2017-07-03 Richard Biener <rguenther@suse.de>
PR tree-optimization/60510
* tree-vect-loop.c (vect_create_epilog_for_reduction): Pass in
the scalar reduction PHI and use it.
(vectorizable_reduction): Properly guard the single_defuse_cycle
path for non-SLP reduction chains where we cannot use it.
Rework reduc_def/index and vector type deduction. Rework
vector operand gathering during reduction op code-gen.
* tree-vect-slp.c (vect_analyze_slp): For failed SLP reduction
chains dissolve the chain and leave it to non-SLP reduction
handling.
* gfortran.dg/vect/pr60510.f: New testcase.
From-SVN: r249919
This combines the information from previous patches to give a guaranteed
alignment for the DR as a whole. This should be a bit safer than using
base_element_aligned, since that only really took the base into account
(not the init or offset).
2017-07-03 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* tree-data-ref.h (dr_alignment): Declare.
* tree-data-ref.c (dr_alignment): New function.
* tree-vectorizer.h (dataref_aux): Remove base_element_aligned.
* tree-vect-data-refs.c (vect_compute_data_ref_alignment): Don't
set it.
* tree-vect-stmts.c (vectorizable_store): Use dr_alignment.
From-SVN: r249917
This patch records the base alignment and misalignment in
innermost_loop_behavior, to avoid the second-guessing that was
previously done in vect_compute_data_ref_alignment. It also makes
vect_analyze_data_refs use dr_analyze_innermost, instead of having an
almost-copy of the same code.
I wasn't sure whether the alignments should be measured in bits
(for consistency with most other interfaces) or in bytes (for consistency
with DR_ALIGNED_TO, now DR_OFFSET_ALIGNMENT, and with *_ptr_info_alignment).
I went for bytes because:
- I think in practice most consumers are going to want bytes.
E.g. using bytes avoids having to mix TYPE_ALIGN and TYPE_ALIGN_UNIT
in vect_compute_data_ref_alignment.
- It means that any bit-level paranoia is dealt with when building
the innermost_loop_behavior and doesn't get pushed down to consumers.
2017-07-03 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* tree-data-ref.h (innermost_loop_behavior): Add base_alignment
and base_misalignment fields.
(DR_BASE_ALIGNMENT, DR_BASE_MISALIGNMENT): New macros.
* tree-data-ref.c: Include builtins.h.
(dr_analyze_innermost): Set up the new innmost_loop_behavior fields.
* tree-vectorizer.h (STMT_VINFO_DR_BASE_ALIGNMENT): New macro.
(STMT_VINFO_DR_BASE_MISALIGNMENT): Likewise.
* tree-vect-data-refs.c: Include tree-cfg.h.
(vect_compute_data_ref_alignment): Use the new innermost_loop_behavior
fields instead of calculating an alignment here.
(vect_analyze_data_refs): Use dr_analyze_innermost. Dump the new
innermost_loop_behavior fields.
From-SVN: r249916
A later patch adds base alignment information to innermost_loop_behavior.
After that, the only remaining piece of alignment information that wasn't
immediately obvious was the step alignment. Adding that allows a minor
simplification to vect_compute_data_ref_alignment, and also potentially
improves the handling of variable strides for outer loop vectorisation.
A later patch will also use it to give the alignment of the DR as a whole.
2017-07-03 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* tree-data-ref.h (innermost_loop_behavior): Add a step_alignment
field.
(DR_STEP_ALIGNMENT): New macro.
* tree-vectorizer.h (STMT_VINFO_DR_STEP_ALIGNMENT): Likewise.
* tree-data-ref.c (dr_analyze_innermost): Initalize step_alignment.
(create_data_ref): Print it.
* tree-vect-stmts.c (vectorizable_load): Use the step alignment
to tell whether the step preserves vector (mis)alignment.
* tree-vect-data-refs.c (vect_compute_data_ref_alignment): Likewise.
Move the check for an integer step and generalise to all INTEGER_CST.
(vect_analyze_data_refs): Set DR_STEP_ALIGNMENT when setting DR_STEP.
Print the outer step alignment.
From-SVN: r249915
This patch renames DR_ALIGNED_TO to DR_OFFSET_ALIGNMENT, to avoid
confusion with the upcoming DR_BASE_ALIGNMENT. Nothing needed the
value as a tree, and the value is clipped to BIGGEST_ALIGNMENT
(maybe it should be MAX_OFILE_ALIGNMENT?) so we might as well use
an unsigned int instead.
2017-07-03 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* tree-data-ref.h (innermost_loop_behavior): Replace aligned_to
with offset_alignment.
(DR_ALIGNED_TO): Delete.
(DR_OFFSET_ALIGNMENT): New macro.
* tree-vectorizer.h (STMT_VINFO_DR_ALIGNED_TO): Delete.
(STMT_VINFO_DR_OFFSET_ALIGNMENT): New macro.
* tree-data-ref.c (dr_analyze_innermost): Update after above changes.
(create_data_ref): Likewise.
* tree-vect-data-refs.c (vect_compute_data_ref_alignment): Likewise.
(vect_analyze_data_refs): Likewise.
* tree-if-conv.c (if_convertible_loop_p_1): Use memset before
creating dummy innermost behavior.
From-SVN: r249914
This means that callers to dr_analyze_innermost don't need a full
data_reference and don't need to fill in any fields beforehand.
2017-07-03 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* tree-data-ref.h (dr_analyze_innermost): Replace the dr argument
with a "innermost_loop_behavior *" and refeence tree.
* tree-data-ref.c (dr_analyze_innermost): Likewise.
(create_data_ref): Update call accordingly.
* tree-predcom.c (find_looparound_phi): Likewise.
From-SVN: r249913
This patch replaces the individual stmt_vinfo dr_* fields with
an innermost_loop_behavior, so that the changes in later patches
get picked up automatically. It also adds a helper function for
getting the behavior of a data reference wrt the vectorised loop.
2017-07-03 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* tree-vectorizer.h (_stmt_vec_info): Replace individual dr_*
fields with dr_wrt_vec_loop.
(STMT_VINFO_DR_BASE_ADDRESS, STMT_VINFO_DR_INIT, STMT_VINFO_DR_OFFSET)
(STMT_VINFO_DR_STEP, STMT_VINFO_DR_ALIGNED_TO): Update accordingly.
(STMT_VINFO_DR_WRT_VEC_LOOP): New macro.
(vect_dr_behavior): New function.
(vect_create_addr_base_for_vector_ref): Remove loop parameter.
* tree-vect-data-refs.c (vect_compute_data_ref_alignment): Use
vect_dr_behavior. Use a step_preserves_misalignment_p boolean to
track whether the step preserves the misalignment.
(vect_create_addr_base_for_vector_ref): Remove loop parameter.
Use vect_dr_behavior.
(vect_setup_realignment): Update call accordingly.
(vect_create_data_ref_ptr): Likewise. Use vect_dr_behavior.
* tree-vect-loop-manip.c (vect_gen_prolog_loop_niters): Update
call to vect_create_addr_base_for_vector_ref.
(vect_create_cond_for_align_checks): Likewise.
* tree-vect-patterns.c (vect_recog_bool_pattern): Copy
STMT_VINFO_DR_WRT_VEC_LOOP as a block.
(vect_recog_mask_conversion_pattern): Likewise.
* tree-vect-stmts.c (compare_step_with_zero): Use vect_dr_behavior.
(new_stmt_vec_info): Remove redundant zeroing.
From-SVN: r249911
The existing code in arm/bpabi.h was quite fragile and relied on matching
specific CPU and/or architecture names. The introduction of the option
format for -mcpu and -march broke that in a way that would be non-trivial
to fix by updating the list. The hook in that file was always a pain
as it required every new CPU being added to be add an update here as well
(easy to miss).
I've fixed that problem once and for all by adding a new callback into
the driver to select the correct BE8 behaviour. This uses features in
the ISA capabilities list to select whether or not to use BE8 format
during linking.
I also noticed that if the user happened to pass both -mbig-endian and
-mlittle-endian on the command line then the linker spec rules would
get somewhat confused and potentially do the wrong thing. I've fixed that
by marking these options as opposites in the option descriptions. The
driver will now automatically suppress overridden options leading to the
correct desired behavior.
Whilst fixing this I noticed a couple of anomolus cases in the
existing BE8 support: we were not generating BE8 format for ARMv6 or
ARMv7-R targets. While the ARMv6 status was probably deliberate at
the time, this is probably not a good idea in the long term as the
alternative, BE32, has been deprecated by ARM. After discussion with
a couple of colleagues I've decided to change this, but to then add an
option to restore the existing behaviour at the user's option. So
this patch introduces two new options (opposites) -mbe8 and -mbe32.
This is a quiet behavior change, so I'll add a comment to the release
notes shortly.
* common/config/arm/arm-common.c (arm_be8_option): New function.
* config/arm/arm-isa.h (isa_feature): Add new feature bit isa_bit_be8.
(ISA_ARMv6): Add isa_bit_be8.
* config/arm/arm.h (arm_be8_option): Add prototype.
(BE8_SPEC_FUNCTION): New define.
(EXTRA_SPEC_FUNCTIONS): Add BE8_SPEC_FUNCTION.
* config/arm/arm.opt (mbig-endian): Mark as Negative of mlittle-endian.
(mlittle-endian): Similarly.
(mbe8, mbe32): New options.
* config/arm/bpabi.h (BE8_LINK_SPEC): Call arm_be8_option.
* doc/invoke.texi (ARM Options): Document -mbe8 and -mbe32.
From-SVN: r249909
* tree-cfgcleanup.c (want_merge_blocks_p): New function.
(cleanup_tree_cfg_bb): Use it.
* profile-count.h (profile_count::of_for_merging, profile_count::merge):
New functions.
* tree-cfg.c (gimple_merge_blocks): Use profile_count::merge.
From-SVN: r249907
2017-07-03 Richard Biener <rguenther@suse.de>
* tree-vect-loop.c (vect_create_epilog_for_reduction): Revert
back to using VIEW_CONVERT_EXPR.
From-SVN: r249902
2017-07-03 Martin Liska <mliska@suse.cz>
* asan.c (asan_emit_stack_protection): Unpoison just red zones
and shadow memory of auto variables which are subject of
use-after-scope sanitization.
(asan_expand_mark_ifn): Add do set only when is_poison.
From-SVN: r249900
dr_analyze_innermost had a "struct loop *nest" parameter that acted
like a boolean. This was added in r179161, with the idea that a
null nest selected BB-level analysis rather than loop analysis.
The handling seemed strange though. If the DR was part of a loop,
we still tried to express the base and offset values as IVs, potentially
giving a nonzero step. If that failed for any reason, we'd revert to
using the original base and offset, just as we would if we hadn't asked
for an IV in the first place.
It seems more natural to use the !in_loop handling whenever nest is null
and always set the step to zero. This actually enables one more SLP
opportunity in bb-slp-pr65935.c.
I checked out r179161 and tried the patch there. The test case added
in that revision still passes, so I don't think there was any particular
need to check simple_iv.
2017-06-28 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* tree-data-ref.c (dr_analyze_innermost): Replace the "nest"
parameter with a "loop" parameter and use it instead of the
loop containing DR_STMT. Don't check simple_iv when doing
BB analysis. Describe the two analysis modes in the comment.
gcc/testsuite/
* gcc.dg/vect/bb-slp-pr65935.c: Expect SLP to be used in main
as well.
From-SVN: r249896
2017-07-03 Tom de Vries <tom@codesourcery.com>
PR tree-optimization/81192
* tree-ssa-tail-merge.c (same_succ_hash): Use bb->loop_father->num in
hash.
(same_succ::equal): Don't find bbs to be equal if bb->loop_father
differs.
(find_same_succ_bb): Remove obsolete test on bb->loop_father->latch.
* gcc.dg/pr81192.c: Update.
From-SVN: r249894
2017-07-02 Thomas Koenig <tkoenig@gcc.gnu.org>
* intrinsics/eoshift0.c: For contiguous arrays, use
block algorithm. Use memcpy where possible.
2017-07-02 Thomas Koenig <tkoenig@gcc.gnu.org>
* gfortran/eoshift_3.f90: New test.
From-SVN: r249882
This patch splits the auto-generated inline functions out of
insn-modes.h and puts them in a new header file, insn-modes-inline.h.
It also makes coretypes.h include these files directly, rather than
indirectly via machmode.h. This in turn allows insn-modes-inline.h
and machmode.h to come later in the include list, after wide-int.h.
This is useful for later patches.
insn-modes.h itself still needs to come first, since it provides
configuration information like MAX_BITSIZE_MODE_ANY_INT, which is
used to control the size of a wide_int.
The patch also makes the generator files include machmode.h
via coretypes.h. Previously they did it by more indirect means.
Finally, the patch makes wide-int-print.h available via coretypes.h
too. There didn't seem to be any reason to force only the print
routines to be included directly, and it would be painful to extend
that approach to the SVE patches.
[Based on the code ARM contributed in branches/ARM/sve-branch@242100]
2017-07-02 Richard Sandiford <richard.sandiford@linaro.org>
Alan Hayward <alan.hayward@arm.com>
David Sherwood <david.sherwood@arm.com>
contrib/header-tools/
* graph-header-logs (ignore): Update coretypes.h header list.
gcc/
* Makefile.in (MACHMODE_H): Remove insn-modes.h
(CORETYPES_H): New define.
(MOSTLYCLEANFILES): Add insn-modes-inline.h.
(insn-modes-inline.h, s-modes-inline-h): New rules.
(generated_files): Add insn-modes-inline.h.
(RTL_BASE_H, TREE_CORE_H): Use CORETYPES_H instead of coretypes.h.
(build/gensupport.o, build/ggc-none.o, build/print-rtl.o): Likewise.
(build/read-md.o, build/read-rtl.o, build/rtl.o): Likewise.
(build/vec.o, build/hash-table.o, build/inchash.o): Likewise.
(build/gencondmd.o, build/genattr.o, build/genattr-common.o): Likewise.
(build/genattrtab.o, build/genautomata.o, build/gencheck.o): Likewise.
(build/gencodes.o, build/genconditions.o): Likewise.
(build/genconfig.o, build/genconstants.o, build/genemit.o): Likewise.
(build/genenums.o, build/genextract.o, build/genflags.o): Likewise.
(build/gentarget-def.o, build/genmddeps.o, build/genopinit.o)
(build/genoutput.o, build/genpeep.o, build/genpreds.o): Likewise.
(build/genrecog.o, build/genmddump.o, build/genmatch.o): Likewise.
(build/gencfn-macros.o, build/gcov-iov.o): Likewise.
* coretypes.h: Include everything up to real.h for generators.
Include insn-modes.h first. Include wide-int-print.h after
wide-int.h. Include insn-modes-inline.h and then machmode.h.
* machmode.h: Don't include insn-modes.h here.
* function-tests.c: Remove includes of signop.h, machmode.h,
double-int.h and wide-int.h.
* rtl.h: Likewise.
* gcc-rich-location.c: Remove includes of machmode.h, double-int.h
and wide-int.h.
* optc-save-gen.awk: Likewise.
* gencheck.c (BITS_PER_UNIT): Delete dummy definition.
* godump.c: Remove include of wide-int-print.h.
* pretty-print.h: Likewise.
* wide-int-print.cc: Likewise.
* wide-int.cc: Likewise.
* hash-map-tests.c: Remove include of signop.h.
* hash-set-tests.c: Likewise.
* rtl-tests.c: Likewise.
* mkconfig.sh: Remove include of machmode.h.
* genmodes.c (emit_insn_modes_h): Split emission of inline functions
into...
(emit_insn_modes_inline_h): ...this new function. Emit the code
into an insn-modes-inline.h header file, adding appropriate
include guards and end comments.
(emit_insn_modes_c_header): Remove include of machmode.h.
(emit_min_insn_modes_c_header): Include coretypes.h rather than
machmode.h.
(main): Handle -i flag and call emit_insn_modes_inline_h when
it is passed.
Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>
From-SVN: r249881
tree-ssa-strlen.c looks for cases in which a string is built up using
operations like:
memcpy (a, "foo", 4);
memcpy (a + 3, "bar", 4);
int x = strlen (a);
As a side-effect, it optimises the non-final memcpys so that they don't
include the nul terminator.
However, after removing some "& ~0x1"s from tree-ssa-dse.c, the DSE pass
does this optimisation itself (because it can tell that later memcpys
overwrite the terminators). The strlen pass wasn't able to handle these
pre-optimised calls in the same way as the unoptimised ones.
This patch adds support for tracking unterminated strings.
[Based on the code ARM contributed in branches/ARM/sve-branch@246236]
2017-07-02 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
* tree-ssa-strlen.c (strinfo): Rename the length field to
nonzero_chars. Add a full_string_p field.
(compare_nonzero_chars, zero_length_string_p): New functions.
(get_addr_stridx): Add an offset_out parameter.
Use compare_nonzero_chars.
(get_stridx): Update accordingly. Use compare_nonzero_chars.
(new_strinfo): Update after above changes to strinfo.
(set_endptr_and_length): Set full_string_p.
(get_string_length): Update after above changes to strinfo.
(unshare_strinfo): Update call to new_strinfo.
(maybe_invalidate): Likewise.
(get_stridx_plus_constant): Change off to unsigned HOST_WIDE_INT.
Use compare_nonzero_chars and zero_string_p. Treat nonzero_chars
as a uhwi instead of an shwi. Update after above changes to
strinfo and new_strinfo.
(zero_length_string): Assert that chainsi contains full strings.
Use zero_length_string_p. Update call to new_strinfo.
(adjust_related_strinfos): Update after above changes to strinfo.
Copy full_string_p from origsi.
(adjust_last_stmt): Use zero_length_string_p.
(handle_builtin_strlen): Update after above changes to strinfo and
new_strinfo. Install the lhs as the string length if the previous
entry didn't describe a full string.
(handle_builtin_strchr): Update after above changes to strinfo
and new_strinfo.
(handle_builtin_strcpy): Likewise.
(handle_builtin_strcat): Likewise.
(handle_builtin_malloc): Likewise.
(handle_pointer_plus): Likewise.
(handle_builtin_memcpy): Likewise. Track nonzero characters
that aren't necessarily followed by a nul terminator.
(handle_char_store): Likewise.
gcc/testsuite/
* gcc.dg/strlenopt-32.c: New testcase.
* gcc.dg/strlenopt-33.c: Likewise.
* gcc.dg/strlenopt-33g.c: Likewise.
* gcc.dg/strlenopt-34.c: Likewise.
* gcc.dg/strlenopt-35.c: Likewise.
From-SVN: r249880
In this testcase, we (correctly) record after:
strcpy (p1, "abcde");
char *p2 = strchr (p1, '\0');
strcpy (p2, q);
that the length of p1 and p2 can be calculated by converting the
second strcpy to:
tmp = stpcpy (p2, q)
and then doing tmp - p1 for p1 and tmp - p2 for p2. This is delayed
until we know whether we actually need it. Then:
char *p3 = strchr (p2, '\0');
forces us to calculate the length of p2 in this way. At this point
we had three related strinfos:
p1: delayed length, calculated from tmp = stpcpy (p2, q)
p2: known length, tmp - p2
p3: known length, 0
After:
memcpy (p3, "x", 2);
we use adjust_related_strinfos to add 1 to each length. However,
that didn't do anything for delayed lengths because:
else if (si->stmt != NULL)
/* Delayed length computation is unaffected. */
;
So after the memcpy we had:
p1: delayed length, calculated from tmp = stpcpy (p2, q)
p2: known length, tmp - p2 + 1
p3: known length, 1
where the length of p1 was no longer correct.
2017-05-16 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
PR tree-optimization/80769
* tree-ssa-strlen.c (strinfo): Document that "stmt" is also used
for malloc and calloc. Document the new invariant that all related
strinfos have delayed lengths or none do.
(verify_related_strinfos): Move earlier in file.
(set_endptr_and_length): New function, split out from...
(get_string_length): ...here. Also set the lengths of related
strinfos.
(zero_length_string): Assert that chainsi has known (rather than
delayed) lengths.
(adjust_related_strinfos): Likewise.
gcc/testsuite/
PR tree-optimization/80769
* gcc.dg/strlenopt-31.c: New test.
* gcc.dg/strlenopt-31g.c: Likewise.
From-SVN: r249879
The test case triggered this assert in vect_update_misalignment_for_peel:
gcc_assert (DR_MISALIGNMENT (dr) / dr_size ==
DR_MISALIGNMENT (dr_peel) / dr_peel_size);
The problem was that:
- one memory reference guaranteed a high base alignment, when considering
that reference in isolation. This meant that we could calculate the
vector misalignment for its DR at compile time.
- the other memory reference only guaranteed a low base alignment,
when considering that reference in isolation. We therefore couldn't
calculate the vector misalignment for its DR at compile time.
- when looking at the values of the two addresses as a pair (rather
than the memory references), it was obvious that they had the same
misalignment, whatever that misalignment happened to be.
This is working as designed, so the patch restricts the assert to cases
in which both addresses have a compile-time misalignment.
In the test case this looks like a missed opportunity. Both references
are unconditional, so it should be possible to use the highest of the
available base alignment guarantees when analyzing each reference.
A later patch does this, but the problem would still remain for
conditional references.
2017-07-02 Richard Sandiford <richard.sandiford@linaro.org>
gcc/
PR tree-optimization/81136
* tree-vect-data-refs.c (vect_update_misalignment_for_peel): Only
assert that two references with the same misalignment have the same
compile-time misalignment if those compile-time misalignments
are known.
gcc/testsuite/
PR tree-optimization/81136
* gcc.dg/vect/pr81136.c: New test.
From-SVN: r249878
A tree type dump currently doesn't print the attributes. Since we have
so many now and they do many interesting things dumping them can be
useful. So dump them by default for tree type dumps.
gcc/:
2017-07-01 Andi Kleen <ak@linux.intel.com>
* print-tree.c (print_node): Print all attributes.
From-SVN: r249877
PR sanitizer/81262
* bb-reorder.c (fix_up_fall_thru_edges): Move variable declarations to
the right scopes, make sure cond_jump isn't preserved between multiple
iterations. Search for fallthru edge whenever there are 3+ edges and
use find_fallthru_edge for it.
From-SVN: r249866
PR sanitizer/81262
* bb-reorder.c (fix_up_fall_thru_edges): Move variable declarations to
the right scopes, make sure cond_jump isn't preserved between multiple
iterations. Search for fallthru edge whenever there are 3+ edges and
use find_fallthru_edge for it.
* gcc.c-torture/compile/pr81262.c: New test.
* g++.dg/ubsan/pr81262.C: New test.
From-SVN: r249865