing
different ways of arranging these datastructure. Having to constantly
adjust client code that is too tied with the internals of line_map and
line_maps would have been even more painful.
Of course, many new public functions have been added to the linemap
module to handle the resolution of virtual locations.
This patch introduces the infrastructure but no part of the compiler
uses virtual locations yet.
However the client code of the linemap data structures has been
adjusted as per the changes. E.g, it's not anymore reliable for a
client code to manipulate struct line_map directly if it just wants to
deal with spelling locations, because struct line_map can now
represent a macro map as well. In that case, it's better to use the
convenient API to resolve the initial (possibly virtual) location to a
spelling location (or to an ordinary map) and use that.
This is the reason why the patch adjusts the Java, Ada and Fortran
front ends.
Also, note that virtual locations are not supposed to be ordered for
relations '<' and '>' anymore. To test if a virtual location appears
"before" another one, one has to use a new operator exposed by the
line map interface. The patch updates the only spot (in the
diagnostics module) I have found that was making the assumption that
locations were ordered for these relations. This is the only change
that introduces a use of the new line map API in this patch, so I am
adding a regression test for it only.
From-SVN: r180081
2011-10-16 Thomas Koenig <tkoenig@gcc.gnu.org>
* frontend-passes.c (current_ns): Make static.
(create_var): Note parent of newly created namespace.
(optimize_namespace): Don't wak sibling namespaces
if they are EXEC_BLOCK because this is handled...
(gfc_code_walker): ... here. Also walk ASSOCIATE lists.
From-SVN: r180063
PR tree-optimization/50727
* tree-vect-patterns.c (vect_operation_fits_smaller_type): Add
DEF_STMT to the list of statements to be replaced by the
pattern statements.
From-SVN: r180060
* tree-vect-stmts.c (vectorizable_load): For SLP without
permutation treat the first load of the node as the first
element in its interleaving chain.
* tree-vect-slp.c (vect_get_and_check_slp_defs): Swap the
operands if necessary and possible.
(vect_build_slp_tree): Add new argument. Allow load groups of
any size in basic blocks. Keep all the loads for further
permutation check. Use the new argument to determine if there
is a permutation. Update the recursive calls.
(vect_supported_load_permutation_p): Allow subchains of
interleaving chains in basic block vectorization.
(vect_analyze_slp_instance): Update the call to
vect_build_slp_tree. Check load permutation based on the new
parameter.
(vect_schedule_slp_instance): Don't start from the first element
in interleaving chain unless the loads are permuted.
From-SVN: r180055
/cp
2011-10-15 Paolo Carlini <paolo.carlini@oracle.com>
PR c++/50732
* semantics.c (finish_trait_expr): Do not try to instantiate the
the base type of an __is_base_of trait.
(check_trait_type): Return a tree; use complete_type_or_else.
/testsuite
2011-10-15 Paolo Carlini <paolo.carlini@oracle.com>
PR c++/50732
* g++.dg/ext/is_base_of_incomplete.C: New.
* g++.dg/ext/is_base_of_diagnostic.C: Adjust dg-errors.
* g++.dg/ext/unary_trait_incomplete.C: Likewise.
From-SVN: r180048
* tree-vect-slp.c: Include langhooks.h.
(vect_create_mask_and_perm): Emit VEC_PERM_EXPR, not a builtin.
(vect_transform_slp_perm_load): Use can_vec_perm_expr_p. Simplify
mask creation for VEC_PERM_EXPR.
* tree-vect-stmts.c (perm_mask_for_reverse): Return the mask,
not the builtin.
(reverse_vec_elements): Emit VEC_PERM_EXPR not a builtin.
* Makefile.in (tree-vect-slp.o): Update dependency.
* optabs.c (can_vec_perm_expr_p): Allow NULL as unknown constant.
From-SVN: r180047
PR rtl-optimization/49941
* jump.c (mark_jump_label_1): Set JUMP_LABEL for simple_return jumps.
* rtl.h (set_return_jump_label): Declare.
* function.c (set_return_jump_label): New function, extracted..
(thread_prologue_and_epilogue_insns): ..from here. Use it in
another instance to set return jump_label.
* cfgrtl.c (force_nonfallthru_and_redirect): Use set_return_jump_label.
* reorg.c (find_end_label): Likewise.
From-SVN: r180027
2011-10-15 Nicolas Roche <roche@adacore.com>
* gcc-interface/lang-specs.h: Ensure -mrtp switch is passed when using
either rtp-smp or ravenscar-cert-rtp runtimes.
From-SVN: r180026
2011-10-15 Bob Duff <duff@adacore.com>
* exp_ch6.adb (Add_Unconstrained_Actuals_To_Build_In_Place_Call):
Do not create a pool formal on unless RE_Root_Storage_Pool_Ptr
is available.
(Expand_N_Extended_Return_Statement): Do not create a renaming of the
build-in-place pool parameter unless RE_Root_Storage_Pool_Ptr is
available.
(Make_Build_In_Place_Call_In_Allocator): Add the user-defined
pool only if RE_Root_Storage_Pool_Ptr is available.
(Make_Build_In_Place_Call_In_Object_Declaration): Do not add a
pool actual unless RE_Root_Storage_Pool_Ptr is available.
* sem_ch6.adb (Create_Extra_Formals): Add build-in-place pool
formal only if RE_Root_Storage_Pool_Ptr is available.
2011-10-15 Matthew Heaney <heaney@adacore.com>
* a-cusyqu.ads, a-cbsyqu.ads, a-cuprqu.ads, a-cbprqu.ads (Queue
type): Specify Priority aspect for protected type.
From-SVN: r180025
* config/sparc/sol2.h: Protect -m{cpu,tune}=native handling
with a more complete cpp test.
* config/sparc/linux64.h: Likewise.
* config/sparc/linux.h: Likewise.
* config/sparc/sparc.opt (sparc_debug): New target variable.
(mdebug): New target option.
* config/sparc/sparc.h (MASK_DEBUG_OPTIONS, MASK_DEBUG_ALL,
TARGET_DEBUG_OPTIONS): New defines.
* config/sparc/sparc.c (debug_target_flag_bits,
debug_target_flags): New functions.
(sparc_option_override): Add name strings back to cpu_table[].
Parse -mdebug string. When TARGET_DEBUG_OPTIONS is true, print
out the target flags before and after override processing as well
as the selected cpu. If MASK_V8PLUS, make sure that the selected
cpu is at least v9.
From-SVN: r180021
PR target/49263
* config/sh/sh.h (ZERO_EXTRACT_ANDMASK): New macro.
* config/sh/sh.c (sh_rtx_costs): Add test instruction case.
* config/sh/sh.md (tstsi_t): Name existing insn. Make inner
and instruction commutative.
(tsthi_t, tstqi_t, tstqi_t_zero, tstsi_t_and_not,
tstsi_t_zero_extract_eq, tstsi_t_zero_extract_xor,
tstsi_t_zero_extract_subreg_xor_little,
tstsi_t_zero_extract_subreg_xor_big): New insns.
(*movsicc_t_false, *movsicc_t_true): Replace space with tab in
asm output.
(*andsi_compact): Reorder alternatives so that K08 is considered
first.
* gcc.target/sh/pr49263.c: New.
From-SVN: r180020
PR target/50354
* config/sparc/linux64.h (TARGET_DEFAULT): Only override if the default
processor is at least V9 and TARGET_64BIT_DEFAULT is defined.
From-SVN: r180013
* gimplify.c (gimplify_expr): Take care that for bitwise-binary
transformation the operands have compatible types.
* gfortran.fortran-torture/compile/logical-2.f90: New test.
From-SVN: r180006
* config/i386/sse.md (vec_widen_smult_hi_v8hi,
vec_widen_smult_lo_v8hi, vec_widen_umult_hi_v8hi,
vec_widen_umult_lo_v8hi): Macroize using VI2_AVX2
mode iterator and any_extend code iterator.
(vec_widen_<s>mult_hi_v8si, vec_widen_<s>mult_lo_v8si): New
expanders.
(vec_widen_smult_hi_v4si, vec_widen_smult_lo_v4si): Enable
also for TARGET_SSE4_1 using pmuldq insn.
(sdot_prodv8hi): Macroize using VI2_AVX2 iterator.
(sse2_sse4_1): New code attr.
(udot_prodv4si): Macroize using any_extend code iterator.
(<s>dot_prodv8si): New expander.
* gcc.target/i386/sse2-mul-1.c: New test.
* gcc.target/i386/sse4_1-mul-1.c: New test.
* gcc.target/i386/avx-mul-1.c: New test.
* gcc.target/i386/xop-mul-1.c: New test.
* gcc.target/i386/avx2-mul-1.c: New test.
From-SVN: r180005
