This patch defines the default parameters and validation for the aarch64
stack clash probing interval and guard sizes. It cleans up the previous
implementation and insures that at no point the invalidate arguments are
present in the pipeline for AArch64. Currently they are only corrected once
cc1 initalizes the back-end.
The default for AArch64 is 64 KB for both of these and we only support 4 KB and 64 KB
probes. We also enforce that any value you set here for the parameters must be
in sync.
If an invalid value is specified an error will be generated and compilation aborted.
gcc/
* common/config/aarch64/aarch64-common.c (TARGET_OPTION_DEFAULT_PARAM,
aarch64_option_default_param): New.
(params.h): Include.
(TARGET_OPTION_VALIDATE_PARAM, aarch64_option_validate_param): New.
* config/aarch64/aarch64.c (aarch64_override_options_internal): Simplify
stack-clash protection validation code.
From-SVN: r264757
This patch changes it so that default parameters are validated during
initialization. This change is needed to ensure parameters set via by the
target specific common initialization routines still keep the parameters within
the valid range.
gcc/
* params.c (validate_param): New.
(add_params): Use it.
(set_param_value): Refactor param validation into validate_param.
(diagnostic.h): Include.
* diagnostic.h (diagnostic_ready_p): New.
From-SVN: r264756
This patch adds the ability for backends to add custom constrains to the param
values by defining a new hook option_validate_param.
This hook is invoked on every set_param_value which allows the back-end to
ensure that the parameters are always within it's desired state.
gcc/
* params.c (set_param_value):
Add index of parameter being validated.
* common/common-target.def (option_validate_param): New.
* common/common-targhooks.h (default_option_validate_param): New.
* common/common-targhooks.c (default_option_validate_param): New.
* doc/tm.texi.in (TARGET_OPTION_VALIDATE_PARAM): New.
* doc/tm.texi: Regenerate.
From-SVN: r264755
This patch cleans up the testsuite when a run is done with stack clash
protection turned on.
Concretely this switches off -fstack-clash-protection for a couple of tests:
* assembler scan: some tests are quite fragile in that they check for exact
assembly output, e.g. check for exact amount of sub etc. These won't
match now.
* vla: Some of the ubsan tests negative array indices. Because the arrays weren't
used before the incorrect $sp wouldn't have been used. The correct value is
restored on ret. Now however we probe the $sp which causes a segfault.
* params: When testing the parameters we have to skip these on AArch64 because of our
custom constraints on them. We already test them separately so this isn't a
loss.
Note that the testsuite is not entire clean due to gdb failure caused by alloca with
stack clash. On AArch64 we output an incorrect .loc directive, but this is already the
case with the current implementation in GCC and is a bug unrelated to this patch series.
gcc/testsuite/
PR target/86486
* gcc.dg/pr82788.c: Skip for AArch64.
* gcc.dg/guality/vla-1.c: Turn off stack-clash.
* gcc.target/aarch64/subsp.c: Likewise.
* gcc.dg/params/blocksort-part.c: Skip stack-clash checks
on AArch64.
* gcc.dg/stack-check-10.c: Add AArch64 specific checks.
* gcc.dg/stack-check-12.c: ILP32 fixup.
* gcc.dg/stack-check-5.c: Add AArch64 specific checks.
* gcc.dg/stack-check-6a.c: Skip on AArch64, we don't support this.
* testsuite/lib/target-supports.exp
(check_effective_target_frame_pointer_for_non_leaf): AArch64 does not
require frame pointer for non-leaf functions.
From-SVN: r264754
This patch enforces that the default guard size for stack-clash protection for
AArch64 be 64KB unless the user has overriden it via configure in which case
the user value is used as long as that value is within the valid range.
It also does some basic validation to ensure that the guard size is only 4KB or
64KB and also enforces that for aarch64 the stack-clash probing interval is
equal to the guard size.
gcc/
PR target/86486
* config/aarch64/aarch64.c (aarch64_override_options_internal):
Add validation for stack-clash parameters and set defaults.
From-SVN: r264753
This patch defines a configure option to allow the setting of the default
guard size via configure flags when building the target.
The new flag is:
* --with-stack-clash-protection-guard-size=<num>
The patch defines a new macro DEFAULT_STK_CLASH_GUARD_SIZE which targets need
to use explicitly is they want to support this configure flag and values that
users may have set.
gcc/
PR target/86486
* configure.ac: Add stack-clash-protection-guard-size.
* doc/install.texi: Document it.
* config.in (DEFAULT_STK_CLASH_GUARD_SIZE): New.
* params.def: Update comment for guard-size.
(PARAM_STACK_CLASH_PROTECTION_GUARD_SIZE,
PARAM_STACK_CLASH_PROTECTION_PROBE_INTERVAL): Update description.
* configure: Regenerate.
From-SVN: r264752
This patch adds a requirement that the number of outgoing arguments for a
function is at least 8 bytes when using stack-clash protection and alloca.
By using this condition we can avoid a check in the alloca code and so have
smaller and simpler code there.
A simplified version of the AArch64 stack frames is:
+-----------------------+
| |
| |
| |
+-----------------------+
|LR |
+-----------------------+
|FP |
+-----------------------+
|dynamic allocations | ---- expanding area which will push the outgoing
+-----------------------+ args down during each allocation.
|padding |
+-----------------------+
|outgoing stack args | ---- safety buffer of 8 bytes (aligned)
+-----------------------+
By always defining an outgoing argument, alloca(0) effectively is safe to probe
at $sp due to the reserved buffer being there. It will never corrupt the stack.
This is also safe for alloca(x) where x is 0 or x % page_size == 0. In the
former it is the same case as alloca(0) while the latter is safe because any
allocation pushes the outgoing stack args down:
|FP |
+-----------------------+
| |
|dynamic allocations | ---- alloca (x)
| |
+-----------------------+
|padding |
+-----------------------+
|outgoing stack args | ---- safety buffer of 8 bytes (aligned)
+-----------------------+
Which means when you probe for the residual, if it's 0 you'll again just probe
in the outgoing stack args range, which we know is non-zero (at least 8 bytes).
gcc/
PR target/86486
* config/aarch64/aarch64.h (STACK_CLASH_MIN_BYTES_OUTGOING_ARGS,
STACK_DYNAMIC_OFFSET): New.
* config/aarch64/aarch64.c (aarch64_layout_frame):
Update outgoing args size.
(aarch64_stack_clash_protection_alloca_probe_range,
TARGET_STACK_CLASH_PROTECTION_ALLOCA_PROBE_RANGE): New.
gcc/testsuite/
PR target/86486
* gcc.target/aarch64/stack-check-alloca-1.c: New.
* gcc.target/aarch64/stack-check-alloca-10.c: New.
* gcc.target/aarch64/stack-check-alloca-2.c: New.
* gcc.target/aarch64/stack-check-alloca-3.c: New.
* gcc.target/aarch64/stack-check-alloca-4.c: New.
* gcc.target/aarch64/stack-check-alloca-5.c: New.
* gcc.target/aarch64/stack-check-alloca-6.c: New.
* gcc.target/aarch64/stack-check-alloca-7.c: New.
* gcc.target/aarch64/stack-check-alloca-8.c: New.
* gcc.target/aarch64/stack-check-alloca-9.c: New.
* gcc.target/aarch64/stack-check-alloca.h: New.
* gcc.target/aarch64/stack-check-14.c: New.
* gcc.target/aarch64/stack-check-15.c: New.
From-SVN: r264751
This patch adds a hook to tell the mid-end about the probing requirements of the
target. On AArch64 we allow a specific range for which no probing needs to
be done. This same range is also the amount that will have to be probed up when
a probe is needed after dropping the stack.
Defining this probe comes with the extra requirement that the outgoing arguments
size of any function that uses alloca and stack clash be at the very least 8
bytes. With this invariant we can skip doing the zero checks for alloca and
save some code.
A simplified version of the AArch64 stack frame is:
+-----------------------+
| |
| |
| |
+-----------------------+
|LR |
+-----------------------+
|FP |
+-----------------------+
|dynamic allocations | -\ probe range hook effects these
+-----------------------+ --\ and ensures that outgoing stack
|padding | -- args is always > 8 when alloca.
+-----------------------+ ---/ Which means it's always safe to probe
|outgoing stack args |-/ at SP
+-----------------------+
This allows us to generate better code than without the hook without affecting
other targets.
With this patch I am also removing the stack_clash_protection_final_dynamic_probe
hook which was added specifically for AArch64 but that is no longer needed.
gcc/
PR target/86486
* explow.c (anti_adjust_stack_and_probe_stack_clash): Support custom
probe ranges.
* target.def (stack_clash_protection_alloca_probe_range): New.
(stack_clash_protection_final_dynamic_probe): Remove.
* targhooks.h (default_stack_clash_protection_alloca_probe_range) New.
(default_stack_clash_protection_final_dynamic_probe): Remove.
* targhooks.c: Likewise.
* doc/tm.texi.in (TARGET_STACK_CLASH_PROTECTION_ALLOCA_PROBE_RANGE): New.
(TARGET_STACK_CLASH_PROTECTION_FINAL_DYNAMIC_PROBE): Remove.
* doc/tm.texi: Regenerate.
From-SVN: r264750
This patch adds basic support for SVE stack clash protection.
It is a first implementation and will use a loop to do the
probing and stack adjustments.
An example sequence is:
.cfi_startproc
mov x15, sp
cntb x16, all, mul #11
add x16, x16, 304
.cfi_def_cfa_register 15
.SVLPSPL0:
cmp x16, 61440
b.lt .SVLPEND0
sub sp, sp, 61440
str xzr, [sp, 0]
sub x16, x16, 61440
b .SVLPSPL0
.SVLPEND0:
sub sp, sp, x16
.cfi_escape 0xf,0xc,0x8f,0,0x92,0x2e,0,0x8,0x58,0x1e,0x23,0xb0,0x2,0x22
for a 64KB guard size, and for a 4KB guard size
.cfi_startproc
mov x15, sp
cntb x16, all, mul #11
add x16, x16, 304
.cfi_def_cfa_register 15
.SVLPSPL0:
cmp x16, 3072
b.lt .SVLPEND0
sub sp, sp, 3072
str xzr, [sp, 0]
sub x16, x16, 3072
b .SVLPSPL0
.SVLPEND0:
sub sp, sp, x16
.cfi_escape 0xf,0xc,0x8f,0,0x92,0x2e,0,0x8,0x58,0x1e,0x23,0xb0,0x2,0x22
This has about the same semantics as alloca, except we prioritize the common case
where no probe is required. We also change the amount we adjust the stack and
the probing interval to be the nearest value to `guard size - abi buffer` that
fits in the 12-bit shifted immediate used by cmp.
While this would mean we probe a bit more often than we require, in practice the
amount of SVE vectors you'd need to spill is significant. Even more so to enter the
loop more than once.
gcc/
PR target/86486
* config/aarch64/aarch64-protos.h (aarch64_output_probe_sve_stack_clash): New.
* config/aarch64/aarch64.c (aarch64_output_probe_sve_stack_clash,
aarch64_clamp_to_uimm12_shift): New.
(aarch64_allocate_and_probe_stack_space): Add SVE specific section.
* config/aarch64/aarch64.md (probe_sve_stack_clash): New.
gcc/testsuite/
PR target/86486
* gcc.target/aarch64/stack-check-prologue-16.c: New test
* gcc.target/aarch64/stack-check-cfa-3.c: New test.
* gcc.target/aarch64/sve/struct_vect_24.c: New test.
* gcc.target/aarch64/sve/struct_vect_24_run.c: New test.
From-SVN: r264749
Since stack clash depends on the LR being saved for non-leaf functions this
patch adds an assert such that if this changes we would notice this.
gcc/
PR target/86486
* config/aarch64/aarch64.c (aarch64_layout_frame): Add assert.
From-SVN: r264748
This patch implements the use of the stack clash mitigation for aarch64.
In Aarch64 we expect both the probing interval and the guard size to be 64KB
and we enforce them to always be equal.
We also probe up by 1024 bytes in the general case when a probe is required.
AArch64 has the following probing conditions:
1a) Any initial adjustment less than 63KB requires no probing. An ABI defined
safe buffer of 1Kbytes is used and a page size of 64k is assumed.
b) Any final adjustment residual requires a probe at SP + 1KB.
We know this to be safe since you would have done at least one page worth
of allocations already to get to that point.
c) Any final adjustment more than remainder (total allocation amount) larger
than 1K - LR offset requires a probe at SP.
safe buffer mentioned in 1a is maintained by the storing of FP/LR.
In the case of -fomit-frame-pointer we can still count on LR being stored
if the function makes a call, even if it's a tail call. The AArch64 frame
layout code guarantees this and tests have been added to check against
this particular case.
2) Any allocations larger than 1 page size, is done in increments of page size
and probed up by 1KB leaving the residuals.
3a) Any residual for initial adjustment that is less than guard-size - 1KB
requires no probing. Essentially this is a sliding window. The probing
range determines the ABI safe buffer, and the amount to be probed up.
Incrementally allocating less than the probing thresholds, e.g. recursive functions will
not be an issue as the storing of LR counts as a probe.
+-------------------+
| ABI SAFE REGION |
+------------------------------
| | |
| | |
| | |
| | |
| | |
| | |
maximum amount | | |
not needing a | | |
probe | | |
| | |
| | |
| | |
| | | Probe offset when
| ---------------------------- probe is required
| | |
+-------- +-------------------+ -------- Point of first probe
| ABI SAFE REGION |
---------------------
| |
| |
| |
Bootstrapped Regtested on aarch64-none-linux-gnu and no issues.
Target was tested with stack clash on and off by default.
GLIBC testsuite also ran with stack clash on by default and no new
regressions.
Co-Authored-By: Richard Sandiford <richard.sandiford@linaro.org>
Co-Authored-By: Tamar Christina <tamar.christina@arm.com>
From-SVN: r264747
Currently some target supports checks such as vect_int cache their
results in a manner that would cause them not to be rechecked when
running the same tests against a different variant in a multi variant
run. This causes tests to be skipped or run when they shouldn't be.
there is already an existing caching mechanism in place that does the
caching correctly, but presumably these weren't used because some of these
tests originally only contained static data. e.g. only checked if the target is
aarch64*-*-* etc.
This patch changes every function that needs to do any caching at all to use
check_cached_effective_target which will cache per variant instead of globally.
For those tests that already parameterize over et_index I have created
check_cached_effective_target_indexed to handle this common case by creating a list
containing the property name and the current value of et_index.
These changes result in a much simpler implementation for most tests and a large
reduction in lines for target-supports.exp.
Regtested on
aarch64-none-elf
x86_64-pc-linux-gnu
powerpc64-unknown-linux-gnu
arm-none-eabi
and no testsuite errors. Difference would depend on your site.exp.
On arm we get about 4500 new testcases and on aarch64 the low 10s.
On PowerPC and x86_64 no changes as expected since the default exp for these
just test the default configuration.
What this means for new target checks is that they should always use either
check_cached_effective_target or check_cached_effective_target_indexed if the
result of the check is to be cached.
As an example the new vect_int looks like
proc check_effective_target_vect_int { } {
return [check_cached_effective_target_indexed <name> {
expr {
<condition>
}}]
}
The debug information that was once there is now all hidden in
check_cached_effective_target, (called from check_cached_effective_target_indexed)
and so the only thing you are required to do is give it a unique cache name and a condition.
The condition doesn't need to be an if statement so simple boolean expressions are enough here:
[istarget i?86-*-*] || [istarget x86_64-*-*]
|| ([istarget powerpc*-*-*]
&& ![istarget powerpc-*-linux*paired*])
|| ...
From-SVN: r264745
Avoid constants to end up in the limm field for particular
instructions when compiling for size.
gcc/
xxxx-xx-xx Claudiu Zissulescu <claziss@synopsys.com>
* config/arc/arc.md (*add_n): Clean up pattern, update instruction
constraints.
(ashlsi3_insn): Update instruction constraints.
(ashrsi3_insn): Likewise.
(rotrsi3): Likewise.
(add_shift): Likewise.
* config/arc/constraints.md (Csz): New 32 bit constraint. It
avoids placing in the limm field small constants which, otherwise,
could end into a small instruction.
testsuite/
xxxx-xx-xx Claudiu Zissulescu <claziss@synopsys.com>
* gcc.target/arc/tph_addx.c: New test.
From-SVN: r264737
gcc/
Claudiu Zissulescu <claziss@synopsys.com>
* config/arc/arc.md (maddsidi4_split): Don't use dmac if the
destination register is not odd-even.
(umaddsidi4_split): Likewise.
gcc/testsuite/
Claudiu Zissulescu <claziss@synopsys.com>
* gcc.target/arc/tmac-3.c: New file.
From-SVN: r264736
2018-10-01 Richard Biener <rguenther@suse.de>
* tree-inline.c (expand_call_inline): Store origin of fn
in BLOCK_ABSTRACT_ORIGIN for the inline BLOCK.
* tree.c (block_ultimate_origin): Simplify and do some
checking.
From-SVN: r264734
for gcc/ada/ChangeLog
* gcc-interface/lang-specs.h (default_compilers): When given
fcompare-debug-second, adjust auxbase like cc1, and pass
gnatd_A.
* gcc-interface/misc.c (flag_compare_debug): Remove variable.
(gnat_post_options): Do not set it.
* lib-writ.adb (flag_compare_debug): Remove import.
(Write_ALI): Do not test it.
From-SVN: r264732
* config/i386/mmx.md (EMMS): New int iterator.
(emms): New int attribute.
(mmx_<emms>): Macroize insn from *mmx_emms and *mmx_femms using
EMMS int iterator. Explicitly declare clobbers.
(mmx_emms): Remove expander.
(mmx_femms): Ditto.
* config/i386/predicates.md (emms_operation): Remove predicate.
(vzeroall_pattern): New predicate.
(vzeroupper_pattern): Rename from vzeroupper_operation.
* config/i386/i386.c (ix86_avx_u128_mode_after): Use
vzeroupper_pattern and vzeroall_pattern predicates.
From-SVN: r264727
gcc/
PR rtl-optimization/86939
* ira-lives.c (make_hard_regno_born): Rename from this...
(make_hard_regno_live): ... to this. Remove update to conflict
information. Update function comment.
(make_hard_regno_dead): Add conflict information update. Update
function comment.
(make_object_born): Rename from this...
(make_object_live): ... to this. Remove update to conflict information.
Update function comment.
(make_object_dead): Add conflict information update. Update function
comment.
(mark_pseudo_regno_live): Call make_object_live.
(mark_pseudo_regno_subword_live): Likewise.
(mark_hard_reg_dead): Update function comment.
(mark_hard_reg_live): Call make_hard_regno_live.
(process_bb_node_lives): Likewise.
* lra-lives.c (make_hard_regno_born): Rename from this...
(make_hard_regno_live): ... to this. Remove update to conflict
information. Remove now uneeded check_pic_pseudo_p argument.
Update function comment.
(make_hard_regno_dead): Add check_pic_pseudo_p argument and add update
to conflict information. Update function comment.
(mark_pseudo_live): Remove update to conflict information. Update
function comment.
(mark_pseudo_dead): Add conflict information update.
(mark_regno_live): Call make_hard_regno_live.
(mark_regno_dead): Call make_hard_regno_dead with new arguement.
(process_bb_lives): Call make_hard_regno_live and make_hard_regno_dead.
From-SVN: r264726
2018-09-30 Paul Thomas <pault@gcc.gnu.org>
PR fortran/87359
* trans-array.c (gfc_is_reallocatable_lhs): Correct the problem
introduced by r264358, which prevented components of associate
names from being reallocated on assignment.
2018-09-30 Paul Thomas <pault@gcc.gnu.org>
PR fortran/87359
* gfortran.dg/associate_40.f90 : New test.
From-SVN: r264725
2018-09-30 Paul Thomas <pault@gcc.gnu.org>
PR fortran/70752
PR fortran/72709
* trans-array.c (gfc_conv_scalarized_array_ref): If this is a
deferred type and the info->descriptor is present, use the
info->descriptor
(gfc_conv_array_ref): Is the se expr is a descriptor type, pass
it as 'decl' rather than the symbol backend_decl.
(gfc_array_allocate): If the se string_length is a component
reference, fix it and use it for the expression string length
if the latter is not a variable type. If it is a variable do
an assignment. Make use of component ref string lengths to set
the descriptor 'span'.
(gfc_conv_expr_descriptor): For pointer assignment, do not set
the span field if gfc_get_array_span returns zero.
* trans.c (get_array_span): If the upper bound a character type
is zero, use the descriptor span if available.
2018-09-30 Paul Thomas <pault@gcc.gnu.org>
PR fortran/70752
PR fortran/72709
* gfortran.dg/deferred_character_25.f90 : New test.
* gfortran.dg/deferred_character_26.f90 : New test.
* gfortran.dg/deferred_character_27.f90 : New test to verify
that PR82617 remains fixed.
From-SVN: r264724
2018-09-30 Paul Thomas <pault@gcc.gnu.org>
PR fortran/70149
* trans-decl.c (gfc_get_symbol_decl): A deferred character
length pointer that is initialized needs the string length to
be initialized as well.
2018-09-30 Paul Thomas <pault@gcc.gnu.org>
PR fortran/70149
* gfortran.dg/deferred_character_24.f90 : New test.
From-SVN: r264721
When passing and returning BLKmode values in 2 integer registers, use
1 TImode register instead of 2 DImode registers. Otherwise, V1TImode
may be used to move and store such BLKmode values, which prevent RTL
optimizations.
gcc/
PR target/87370
* config/i386/i386.c (construct_container): Use TImode for
BLKmode values in 2 integer registers.
gcc/testsuite/
PR target/87370
* gcc.target/i386/pr87370.c: New test.
From-SVN: r264716
2018-09-29 Paul Thomas <pault@gcc.gnu.org>
PR fortran/65667
* trans-expr.c (gfc_trans_assignment_1): If there is dependency
fix the rse stringlength.
2018-09-29 Paul Thomas <pault@gcc.gnu.org>
PR fortran/65667
* gfortran.dg/dependency_52.f90 : New test.
From-SVN: r264715
* builtins.c (unterminated_array): Pass in c_strlen_data * to
c_strlen rather than just a tree *.
(c_strlen): Change NONSTR argument to a c_strlen_data pointer.
Update recursive calls appropriately. If caller did not provide a
suitable data pointer, create a local one. When a non-terminated
string is discovered, bubble up information about the string via the
c_strlen_data object.
* builtins.h (c_strlen): Update prototype.
(c_strlen_data): New structure.
* gimple-fold.c (get_range_strlen): Update calls to c_strlen.
For a type 2 call, if c_strlen indicates a non-terminated string
use the length of the non-terminated string.
(gimple_fold_builtin_stpcpy): Update calls to c_strlen.
From-SVN: r264712
PR target/87467
* config/i386/avx512fintrin.h (_mm512_abs_pd, _mm512_mask_abs_pd): Use
__m512d type for __A argument rather than __m512.
* gcc.target/i386/avx512f-abspd-1.c (SIZE): Divide by two.
(CALC): Use double instead of float.
(TEST): Adjust to test _mm512_abs_pd and _mm512_mask_abs_pd rather than
_mm512_abs_ps and _mm512_mask_abs_ps.
From-SVN: r264711
* doc/xml/gnu/fdl-1.3.xml: The Free Software Foundation web
site now uses https. Also omit the unnecessary trailing slash.
* doc/xml/gnu/gpl-3.0.xml: Ditto.
From-SVN: r264710
* match.pd (simple_comparison): Don't optimize if either operand is
a function pointer when target needs function pointer canonicalization.
From-SVN: r264705
Now that e.g. ASM_CPU_POWER5_SPEC is always "-mpower5" it is clearer and
easier to just write that directly.
* config/rs6000/driver-rs6000.c (asm_names): Adjust the entries for
power5 .. power9 to remove indirection.
* config/rs6000/rs6000.h (ASM_CPU_POWER5_SPEC, ASM_CPU_POWER6_SPEC,
ASM_CPU_POWER7_SPEC, ASM_CPU_POWER8_SPEC, ASM_CPU_POWER9_SPEC,
ASM_CPU_476_SPEC): Delete.
(ASM_CPU_SPEC): Adjust.
(EXTRA_SPECS): Delete asm_cpu_power5, asm_cpu_power6, asm_cpu_power7,
asm_cpu_power8, asm_cpu_power9, asm_cpu_476.
From-SVN: r264704
All supported assemblers know lwsync, so we never need to generate this
instruction using the .long escape hatch.
* config.in (HAVE_AS_LWSYNC): Delete.
* config/powerpcspe/powerpcspe.h (TARGET_LWSYNC_INSTRUCTION): Delete.
* config/powerpcspe/sync.md (*lwsync): Always generate lwsync, never
do it as a .long .
* config/rs6000/rs6000.h (TARGET_LWSYNC_INSTRUCTION): Delete.
* config/rs6000/sync.md (*lwsync): Always generate lwsync, never do it
as a .long .
* configure.ac: Delete HAVE_AS_LWSYNC.
* configure: Regenerate.
From-SVN: r264702
* calls.c (expand_call): Try to do a tail call for thunks at -O0 too.
* cgraph.h (struct cgraph_thunk_info): Add indirect_offset.
(cgraph_node::create_thunk): Add indirect_offset parameter.
(thunk_adjust): Likewise.
* cgraph.c (cgraph_node::create_thunk): Add indirect_offset parameter
and initialize the corresponding field with it.
(cgraph_node::dump): Dump indirect_offset field.
* cgraphclones.c (duplicate_thunk_for_node): Deal with indirect_offset.
* cgraphunit.c (cgraph_node::analyze): Be prepared for external thunks.
(thunk_adjust): Add indirect_offset parameter and deal with it.
(cgraph_node::expand_thunk): Deal with the indirect_offset field and
pass it to thunk_adjust. Do not call the target hook if it's non-zero
or if the thunk is external or local. Fix formatting. Do not chain
the RESULT_DECL to BLOCK_VARS. Pass the static chain to the target,
if any, in the GIMPLE representation.
* ipa-icf.c (sem_function::equals_wpa): Deal with indirect_offset.
* lto-cgraph.c (lto_output_node): Write indirect_offset field.
(input_node): Read indirect_offset field.
* tree-inline.c (expand_call_inline): Pass indirect_offset field in the
call to thunk_adjust.
* tree-nested.c (struct nesting_info): Add thunk_p field.
(create_nesting_tree): Set it.
(convert_all_function_calls): Copy static chain from targets to thunks.
(finalize_nesting_tree_1): Return early for thunks.
(unnest_nesting_tree_1): Do not finalize thunks.
(gimplify_all_functions): Do not gimplify thunks.
cp/
* method.c (use_thunk): Adjust call to cgraph_node::create_thunk.
ada/
* gcc-interface/decl.c (is_cplusplus_method): Do not require C++
convention on Interfaces.
* gcc-interface/trans.c (Subprogram_Body_to_gnu): Try to create a
bona-fide thunk and hand it over to the middle-end.
(get_controlling_type): New function.
(use_alias_for_thunk_p): Likewise.
(thunk_labelno): New static variable.
(make_covariant_thunk): New function.
(maybe_make_gnu_thunk): Likewise.
* gcc-interface/utils.c (finish_subprog_decl): Set DECL_CONTEXT of the
result DECL here instead of...
(end_subprog_body): ...here.
Co-Authored-By: Pierre-Marie de Rodat <derodat@adacore.com>
From-SVN: r264701
As noted at Cauldron, dumpfile.c currently emits "note: " for all kinds
of dump message, so that (after filtering) there's no distinction between
MSG_OPTIMIZED_LOCATIONS vs MSG_NOTE vs MSG_MISSED_OPTIMIZATION in the
textual output.
This patch changes dumpfile.c so that the "note: " varies to show
which MSG_* was used, with the string prefix matching that used for
filtering in -fopt-info, hence e.g.
directive_unroll_3.f90:24:0: optimized: loop unrolled 7 times
and:
pr19210-1.c:24:3: missed: missed loop optimization: niters analysis ends up with assumptions.
The patch adds "dg-optimized" and "dg-missed" directives for use
in the testsuite for matching these (with -fopt-info on stderr; they
don't help for dumpfile output).
The patch also converts the various problem-reporting dump messages
in coverage.c:get_coverage_counts to use MSG_MISSED_OPTIMIZATION
rather than MSG_OPTIMIZED_LOCATIONS, as the docs call out "optimized"
as
"information when an optimization is successfully applied",
whereas "missed" is for
"information about missed optimizations",
and problems with profile data seem to me to fall much more into the
latter category than the former. Doing so requires converting a few
tests from using "-fopt-info" (which is implicitly
"-fopt-info-optimized-optall") to getting the "missed" optimizations.
Changing them to "-fopt-info-missed" added lots of noise from the
vectorizer, so I changed these tests to use "-fopt-info-missed-ipa".
gcc/ChangeLog:
* coverage.c (get_coverage_counts): Convert problem-reporting dump
messages from MSG_OPTIMIZED_LOCATIONS to MSG_MISSED_OPTIMIZATION.
* dumpfile.c (kind_as_string): New function.
(dump_loc): Rather than a hardcoded prefix of "note: ", use
kind_as_string to vary the prefix based on dump_kind.
(selftest::test_capture_of_dump_calls): Update for above.
gcc/testsuite/ChangeLog:
* c-c++-common/unroll-1.c: Update expected output from "note" to
"optimized".
* c-c++-common/unroll-2.c: Likewise.
* c-c++-common/unroll-3.c: Likewise.
* g++.dg/tree-ssa/dom-invalid.C: Update expected output from
dg-message to dg-missed. Convert param from -fopt-info to
-fopt-info-missed-ipa.
* g++.dg/tree-ssa/pr81408.C: Update expected output from
dg-message to dg-missed.
* g++.dg/vect/slp-pr56812.cc: Update expected output from
dg-message to dg-optimized.
* gcc.dg/pr26570.c: Update expected output from dg-message to
dg-missed. Convert param from -fopt-info to
-fopt-info-missed-ipa.
* gcc.dg/pr32773.c: Likewise.
* gcc.dg/tree-ssa/pr19210-1.c: Update expected output from
dg-message to dg-missed.
* gcc.dg/unroll-2.c: Update expected output from dg-message to
dg-optimized.
* gcc.dg/vect/nodump-vect-opt-info-1.c: Likewise. Convert param
from -fopt-info to -fopt-info-vec.
* gfortran.dg/directive_unroll_1.f90: Update expected output from
"note" to "optimized".
* gfortran.dg/directive_unroll_2.f90: Likewise.
* gfortran.dg/directive_unroll_3.f90: Likewise.
* gnat.dg/unroll4.adb: Likewise.
* lib/gcc-dg.exp (dg-optimized): New procedure.
(dg-missed): New procedure.
From-SVN: r264697
As reported in
<https://gcc.gnu.org/ml/gcc-patches/2018-09/msg01684.html>, some
fp-int-convert tests fail after my fix for PR c/87390, in Arm /
AArch64 configurations where _Float16 uses excess precision by
default. The issue is comparisons of the results of a conversion by
assignment (compile-time or run-time) from integer to floating-point
with the original integer value; previously this would compare against
an implicit compile-time conversion to the target type, but now, for
C11 and later, it compares against an implicit compile-time conversion
to a possibly wider evaluation format. This is fixed by adding casts
to the test so that the comparison is with a value converted
explicitly to the target type at compile time, without any use of a
wider evaluation format.
PR c/87390
* gcc.dg/torture/fp-int-convert.h (TEST_I_F_VAL): Convert integer
values explicitly to target type for comparison.
From-SVN: r264696
* config/i386/i386.h (SSE_REGNO): Fix check for FIRST_REX_SSE_REG.
(GET_SSE_REGNO): Rename from SSE_REGNO. Update all uses for rename.
From-SVN: r264695
* config/i386/i386.h (CC_REGNO): Remove FPSR_REGS.
* config/i386/i386.c (ix86_fixed_condition_code_regs): Use
INVALID_REGNUM instead of FPSR_REG.
(ix86_md_asm_adjust): Do not clobber FPSR_REG.
* config/i386/i386.md: Update comment of FP compares.
(fldenv): Do not clobber FPSR_REG.
From-SVN: r264694
Fix a bug in the parser code that decides whether a given name should
be considered exported or not. The function Lex::is_exported_name
(which assumes that its input is a mangled name) was being called on
non-mangled (raw utf-8) names in various places. For the bug in
question this caused an imported package to be registered under the
wrong name. To fix the issue, rename 'Lex::is_exported_name' to
'Lex::is_exported_mangled_name', and add a new 'Lex::is_exported_name'
that works on utf-8 strings.
Fixesgolang/go#27836.
Reviewed-on: https://go-review.googlesource.com/137736
From-SVN: r264690
This patch was part of the original patch we acquired from Honza and Martin.
It simplifies nested vec_merge operations using the same mask.
Self-tests are included.
2018-09-28 Andrew Stubbs <ams@codesourcery.com>
Jan Hubicka <jh@suse.cz>
Martin Jambor <mjambor@suse.cz>
* simplify-rtx.c (simplify_merge_mask): New function.
(simplify_ternary_operation): Use it, also see if VEC_MERGEs with the
same masks are used in op1 or op2.
(test_vec_merge): New function.
(test_vector_ops): Call test_vec_merge.
Co-Authored-By: Jan Hubicka <jh@suse.cz>
Co-Authored-By: Martin Jambor <mjambor@suse.cz>
From-SVN: r264688
