]> git.ipfire.org Git - thirdparty/gcc.git/log
thirdparty/gcc.git
21 months agofortran/intrinsic.texi: Add 'intrinsic' to SIGNAL example
Tobias Burnus [Tue, 17 Oct 2023 17:35:18 +0000 (19:35 +0200)] 
fortran/intrinsic.texi: Add 'intrinsic' to SIGNAL example

gcc/fortran/ChangeLog:

* intrinsic.texi (signal): Add 'intrinsic :: signal, sleep' to
the example to make it safer.

21 months agoInitial Panther Lake Support
Haochen Jiang [Wed, 18 Oct 2023 06:40:45 +0000 (14:40 +0800)] 
Initial Panther Lake Support

gcc/ChangeLog:

* common/config/i386/cpuinfo.h (get_intel_cpu): Add Panther
Lake.
* common/config/i386/i386-common.cc (processor_name):
Ditto.
(processor_alias_table): Ditto.
* common/config/i386/i386-cpuinfo.h (enum processor_types):
Add INTEL_PANTHERLAKE.
* config.gcc: Add -march=pantherlake.
* config/i386/driver-i386.cc (host_detect_local_cpu): Refactor
the if clause. Handle pantherlake.
* config/i386/i386-c.cc (ix86_target_macros_internal):
Handle pantherlake.
* config/i386/i386-options.cc (processor_cost_table): Ditto.
(m_PANTHERLAKE): New.
(m_CORE_HYBRID): Add pantherlake.
* config/i386/i386.h (enum processor_type): Ditto.
* doc/extend.texi: Ditto.
* doc/invoke.texi: Ditto.

gcc/testsuite/ChangeLog:

* g++.target/i386/mv16.C: Ditto.
* gcc.target/i386/funcspec-56.inc: Handle new march.

21 months agox86: Add m_CORE_HYBRID for hybrid clients tuning
Haochen Jiang [Mon, 16 Oct 2023 05:51:02 +0000 (13:51 +0800)] 
x86: Add m_CORE_HYBRID for hybrid clients tuning

gcc/Changelog:

* config/i386/i386-options.cc (m_CORE_HYBRID): New.
* config/i386/x86-tune.def: Replace hybrid client tune to
m_CORE_HYBRID.

21 months agoInitial Clearwater Forest Support
Haochen Jiang [Wed, 18 Oct 2023 06:39:53 +0000 (14:39 +0800)] 
Initial Clearwater Forest Support

gcc/ChangeLog:

* common/config/i386/cpuinfo.h
(get_intel_cpu): Handle Clearwater Forest.
* common/config/i386/i386-common.cc (processor_name):
Add Clearwater Forest.
(processor_alias_table): Ditto.
* common/config/i386/i386-cpuinfo.h (enum processor_types):
Add INTEL_CLEARWATERFOREST.
* config.gcc: Add -march=clearwaterforest.
* config/i386/driver-i386.cc (host_detect_local_cpu): Handle
clearwaterforest.
* config/i386/i386-c.cc (ix86_target_macros_internal): Ditto.
* config/i386/i386-options.cc (processor_cost_table): Ditto.
(m_CLEARWATERFOREST): New.
(m_CORE_ATOM): Add clearwaterforest.
* config/i386/i386.h (enum processor_type): Ditto.
* doc/extend.texi: Ditto.
* doc/invoke.texi: Ditto.

gcc/testsuite/ChangeLog:

* g++.target/i386/mv16.C: Ditto.
* gcc.target/i386/funcspec-56.inc: Handle new march.

21 months agoSupport 32/64-bit vectorization for _Float16 fma related operations.
liuhongt [Mon, 16 Oct 2023 08:22:04 +0000 (16:22 +0800)] 
Support 32/64-bit vectorization for _Float16 fma related operations.

gcc/ChangeLog:

* config/i386/mmx.md (fma<mode>4): New expander.
(fms<mode>4): Ditto.
(fnma<mode>4): Ditto.
(fnms<mode>4): Ditto.
(vec_fmaddsubv4hf4): Ditto.
(vec_fmsubaddv4hf4): Ditto.

gcc/testsuite/ChangeLog:

* gcc.target/i386/part-vect-fmaddsubhf-1.c: New test.
* gcc.target/i386/part-vect-fmahf-1.c: New test.

21 months agoRISC-V: Enable more tests for dynamic LMUL and bug fix[PR111832]
Juzhe-Zhong [Tue, 17 Oct 2023 07:30:39 +0000 (15:30 +0800)] 
RISC-V: Enable more tests for dynamic LMUL and bug fix[PR111832]

Last time, Robin has mentioned that dynamic LMUL will cause ICE in SPEC:

https://gcc.gnu.org/pipermail/gcc-patches/2023-September/629992.html

which is caused by assertion FAIL.

When we enable more currents in rvv.exp with dynamic LMUL, such issue can be
reproduced and has a PR: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111832

Now, we enable more tests in rvv.exp in this patch and fix the bug.

PR target/111832

gcc/ChangeLog:

* config/riscv/riscv-vector-costs.cc (get_biggest_mode): New function.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/rvv.exp: Enable more dynamic tests.

21 months agoDaily bump.
GCC Administrator [Wed, 18 Oct 2023 00:17:58 +0000 (00:17 +0000)] 
Daily bump.

21 months agoaarch64: Put LR save slot first in more cases
Richard Sandiford [Tue, 17 Oct 2023 22:46:33 +0000 (23:46 +0100)] 
aarch64: Put LR save slot first in more cases

Now that the prologue and epilogue code iterates over saved
registers in offset order, we can put the LR save slot first
without compromising LDP/STP formation.

This isn't worthwhile when shadow call stacks are enabled, since the
first two registers are also push/pop candidates, and LR cannot be
popped when shadow call stacks are enabled.  (LR is instead loaded
first and compared against the shadow stack's value.)

But otherwise, it seems better to put the LR save slot first,
to reduce unnecessary variation with the layout for stack clash
protection.

gcc/
* config/aarch64/aarch64.cc (aarch64_layout_frame): Don't make
the position of the LR save slot dependent on stack clash
protection unless shadow call stacks are enabled.

gcc/testsuite/
* gcc.target/aarch64/test_frame_2.c: Expect x30 to come before x19.
* gcc.target/aarch64/test_frame_4.c: Likewise.
* gcc.target/aarch64/test_frame_7.c: Likewise.
* gcc.target/aarch64/test_frame_10.c: Likewise.

21 months agoaarch64: Use vecs to store register save order
Richard Sandiford [Tue, 17 Oct 2023 22:46:33 +0000 (23:46 +0100)] 
aarch64: Use vecs to store register save order

aarch64_save/restore_callee_saves looped over registers in register
number order.  This in turn meant that we could only use LDP and STP
for registers that were consecutive both number-wise and
offset-wise (after unsaved registers are excluded).

This patch instead builds lists of the registers that we've decided to
save, in offset order.  We can then form LDP/STP pairs regardless of
register number order, which in turn means that we can put the LR save
slot first without losing LDP/STP opportunities.

gcc/
* config/aarch64/aarch64.h (aarch64_frame): Add vectors that
store the list saved GPRs, FPRs and predicate registers.
* config/aarch64/aarch64.cc (aarch64_layout_frame): Initialize
the lists of saved registers.  Use them to choose push candidates.
Invalidate pop candidates if we're not going to do a pop.
(aarch64_next_callee_save): Delete.
(aarch64_save_callee_saves): Take a list of registers,
rather than a range.  Make !skip_wb select only write-back
candidates.
(aarch64_expand_prologue): Update calls accordingly.
(aarch64_restore_callee_saves): Take a list of registers,
rather than a range.  Always skip pop candidates.  Also skip
LR if shadow call stacks are enabled.
(aarch64_expand_epilogue): Update calls accordingly.

gcc/testsuite/
* gcc.target/aarch64/sve/pcs/stack_clash_2.c: Expect restores
to happen in offset order.
* gcc.target/aarch64/sve/pcs/stack_clash_2_128.c: Likewise.
* gcc.target/aarch64/sve/pcs/stack_clash_2_256.c: Likewise.
* gcc.target/aarch64/sve/pcs/stack_clash_2_512.c: Likewise.
* gcc.target/aarch64/sve/pcs/stack_clash_2_1024.c: Likewise.
* gcc.target/aarch64/sve/pcs/stack_clash_2_2048.c: Likewise.

21 months agoHandle epilogues that contain jumps
Richard Sandiford [Tue, 17 Oct 2023 22:45:43 +0000 (23:45 +0100)] 
Handle epilogues that contain jumps

The prologue/epilogue pass allows the prologue sequence to contain
jumps.  The sequence is then partitioned into basic blocks using
find_many_sub_basic_blocks.

This patch treats epilogues in a similar way.  Since only one block
might need to be split, the patch (re)introduces a find_sub_basic_blocks
routine to handle a single block.

The new routine hard-codes the assumption that split_block will chain
the new block immediately after the original block.  The routine doesn't
try to replicate the fix for PR81030, since that was specific to
gimple->rtl expansion.

The patch is needed for follow-on aarch64 patches that add conditional
code to the epilogue.  The tests are part of those patches.

gcc/
* cfgbuild.h (find_sub_basic_blocks): Declare.
* cfgbuild.cc (update_profile_for_new_sub_basic_block): New function,
split out from...
(find_many_sub_basic_blocks): ...here.
(find_sub_basic_blocks): New function.
* function.cc (thread_prologue_and_epilogue_insns): Handle
epilogues that contain jumps.

21 months agossa_name_has_boolean_range vs signed-boolean:31 types
Andrew Pinski [Fri, 1 Sep 2023 22:09:08 +0000 (22:09 +0000)] 
ssa_name_has_boolean_range vs signed-boolean:31 types

This turns out to be a latent bug in ssa_name_has_boolean_range
where it would return true for all boolean types but all of the
uses of ssa_name_has_boolean_range was expecting 0/1 as the range
rather than [-1,0].
So when I fixed vector lower to do all comparisons in boolean_type
rather than still in the signed-boolean:31 type (to fix a different issue),
the pattern in match for `-(type)!A -> (type)A - 1.` would assume A (which
was signed-boolean:31) had a range of [0,1] which broke down and sometimes
gave us -1/-2 as values rather than what we were expecting of -1/0.

This was the simpliest patch I found while testing.

We have another way of matching [0,1] range which we could use instead
of ssa_name_has_boolean_range except that uses only the global ranges
rather than the local range (during VRP).
I tried to clean this up slightly by using gimple_match_zero_one_valuedp
inside ssa_name_has_boolean_range but that failed because due to using
only the global ranges. I then tried to change get_nonzero_bits to use
the local ranges at the optimization time but that failed also because
we would remove branches to __builtin_unreachable during evrp and lose
information as we don't set the global ranges during evrp.

OK? Bootstrapped and tested on x86_64-linux-gnu.

PR tree-optimization/110817

gcc/ChangeLog:

* tree-ssanames.cc (ssa_name_has_boolean_range): Remove the
check for boolean type as they don't have "[0,1]" range.

gcc/testsuite/ChangeLog:

* gcc.c-torture/execute/pr110817-1.c: New test.
* gcc.c-torture/execute/pr110817-2.c: New test.
* gcc.c-torture/execute/pr110817-3.c: New test.

21 months agoc++: accepts-invalid with =delete("") [PR111840]
Marek Polacek [Tue, 17 Oct 2023 20:11:11 +0000 (16:11 -0400)] 
c++: accepts-invalid with =delete("") [PR111840]

r6-2367 added a DECL_INITIAL check to cp_parser_simple_declaration
so that we don't emit multiple errors in g++.dg/parse/error57.C.
But that means we don't diagnose

  int f1() = delete("george_crumb");

anymore, because fn decls often have error_mark_node in their
DECL_INITIAL.  (The code may be allowed one day via https://wg21.link/P2573R0.)

I was hoping I could use cp_parser_error_occurred but that would
regress error57.C.

PR c++/111840

gcc/cp/ChangeLog:

* parser.cc (cp_parser_simple_declaration): Do cp_parser_error
for FUNCTION_DECLs.

gcc/testsuite/ChangeLog:

* g++.dg/parse/error65.C: New test.

21 months agoc++: Fix compile-time-hog in cp_fold_immediate_r [PR111660]
Marek Polacek [Thu, 12 Oct 2023 19:58:05 +0000 (15:58 -0400)] 
c++: Fix compile-time-hog in cp_fold_immediate_r [PR111660]

My recent patch introducing cp_fold_immediate_r caused exponential
compile time with nested COND_EXPRs.  The problem is that the COND_EXPR
case recursively walks the arms of a COND_EXPR, but after processing
both arms it doesn't end the walk; it proceeds to walk the
sub-expressions of the outermost COND_EXPR, triggering again walking
the arms of the nested COND_EXPR, and so on.  This patch brings the
compile time down to about 0m0.030s.

The ff_fold_immediate flag is unused after this patch but since I'm
using it in the P2564 patch, I'm not removing it now.  Maybe at_eof
can be used instead and then we can remove ff_fold_immediate.

PR c++/111660

gcc/cp/ChangeLog:

* cp-gimplify.cc (cp_fold_immediate_r) <case COND_EXPR>: Don't
handle it here.
(cp_fold_r): Handle COND_EXPR here.

gcc/testsuite/ChangeLog:

* g++.dg/cpp0x/hog1.C: New test.
* g++.dg/cpp2a/consteval36.C: New test.

21 months agoc++: mangling tweaks
Jason Merrill [Sun, 24 Sep 2023 14:33:00 +0000 (15:33 +0100)] 
c++: mangling tweaks

Most of this is introducing the abi_check function to reduce the verbosity
of most places that check -fabi-version.

The start_mangling change is to avoid needing to zero-initialize additional
members of the mangling globals, though I'm not actually adding any.

The comment documents existing semantics.

gcc/cp/ChangeLog:

* mangle.cc (abi_check): New.
(write_prefix, write_unqualified_name, write_discriminator)
(write_type, write_member_name, write_expression)
(write_template_arg, write_template_param): Use it.
(start_mangling): Assign from {}.
* cp-tree.h: Update comment.

21 months agoc++: Add missing auto_diagnostic_groups to constexpr.cc
Nathaniel Shead [Tue, 17 Oct 2023 10:35:21 +0000 (21:35 +1100)] 
c++: Add missing auto_diagnostic_groups to constexpr.cc

gcc/cp/ChangeLog:

* constexpr.cc (cxx_eval_dynamic_cast_fn): Add missing
auto_diagnostic_group.
(cxx_eval_call_expression): Likewise.
(diag_array_subscript): Likewise.
(outside_lifetime_error): Likewise.
(potential_constant_expression_1): Likewise.

Signed-off-by: Nathaniel Shead <nathanieloshead@gmail.com>
Reviewed-by: Marek Polacek <polacek@redhat.com>
21 months agoRISC-V/testsuite/pr111466.c: update test and expected output
Vineet Gupta [Tue, 17 Oct 2023 17:05:41 +0000 (10:05 -0700)] 
RISC-V/testsuite/pr111466.c: update test and expected output

Update the test to potentially generate two SEXT.W instructions: one for
incoming function arg, other for function return.

But after commit 8eb9cdd14218
("expr: don't clear SUBREG_PROMOTED_VAR_P flag for a promoted subreg")
the test is not supposed to generate either of them so fix the expected
assembler output which was errorneously introduced by commit above.

gcc/testsuite/ChangeLog:
* gcc.target/riscv/pr111466.c (foo2): Change return to unsigned
int as that will potentially generate two SEXT.W instructions.
dg-final: Change to scan-assembler-not SEXT.W.

Signed-off-by: Vineet Gupta <vineetg@rivosinc.com>
21 months agoc: error for function with external and internal linkage [PR111708]
Martin Uecker [Sat, 14 Oct 2023 07:09:07 +0000 (09:09 +0200)] 
c: error for function with external and internal linkage [PR111708]

Declaring a function with both external and internal linkage
in the same TU is translation-time UB.  Add an error for this
case as already done for objects.

PR c/111708

gcc/c/ChangeLog:

* c-decl.cc (grokdeclarator): Add error.

gcc/testsuite/ChangeLog:

* gcc.dg/pr111708-1.c: New test.
* gcc.dg/pr111708-2.c: New test.

21 months agoFortran: out of bounds access with nested implied-do IO [PR111837]
Harald Anlauf [Mon, 16 Oct 2023 19:02:20 +0000 (21:02 +0200)] 
Fortran: out of bounds access with nested implied-do IO [PR111837]

gcc/fortran/ChangeLog:

PR fortran/111837
* frontend-passes.cc (traverse_io_block): Dependency check of loop
nest shall be triangular, not banded.

gcc/testsuite/ChangeLog:

PR fortran/111837
* gfortran.dg/implied_do_io_8.f90: New test.

21 months agofortran/intrinsic.texi: Improve SIGNAL intrinsic entry
Tobias Burnus [Tue, 17 Oct 2023 16:23:09 +0000 (18:23 +0200)] 
fortran/intrinsic.texi: Improve SIGNAL intrinsic entry

gcc/fortran/ChangeLog:

* intrinsic.texi (signal): Mention that the argument
passed to the signal handler procedure is passed by reference.
Extend example.

21 months agoMATCH: [PR111432] Simplify `a & (x | CST)` to a when we know that (a & ~CST) == 0
Andrew Pinski [Fri, 13 Oct 2023 20:27:18 +0000 (13:27 -0700)] 
MATCH: [PR111432] Simplify `a & (x | CST)` to a when we know that (a & ~CST) == 0

This adds the simplification `a & (x | CST)` to a when we know that
`(a & ~CST) == 0`. In a similar fashion as `a & CST` is handle.

I looked into handling `a | (x & CST)` but that I don't see any decent
simplifications happening.

OK? Bootstrapped and tested on x86_linux-gnu with no regressions.

PR tree-optimization/111432

gcc/ChangeLog:

* match.pd (`a & (x | CST)`): New pattern.

gcc/testsuite/ChangeLog:

* gcc.dg/tree-ssa/bitops-7.c: New test.

21 months agoLibF7: Re-generate f7-renames.h to pick up white-space from f7renames.sh.
Georg-Johann Lay [Tue, 17 Oct 2023 15:18:18 +0000 (17:18 +0200)] 
LibF7: Re-generate f7-renames.h to pick up white-space from f7renames.sh.

libgcc/config/avr/libf7/
* f7-renames.h: Re-renerate.

21 months agotree-cfg: Add count information when creating new bb in move_sese_region_to_fn
Andre Vieira [Tue, 17 Oct 2023 14:02:29 +0000 (15:02 +0100)] 
tree-cfg: Add count information when creating new bb in move_sese_region_to_fn

This patch makes sure the profile_count information is initialized for the new
bb created in move_sese_region_to_fn.

gcc/ChangeLog:

* tree-cfg.cc (move_sese_region_to_fn): Initialize profile_count for
new basic block.

21 months agoPR modula2/111756: Re-building all-gcc after source changes fails to link
Gaius Mulley [Tue, 17 Oct 2023 13:11:24 +0000 (14:11 +0100)] 
PR modula2/111756: Re-building all-gcc after source changes fails to link

When having modula-2 enabled in a development tree and there are any
changes that trigger rebuilds in m2/ doing a 'make all-gcc' in the
build directory might fail due to lack of dependency tracking.  This
patch introduces build dependencies into gcc/m2/Make-lang.in using -M*
options.  The patch also introduces all -M* options to cc1gm2 and gm2.

gcc/m2/ChangeLog:

PR modula2/111756
* Make-lang.in (CM2DEP): New define conditionally set if
($(CXXDEPMODE),depmode=gcc3).
(GM2_1): Use $(CM2DEP).
(m2/gm2-gcc/%.o): Ensure $(@D)/$(DEPDIR) is created.
Add $(CM2DEP) to the $(COMPILER) command and use $(POSTCOMPILE).
(m2/gm2-gcc/m2configure.o): Ditto.
(m2/gm2-lang.o): Ditto.
(m2/m2pp.o): Ditto.
(m2/gm2-gcc/rtegraph.o): Ditto.
(m2/mc-boot/$(SRC_PREFIX)%.o): Ditto.
(m2/mc-boot-ch/$(SRC_PREFIX)%.o): Ditto.
(m2/mc-boot-ch/$(SRC_PREFIX)%.o): Ditto.
(m2/mc-boot/main.o): Ditto.
(mcflex.o): Ditto.
(m2/gm2-libs-boot/M2RTS.o): Ditto.
(m2/gm2-libs-boot/%.o): Ditto.
(m2/gm2-libs-boot/%.o): Ditto.
(m2/gm2-libs-boot/RTcodummy.o): Ditto.
(m2/gm2-libs-boot/RTintdummy.o): Ditto.
(m2/gm2-libs-boot/wrapc.o): Ditto.
(m2/gm2-libs-boot/UnixArgs.o): Ditto.
(m2/gm2-libs-boot/choosetemp.o): Ditto.
(m2/gm2-libs-boot/errno.o): Ditto.
(m2/gm2-libs-boot/dtoa.o): Ditto.
(m2/gm2-libs-boot/ldtoa.o): Ditto.
(m2/gm2-libs-boot/termios.o): Ditto.
(m2/gm2-libs-boot/SysExceptions.o): Ditto.
(m2/gm2-libs-boot/SysStorage.o): Ditto.
(m2/gm2-compiler-boot/M2GCCDeclare.o): Ditto.
(m2/gm2-compiler-boot/M2Error.o): Ditto.
(m2/gm2-compiler-boot/%.o): Ditto.
(m2/gm2-compiler-boot/%.o): Ditto.
(m2/gm2-compiler-boot/m2flex.o): Ditto.
(m2/gm2-compiler/%.o): Ditto.
(m2/gm2-compiler/m2flex.o): Ditto.
(m2/gm2-libs-iso/%.o): Ditto.
(m2/gm2-libs/%.o): Ditto.
(m2/gm2-libs/%.o): Ditto.
(m2/gm2-libs/choosetemp.o): Ditto.
(m2/boot-bin/mklink$(exeext)): Ditto.
(m2/pge-boot/%.o): Ditto.
(m2/pge-boot/%.o): Ditto.
(m2/gm2-compiler/%.o): Ensure $(@D)/$(DEPDIR) is created and use
$(POSTCOMPILE).
(m2/gm2-compiler/%.o): Ditto.
(m2/gm2-libs-iso/%.o): Ditto.
(m2/gm2-libs/%.o): Ditto.
* README: Purge out of date info.
* gm2-compiler/M2Comp.mod (MakeSaveTempsFileNameExt): Import.
(OnExitDelete): Import.
(GetModuleDefImportStatementList): Import.
(GetModuleModImportStatementList): Import.
(GetImportModule): Import.
(IsImportStatement): Import.
(IsImport): Import.
(GetImportStatementList): Import.
(File): Import.
(Close): Import.
(EOF): Import.
(IsNoError): Import.
(WriteLine): Import.
(WriteChar): Import.
(FlushOutErr): Import.
(WriteS): Import.
(OpenToRead): Import.
(OpenToWrite): Import.
(ReadS): Import.
(WriteS): Import.
(GetM): Import.
(GetMM): Import.
(GetDepTarget): Import.
(GetMF): Import.
(GetMP): Import.
(GetObj): Import.
(GetMD): Import.
(GetMMD): Import.
(GenerateDefDependency): New procedure.
(GenerateDependenciesFromImport): New procedure.
(GenerateDependenciesFromList): New procedure.
(GenerateDependencies): New procedure.
(Compile): Re-write.
(compile): Re-format.
(CreateFileStem): New procedure function.
(DoPass0): Re-write.
(IsLibrary): New procedure function.
(IsUnique): New procedure function.
(Append): New procedure.
(MergeDep): New procedure.
(GetRuleTarget): New procedure function.
(ReadDepContents): New procedure function.
(WriteDep): New procedure.
(WritePhonyDep): New procedure.
(WriteDepContents): New procedure.
(CreateDepFilename): New procedure function.
(Pass0CheckDef): New procedure function.
(Pass0CheckMod): New procedure function.
(DoPass0): Re-write.
(DepContent): New variable.
(DepOutput): New variable.
(BaseName): New procedure function.
* gm2-compiler/M2GCCDeclare.mod (PrintTerse): Handle IsImport.
Replace IsGnuAsmVolatile with IsGnuAsm.
* gm2-compiler/M2Options.def (EXPORT QUALIFIED): Remove list.
(SetM): New procedure.
(GetM): New procedure function.
(SetMM): New procedure.
(GetMM): New procedure function.
(SetMF): New procedure.
(GetMF): New procedure function.
(SetPPOnly): New procedure.
(GetB): New procedure function.
(SetMD): New procedure.
(GetMD): New procedure function.
(SetMMD): New procedure.
(GetMMD): New procedure function.
(SetMQ): New procedure.
(SetMT): New procedure.
(GetMT): New procedure function.
(GetDepTarget): New procedure function.
(SetMP): New procedure.
(GetMP): New procedure function.
(SetObj): New procedure.
(SetSaveTempsDir): New procedure.
* gm2-compiler/M2Options.mod (SetM): New procedure.
(GetM): New procedure function.
(SetMM): New procedure.
(GetMM): New procedure function.
(SetMF): New procedure.
(GetMF): New procedure function.
(SetPPOnly): New procedure.
(GetB): New procedure function.
(SetMD): New procedure.
(GetMD): New procedure function.
(SetMMD): New procedure.
(GetMMD): New procedure function.
(SetMQ): New procedure.
(SetMT): New procedure.
(GetMT): New procedure function.
(GetDepTarget): New procedure function.
(SetMP): New procedure.
(GetMP): New procedure function.
(SetObj): New procedure.
(SetSaveTempsDir): New procedure.
* gm2-compiler/M2Preprocess.def (PreprocessModule): New parameters
topSource and outputDep.  Re-write.
(MakeSaveTempsFileNameExt): New procedure function.
(OnExitDelete): New procedure function.
* gm2-compiler/M2Preprocess.mod (GetM): Import.
(GetMM): Import.
(OnExitDelete): Add debugging message.
(RemoveFile): Add debugging message.
(BaseName): Remove.
(BuildCommandLineExecute): New procedure function.
* gm2-compiler/M2Search.def (SetDefExtension): Remove unnecessary
spacing.
* gm2-compiler/SymbolTable.mod (GetSymName): Handle ImportSym and
ImportStatementSym.
* gm2-gcc/m2options.h (M2Options_SetMD): New function.
(M2Options_GetMD): New function.
(M2Options_SetMMD): New function.
(M2Options_GetMMD): New function.
(M2Options_SetM): New function.
(M2Options_GetM): New function.
(M2Options_SetMM): New function.
(M2Options_GetMM): New function.
(M2Options_GetMQ): New function.
(M2Options_SetMF): New function.
(M2Options_GetMF): New function.
(M2Options_SetMT): New function.
(M2Options_SetMP): New function.
(M2Options_GetMP): New function.
(M2Options_GetDepTarget): New function.
* gm2-lang.cc (gm2_langhook_init): Correct comment case.
(gm2_langhook_init_options): Add case OPT_M and
OPT_MM.
(gm2_langhook_post_options): Add case OPT_MF, OPT_MT,
OPT_MD and OPT_MMD.
* lang-specs.h (M2CPP): Pass though MF option.
(MDMMD): New define.  Add MDMMD to "@modula-2".

Signed-off-by: Gaius Mulley <gaiusmod2@gmail.com>
21 months agotree-optimization/111846 - put simd-clone-info into SLP tree
Richard Biener [Tue, 17 Oct 2023 11:42:59 +0000 (13:42 +0200)] 
tree-optimization/111846 - put simd-clone-info into SLP tree

The following avoids bogously re-using the simd-clone-info we
currently hang off stmt_info from two different SLP contexts where
a different number of lanes should have chosen a different best
simdclone.

PR tree-optimization/111846
* tree-vectorizer.h (_slp_tree::simd_clone_info): Add.
(SLP_TREE_SIMD_CLONE_INFO): New.
* tree-vect-slp.cc (_slp_tree::_slp_tree): Initialize
SLP_TREE_SIMD_CLONE_INFO.
(_slp_tree::~_slp_tree): Release it.
* tree-vect-stmts.cc (vectorizable_simd_clone_call): Use
SLP_TREE_SIMD_CLONE_INFO or STMT_VINFO_SIMD_CLONE_INFO
dependent on if we're doing SLP.

* gcc.dg/vect/pr111846.c: New testcase.

21 months agowide-int-print: Don't print large numbers hexadecimally for print_dec{,s,u}
Jakub Jelinek [Tue, 17 Oct 2023 12:25:00 +0000 (14:25 +0200)] 
wide-int-print: Don't print large numbers hexadecimally for print_dec{,s,u}

The following patch implements printing of wide_int/widest_int numbers
decimally when asked for that using print_dec{,s,u}, even if they have
precision larger than 64 and get_len () above 1 (right now we printed
them hexadecimally and even negative numbers as huge positive hexadecimal).

In order to avoid the expensive division/modulo by 10^19 twice, once to
estimate how many will be needed and another to actually print it, the
patch prints the 19 digit chunks in reverse order (from least significant
to most significant) and then reorders those with linear complexity to form
the right printed number.
Tested with printing both 256 and 320 bit numbers (first as an example
of even number of 19 digit chunks plus one shorter above it, the second
as an example of odd number of 19 digit chunks plus one shorter above it).

The l * HOST_BITS_PER_WIDE_INT / 3 + 3 estimatition thinking about it now
is one byte too much (one byte for -, one for '\0') and too conservative,
so we could go with l * HOST_BITS_PER_WIDE_INT / 3 + 2 as well, or e.g.
l * HOST_BITS_PER_WIDE_INT * 10 / 33 + 3 as even less conservative
estimation (though more expensive to compute in inline code).
But that l * HOST_BITS_PER_WIDE_INT / 4 + 4; is likely one byte too much
as well, 2 bytes for 0x, one byte for '\0' and where does the 4th one come
from?  Of course all of these assuming HOST_BITS_PER_WIDE_INT is a multiple
of 64...

2023-10-17  Jakub Jelinek  <jakub@redhat.com>

* wide-int-print.h (print_dec_buf_size): For length, divide number
of bits by 3 and add 3 instead of division by 4 and adding 4.
* wide-int-print.cc (print_decs): Remove superfluous ()s.  Don't call
print_hex, instead call print_decu on either negated value after
printing - or on wi itself.
(print_decu): Don't call print_hex, instead print even large numbers
decimally.
(pp_wide_int_large): Assume len from print_dec_buf_size is big enough
even if it returns false.
* pretty-print.h (pp_wide_int): Use print_dec_buf_size to check if
pp_wide_int_large should be used.
* tree-pretty-print.cc (dump_generic_node): Use print_hex_buf_size
to compute needed buffer size.

21 months agoRISC-V: Fix failed testcase when use -cmodel=medany
Lehua Ding [Tue, 17 Oct 2023 09:48:25 +0000 (17:48 +0800)] 
RISC-V: Fix failed testcase when use -cmodel=medany

This little path fix a failed testcase when use -cmodel=medany.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/base/cpymem-1.c: Split check.

21 months agoLibF7: Implement fma / fmal.
Georg-Johann Lay [Tue, 17 Oct 2023 09:36:52 +0000 (11:36 +0200)] 
LibF7: Implement fma / fmal.

libgcc/config/avr/libf7/
* libf7.h (F7_SIZEOF): New macro.
* libf7-asm.sx: Use F7_SIZEOF instead of magic number "10".
(F7MOD_D_fma_, __fma): New module and function.
(fma) [-mdouble=64]: Define as alias for __fma.
(fmal) [-mlong-double=64]: Define as alias for __fma.
* libf7-common.mk (F7_ASM_PARTS): Add D_fma.

21 months agomiddle-end/111818 - failed DECL_NOT_GIMPLE_REG_P setting of volatile
Richard Biener [Mon, 16 Oct 2023 10:50:46 +0000 (12:50 +0200)] 
middle-end/111818 - failed DECL_NOT_GIMPLE_REG_P setting of volatile

The following addresses a missed DECL_NOT_GIMPLE_REG_P setting of
a volatile declared parameter which causes inlining to substitute
a constant parameter into a context where its address is required.

The main issue is in update_address_taken which clears
DECL_NOT_GIMPLE_REG_P from the parameter but fails to rewrite it
because is_gimple_reg returns false for volatiles.  The following
changes maybe_optimize_var to make the 1:1 correspondence between
clearing DECL_NOT_GIMPLE_REG_P of a register typed decl and
actually rewriting it to SSA.

PR middle-end/111818
* tree-ssa.cc (maybe_optimize_var): When clearing
DECL_NOT_GIMPLE_REG_P always rewrite into SSA.

* gcc.dg/torture/pr111818.c: New testcase.

21 months agotree-optimization/111807 - ICE in verify_sra_access_forest
Richard Biener [Mon, 16 Oct 2023 10:54:28 +0000 (12:54 +0200)] 
tree-optimization/111807 - ICE in verify_sra_access_forest

The following addresses build_reconstructed_reference failing to
build references with a different offset than the models and thus
the caller conditional being off.  This manifests when attempting
to build a ref with offset 160 from the model BIT_FIELD_REF <l_4827[9], 8, 0>
onto the same base l_4827 but the models offset being 288.  This
cannot work for any kind of ref I can think of, not just with
BIT_FIELD_REFs.

PR tree-optimization/111807
* tree-sra.cc (build_ref_for_model): Only call
build_reconstructed_reference when the offsets are the same.

* gcc.dg/torture/pr111807.c: New testcase.

21 months agoexpr: don't clear SUBREG_PROMOTED_VAR_P flag for a promoted subreg [target/111466]
Vineet Gupta [Tue, 17 Oct 2023 03:59:09 +0000 (21:59 -0600)] 
expr: don't clear SUBREG_PROMOTED_VAR_P flag for a promoted subreg [target/111466]

RISC-V suffers from extraneous sign extensions, despite/given the ABI
guarantee that 32-bit quantities are sign-extended into 64-bit registers,
meaning incoming SI function args need not be explicitly sign extended
(so do SI return values as most ALU insns implicitly sign-extend too.)

Existing REE doesn't seem to handle this well and there are various ideas
floating around to smarten REE about it.

RISC-V also seems to correctly implement middle-end hook PROMOTE_MODE
etc.

Another approach would be to prevent EXPAND from generating the
sign_extend in the first place which this patch tries to do.

The hunk being removed was introduced way back in 1994 as
   5069803972 ("expand_expr, case CONVERT_EXPR .. clear the promotion flag")

This survived full testsuite run for RISC-V rv64gc with surprisingly no
fallouts: test results before/after are exactly same.

|                               | # of unexpected case / # of unique unexpected case
|                               |          gcc |          g++ |     gfortran |
| rv64imafdc_zba_zbb_zbs_zicond/|  264 /    87 |    5 /     2 |   72 /    12 |
|    lp64d/medlow

Granted for something so old to have survived, there must be a valid
reason. Unfortunately the original change didn't have additional
commentary or a test case. That is not to say it can't/won't possibly
break things on other arches/ABIs, hence the RFC for someone to scream
that this is just bonkers, don't do this ðŸ™‚

I've explicitly CC'ed Jakub and Roger who have last touched subreg
promoted notes in expr.cc for insight and/or screaming ðŸ˜‰

Thanks to Robin for narrowing this down in an amazing debugging session
@ GNU Cauldron.

```
foo2:
sext.w a6,a1             <-- this goes away
beq a1,zero,.L4
li a5,0
li a0,0
.L3:
addw a4,a2,a5
addw a5,a3,a5
addw a0,a4,a0
bltu a5,a6,.L3
ret
.L4:
li a0,0
ret
```

Signed-off-by: Vineet Gupta <vineetg@rivosinc.com>
Co-developed-by: Robin Dapp <rdapp.gcc@gmail.com>
PR target/111466
gcc/
* expr.cc (expand_expr_real_2): Do not clear SUBREG_PROMOTED_VAR_P.

gcc/testsuite
* gcc.target/riscv/pr111466.c: New test.

21 months agoLoongArch: Fix vec_initv32qiv16qi template to avoid ICE.
Chenghui Pan [Wed, 11 Oct 2023 08:41:25 +0000 (16:41 +0800)] 
LoongArch: Fix vec_initv32qiv16qi template to avoid ICE.

Following test code triggers unrecognized insn ICE on LoongArch target
with "-O3 -mlasx":

void
foo (unsigned char *dst, unsigned char *src)
{
  for (int y = 0; y < 16; y++)
    {
      for (int x = 0; x < 16; x++)
        dst[x] = src[x] + 1;
      dst += 32;
      src += 32;
    }
}

ICE info:
./test.c: In function â€˜foo’:
./test.c:8:1: error: unrecognizable insn:
    8 | }
      | ^
(insn 15 14 16 4 (set (reg:V32QI 185 [ vect__24.7 ])
        (vec_concat:V32QI (reg:V16QI 186)
            (const_vector:V16QI [
                    (const_int 0 [0]) repeated x16
                ]))) "./test.c":4:19 -1
     (nil))
during RTL pass: vregs
./test.c:8:1: internal compiler error: in extract_insn, at recog.cc:2791
0x12028023b _fatal_insn(char const*, rtx_def const*, char const*, int, char const*)
        /home/panchenghui/upstream/gcc/gcc/rtl-error.cc:108
0x12028026f _fatal_insn_not_found(rtx_def const*, char const*, int, char const*)
        /home/panchenghui/upstream/gcc/gcc/rtl-error.cc:116
0x120a03c5b extract_insn(rtx_insn*)
        /home/panchenghui/upstream/gcc/gcc/recog.cc:2791
0x12067ff73 instantiate_virtual_regs_in_insn
        /home/panchenghui/upstream/gcc/gcc/function.cc:1610
0x12067ff73 instantiate_virtual_regs
        /home/panchenghui/upstream/gcc/gcc/function.cc:1983
0x12067ff73 execute
        /home/panchenghui/upstream/gcc/gcc/function.cc:2030

This RTL is generated inside loongarch_expand_vector_group_init function (related
to vec_initv32qiv16qi template). Original impl doesn't ensure all vec_concat arguments
are register type. This patch adds force_reg() to the vec_concat argument generation.

gcc/ChangeLog:

* config/loongarch/loongarch.cc (loongarch_expand_vector_group_init):
fix impl related to vec_initv32qiv16qi template to avoid ICE.

gcc/testsuite/ChangeLog:

* gcc.target/loongarch/vector/lasx/lasx-vec-init-1.c: New test.

21 months agoLoongArch: Delete macro definition ASM_OUTPUT_ALIGN_WITH_NOP.
Lulu Cheng [Fri, 15 Sep 2023 03:56:01 +0000 (11:56 +0800)] 
LoongArch: Delete macro definition ASM_OUTPUT_ALIGN_WITH_NOP.

There are two reasons for removing this macro definition:
1. The default in the assembler is to use the nop instruction for filling.
2. For assembly directives: .align [abs-expr[, abs-expr[, abs-expr]]]
   The third expression it is the maximum number of bytes that should be
   skipped by this alignment directive.
   Therefore, it will affect the display of the specified alignment rules
   and affect the operating efficiency.

This modification relies on binutils commit 1fb3cdd87ec61715a5684925fb6d6a6cf53bb97c.
(Since the assembler will add nop based on the .align information when doing relax,
it will cause the conditional branch to go out of bounds during the assembly process.
This submission of binutils solves this problem.)

gcc/ChangeLog:

* config/loongarch/loongarch.h (ASM_OUTPUT_ALIGN_WITH_NOP):
Delete.

Co-authored-by: Chenghua Xu <xuchenghua@loongson.cn>
21 months agoRISC-V: Fix unexpected big LMUL choosing in dynamic LMUL model for non-adjacent load...
Juzhe-Zhong [Mon, 16 Oct 2023 23:20:38 +0000 (07:20 +0800)] 
RISC-V: Fix unexpected big LMUL choosing in dynamic LMUL model for non-adjacent load/store

Consider this following case:
int
bar (int *x, int a, int b, int n)
{
  x = __builtin_assume_aligned (x, __BIGGEST_ALIGNMENT__);
  int sum1 = 0;
  int sum2 = 0;
  for (int i = 0; i < n; ++i)
    {
      sum1 += x[2*i] - a;
      sum1 += x[2*i+1] * b;
      sum2 += x[2*i] - b;
      sum2 += x[2*i+1] * a;
    }
  return sum1 + sum2;
}

Before this patch:

bar:
        ble     a3,zero,.L5
        csrr    t0,vlenb
        csrr    a6,vlenb
        slli    t1,t0,3
        vsetvli a5,zero,e32,m4,ta,ma
        sub     sp,sp,t1
        vid.v   v20
        vmv.v.x v12,a1
        vand.vi v4,v20,1
        vmv.v.x v16,a2
        vmseq.vi        v4,v4,1
        slli    t3,a6,2
        vsetvli zero,a5,e32,m4,ta,ma
        vmv1r.v v0,v4
        viota.m v8,v4
        add     a7,t3,sp
        vsetvli a5,zero,e32,m4,ta,mu
        vand.vi v28,v20,-2
        vadd.vi v4,v28,1
        vs4r.v  v20,0(a7)                        -----  spill
        vrgather.vv     v24,v12,v8
        vrgather.vv     v20,v16,v8
        vrgather.vv     v24,v16,v8,v0.t
        vrgather.vv     v20,v12,v8,v0.t
        vs4r.v  v4,0(sp)                          ----- spill
        slli    a3,a3,1
        addi    t4,a6,-1
        neg     t1,a6
        vmv4r.v v0,v20
        vmv.v.i v4,0
        j       .L4
.L13:
        vsetvli a5,zero,e32,m4,ta,ma
.L4:
        mv      a7,a3
        mv      a4,a3
        bleu    a3,a6,.L3
        csrr    a4,vlenb
.L3:
        vmv.v.x v8,t4
        vl4re32.v       v12,0(sp)                ---- spill
        vand.vv v20,v28,v8
        vand.vv v8,v12,v8
        vsetvli zero,a4,e32,m4,ta,ma
        vle32.v v16,0(a0)
        vsetvli a5,zero,e32,m4,ta,ma
        add     a3,a3,t1
        vrgather.vv     v12,v16,v20
        add     a0,a0,t3
        vrgather.vv     v20,v16,v8
        vsub.vv v12,v12,v0
        vsetvli zero,a4,e32,m4,tu,ma
        vadd.vv v4,v4,v12
        vmacc.vv        v4,v24,v20
        bgtu    a7,a6,.L13
        csrr    a1,vlenb
        slli    a1,a1,2
        add     a1,a1,sp
        li      a4,-1
        csrr    t0,vlenb
        vsetvli a5,zero,e32,m4,ta,ma
        vl4re32.v       v12,0(a1)               ---- spill
        vmv.v.i v8,0
        vmul.vx v0,v12,a4
        li      a2,0
        slli    t1,t0,3
        vadd.vi v0,v0,-1
        vand.vi v0,v0,1
        vmseq.vv        v0,v0,v8
        vand.vi v12,v12,1
        vmerge.vvm      v16,v8,v4,v0
        vmseq.vv        v12,v12,v8
        vmv.s.x v1,a2
        vmv1r.v v0,v12
        vredsum.vs      v16,v16,v1
        vmerge.vvm      v8,v8,v4,v0
        vmv.x.s a0,v16
        vredsum.vs      v8,v8,v1
        vmv.x.s a5,v8
        add     sp,sp,t1
        addw    a0,a0,a5
        jr      ra
.L5:
        li      a0,0
        ret

We can there are multiple horrible register spillings.
The root cause of this issue is for a scalar IR load:

_5 = *_4;

We didn't check whether it is a continguous load/store or gather/scatter load/store

Since it will be translate into:

   1. MASK_LEN_GATHER_LOAD (..., perm indice).
   2. Continguous load/store + VEC_PERM (..., perm indice)

It's obvious that no matter which situation, we will end up with consuming one vector register group (perm indice)
that we didn't count it before.

So this case we pick LMUL = 4 which is incorrect choice for dynamic LMUL cost model.

The key of this patch is:

  if ((type == load_vec_info_type || type == store_vec_info_type)
      && !adjacent_dr_p (STMT_VINFO_DATA_REF (stmt_info)))
    {
               ...
            }

Add one more register consumption if it is not an adjacent load/store.

After this patch, it pick LMUL = 2 which is optimal:

bar:
ble a3,zero,.L4
csrr a6,vlenb
vsetvli a5,zero,e32,m2,ta,ma
vmv.v.x v6,a2
srli a2,a6,1
vmv.v.x v4,a1
vid.v v12
slli a3,a3,1
vand.vi v0,v12,1
addi t1,a2,-1
vmseq.vi v0,v0,1
slli a6,a6,1
vsetvli zero,a5,e32,m2,ta,ma
neg a7,a2
viota.m v2,v0
vsetvli a5,zero,e32,m2,ta,mu
vrgather.vv v16,v4,v2
vrgather.vv v14,v6,v2
vrgather.vv v16,v6,v2,v0.t
vrgather.vv v14,v4,v2,v0.t
vand.vi v18,v12,-2
vmv.v.i v2,0
vadd.vi v20,v18,1
.L3:
minu a4,a3,a2
vsetvli zero,a4,e32,m2,ta,ma
vle32.v v8,0(a0)
vsetvli a5,zero,e32,m2,ta,ma
vmv.v.x v4,t1
vand.vv v10,v18,v4
vrgather.vv v6,v8,v10
vsub.vv v6,v6,v14
vsetvli zero,a4,e32,m2,tu,ma
vadd.vv v2,v2,v6
vsetvli a1,zero,e32,m2,ta,ma
vand.vv v4,v20,v4
vrgather.vv v6,v8,v4
vsetvli zero,a4,e32,m2,tu,ma
mv a4,a3
add a0,a0,a6
add a3,a3,a7
vmacc.vv v2,v16,v6
bgtu a4,a2,.L3
vsetvli a1,zero,e32,m2,ta,ma
vand.vi v0,v12,1
vmv.v.i v4,0
li a3,-1
vmseq.vv v0,v0,v4
vmv.s.x v1,zero
vmerge.vvm v6,v4,v2,v0
vredsum.vs v6,v6,v1
vmul.vx v0,v12,a3
vadd.vi v0,v0,-1
vand.vi v0,v0,1
vmv.x.s a4,v6
vmseq.vv v0,v0,v4
vmv.s.x v1,zero
vmerge.vvm v4,v4,v2,v0
vredsum.vs v4,v4,v1
vmv.x.s a0,v4
addw a0,a0,a4
ret
.L4:
li a0,0
ret

No spillings.

gcc/ChangeLog:

* config/riscv/riscv-vector-costs.cc (max_number_of_live_regs): Fix big LMUL issue.
(get_store_value): New function.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/costmodel/riscv/rvv/dynamic-lmul2-7.c: New test.

21 months agod: Forbid taking the address of an intrinsic with no implementation
Iain Buclaw [Tue, 17 Oct 2023 00:06:34 +0000 (02:06 +0200)] 
d: Forbid taking the address of an intrinsic with no implementation

This code fails to link:

    import core.math;
    real function(real) fn = &sin;

However, when called directly, the D intrinsic `sin()' is expanded by
the front-end into the GCC built-in `__builtin_sin()'.  This has been
fixed to now also expand the function when a reference is taken.

As there are D intrinsics and GCC built-ins that don't have a fallback
implementation, raise an error if taking the address is not possible.

gcc/d/ChangeLog:

* d-tree.h (intrinsic_code): Update define for DEF_D_INTRINSIC.
(maybe_reject_intrinsic): New prototype.
* expr.cc (ExprVisitor::visit (SymOffExp *)): Call
maybe_reject_intrinsic.
* intrinsics.cc (intrinsic_decl): Add fallback field.
(intrinsic_decls): Update define for DEF_D_INTRINSIC.
(maybe_reject_intrinsic): New function.
* intrinsics.def (DEF_D_LIB_BUILTIN): Update.
(DEF_CTFE_BUILTIN): Update.
(INTRINSIC_BSF): Declare as library builtin.
(INTRINSIC_BSR): Likewise.
(INTRINSIC_BT): Likewise.
(INTRINSIC_BSF64): Likewise.
(INTRINSIC_BSR64): Likewise.
(INTRINSIC_BT64): Likewise.
(INTRINSIC_POPCNT32): Likewise.
(INTRINSIC_POPCNT64): Likewise.
(INTRINSIC_ROL): Likewise.
(INTRINSIC_ROL_TIARG): Likewise.
(INTRINSIC_ROR): Likewise.
(INTRINSIC_ROR_TIARG): Likewise.
(INTRINSIC_ADDS): Likewise.
(INTRINSIC_ADDSL): Likewise.
(INTRINSIC_ADDU): Likewise.
(INTRINSIC_ADDUL): Likewise.
(INTRINSIC_SUBS): Likewise.
(INTRINSIC_SUBSL): Likewise.
(INTRINSIC_SUBU): Likewise.
(INTRINSIC_SUBUL): Likewise.
(INTRINSIC_MULS): Likewise.
(INTRINSIC_MULSL): Likewise.
(INTRINSIC_MULU): Likewise.
(INTRINSIC_MULUI): Likewise.
(INTRINSIC_MULUL): Likewise.
(INTRINSIC_NEGS): Likewise.
(INTRINSIC_NEGSL): Likewise.
(INTRINSIC_TOPRECF): Likewise.
(INTRINSIC_TOPREC): Likewise.
(INTRINSIC_TOPRECL): Likewise.

gcc/testsuite/ChangeLog:

* gdc.dg/builtins_reject.d: New test.
* gdc.dg/intrinsics_reject.d: New test.

21 months agoDaily bump.
GCC Administrator [Tue, 17 Oct 2023 00:17:33 +0000 (00:17 +0000)] 
Daily bump.

21 months agoFix minor problem in stack probing
Jeff Law [Mon, 16 Oct 2023 23:14:38 +0000 (17:14 -0600)] 
Fix minor problem in stack probing

probe_stack_range has an assert to capture the possibility that that
expand_binop might not construct its result in the provided target.

We triggered that internally a little while ago.  I'm pretty sure it was in the
testsuite, so no new testcase.  The fix is easy, copy the result into the
proper target when needed.

Bootstrapped and regression tested on x86.

gcc/
* explow.cc (probe_stack_range): Handle case when expand_binop
does not construct its result in the expected location.

21 months agodiagnostics: special-case -fdiagnostics-text-art-charset=ascii for LANG=C
David Malcolm [Mon, 16 Oct 2023 23:04:50 +0000 (19:04 -0400)] 
diagnostics: special-case -fdiagnostics-text-art-charset=ascii for LANG=C

In the LWN discussion of the "ASCII" art in GCC 14
  https://lwn.net/Articles/946733/#Comments
there was some concern about the use of non-ASCII characters in the
output.

Currently -fdiagnostics-text-art-charset defaults to "emoji".
To better handle older terminals by default, this patch special-cases
LANG=C to use -fdiagnostics-text-art-charset=ascii.

gcc/ChangeLog:
* diagnostic.cc (diagnostic_initialize): When LANG=C, update
default for -fdiagnostics-text-art-charset from emoji to ascii.
* doc/invoke.texi (fdiagnostics-text-art-charset): Document the above.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>
21 months agodiagnostics: fix missing initialization of context->extra_output_kind
David Malcolm [Mon, 16 Oct 2023 23:02:34 +0000 (19:02 -0400)] 
diagnostics: fix missing initialization of context->extra_output_kind

gcc/ChangeLog:
* diagnostic.cc (diagnostic_initialize): Ensure
context->extra_output_kind is initialized.

Signed-off-by: David Malcolm <dmalcolm@redhat.com>
21 months agoi386: Allow -mlarge-data-threshold with -mcmodel=large
Uros Bizjak [Mon, 16 Oct 2023 20:22:28 +0000 (22:22 +0200)] 
i386: Allow -mlarge-data-threshold with -mcmodel=large

From: Fangrui Song <maskray@google.com>

When using -mcmodel=medium, large data objects larger than the
-mlarge-data-threshold threshold are placed into large data sections
(.lrodata, .ldata, .lbss and some variants).  GNU ld and ld.lld 17 place
.l* sections into separate output sections.  If small and medium code
model object files are mixed, the .l* sections won't exert relocation
overflow pressure on sections in object files built with -mcmodel=small.

However, when using -mcmodel=large, -mlarge-data-threshold doesn't
apply.  This means that the .rodata/.data/.bss sections may exert
relocation overflow pressure on sections in -mcmodel=small object files.

This patch allows -mcmodel=large to generate .l* sections and drops an
unneeded documentation restriction that the value must be the same.

Link: https://groups.google.com/g/x86-64-abi/c/jnQdJeabxiU
("Large data sections for the large code model")

Signed-off-by: Fangrui Song <maskray@google.com>
gcc/ChangeLog:

* config/i386/i386.cc (ix86_can_inline_p):
Handle CM_LARGE and CM_LARGE_PIC.
(x86_elf_aligned_decl_common): Ditto.
(x86_output_aligned_bss): Ditto.
* config/i386/i386.opt: Update doc for -mlarge-data-threshold=.
* doc/invoke.texi: Update doc for -mlarge-data-threshold=.

gcc/testsuite/ChangeLog:

* gcc.target/i386/large-data.c: New test.

21 months agoRISC-V: NFC: Move scalar block move expansion code into riscv-string.cc
Christoph Müllner [Mon, 16 Oct 2023 19:57:43 +0000 (13:57 -0600)] 
RISC-V: NFC: Move scalar block move expansion code into riscv-string.cc

This just moves a few functions out of riscv.cc into riscv-string.cc in an
attempt to keep riscv.cc manageable.  This was originally Christoph's code and
I'm just pushing it on his behalf.

Full disclosure: I built rv64gc after changing to verify everything still
builds.  Given it was just lifting code from one place to another, I didn't run
the testsuite.

gcc/
* config/riscv/riscv-protos.h (emit_block_move): Remove redundant
prototype.  Improve comment.
* config/riscv/riscv.cc (riscv_block_move_straight): Move from riscv.cc
into riscv-string.cc.
(riscv_adjust_block_mem, riscv_block_move_loop): Likewise.
(riscv_expand_block_move): Likewise.
* config/riscv/riscv-string.cc (riscv_block_move_straight): Add moved
function.
(riscv_adjust_block_mem, riscv_block_move_loop): Likewise.
(riscv_expand_block_move): Likewise.

21 months agoRISC-V/testsuite: add a default march (lacking zfa) to some fp tests
Vineet Gupta [Mon, 16 Oct 2023 19:45:41 +0000 (12:45 -0700)] 
RISC-V/testsuite: add a default march (lacking zfa) to some fp tests

A bunch of FP tests expecting specific FP asm output fail when built
with zfa because different insns are generated. And this happens
because those tests don't have an explicit -march and the default
used to configure gcc could end up with zfa causing the false fails.

Fix that by adding the -march explicitly which doesn't have zfa.

BTW it seems we have some duplication in tests for zfa and non-zfa and
it would have been better if they were consolidated, but oh well.

gcc/testsuite:
* gcc.target/riscv/fle-ieee.c: Updates dg-options with
explicit -march=rv64gc and -march=rv32gc.
* gcc.target/riscv/fle-snan.c: Ditto.
* gcc.target/riscv/fle.c: Ditto.
* gcc.target/riscv/flef-ieee.c: Ditto.
* gcc.target/riscv/flef.c: Ditto.
* gcc.target/riscv/flef-snan.c: Ditto.
* gcc.target/riscv/flt-ieee.c: Ditto.
* gcc.target/riscv/flt-snan.c: Ditto.
* gcc.target/riscv/fltf-ieee.c: Ditto.
* gcc.target/riscv/fltf-snan.c: Ditto.

Signed-off-by: Vineet Gupta <vineetg@rivosinc.com>
21 months agoImplement new RTL optimizations pass: fold-mem-offsets
Manolis Tsamis [Mon, 16 Oct 2023 19:08:12 +0000 (13:08 -0600)] 
Implement new RTL optimizations pass: fold-mem-offsets

This is a new RTL pass that tries to optimize memory offset calculations
by moving them from add immediate instructions to the memory loads/stores.
For example it can transform this:

  addi t4,sp,16
  add  t2,a6,t4
  shl  t3,t2,1
  ld   a2,0(t3)
  addi a2,1
  sd   a2,8(t2)

into the following (one instruction less):

  add  t2,a6,sp
  shl  t3,t2,1
  ld   a2,32(t3)
  addi a2,1
  sd   a2,24(t2)

Although there are places where this is done already, this pass is more
powerful and can handle the more difficult cases that are currently not
optimized. Also, it runs late enough and can optimize away unnecessary
stack pointer calculations.

gcc/ChangeLog:

* Makefile.in: Add fold-mem-offsets.o.
* passes.def: Schedule a new pass.
* tree-pass.h (make_pass_fold_mem_offsets): Declare.
* common.opt: New options.
* doc/invoke.texi: Document new option.
* fold-mem-offsets.cc: New file.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/fold-mem-offsets-1.c: New test.
* gcc.target/riscv/fold-mem-offsets-2.c: New test.
* gcc.target/riscv/fold-mem-offsets-3.c: New test.
* gcc.target/i386/pr52146.c: Adjust expected output.

Signed-off-by: Manolis Tsamis <manolis.tsamis@vrull.eu>
21 months agod: Merge upstream dmd, druntime 4c18eed967, phobos d945686a4.
Iain Buclaw [Sun, 15 Oct 2023 17:09:05 +0000 (19:09 +0200)] 
d: Merge upstream dmd, druntime 4c18eed967, phobos d945686a4.

D front-end changes:

- Import latest fixes to mainline.

D runtime changes:

- Import latest fixes to mainline.

Phobos changes:

- Import latest fixes to mainline.

gcc/d/ChangeLog:

* dmd/MERGE: Merge upstream dmd 4c18eed967.
* d-diagnostic.cc (verrorReport): Update for new front-end interface.
(verrorReportSupplemental): Likewise.
* d-lang.cc (d_init_options): Likewise.
(d_handle_option): Likewise.
(d_post_options): Likewise.
(d_parse_file): Likewise.
* decl.cc (get_symbol_decl): Likewise.

libphobos/ChangeLog:

* libdruntime/MERGE: Merge upstream druntime 4c18eed967.
* src/MERGE: Merge upstream phobos d945686a4.

21 months agoMATCH: Improve `A CMP 0 ? A : -A` set of patterns to use bitwise_equal_p.
Andrew Pinski [Sun, 15 Oct 2023 17:36:56 +0000 (10:36 -0700)] 
MATCH: Improve `A CMP 0 ? A : -A` set of patterns to use bitwise_equal_p.

This improves the `A CMP 0 ? A : -A` set of match patterns to use
bitwise_equal_p which allows an nop cast between signed and unsigned.
This allows catching a few extra cases which were not being caught before.

OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.

gcc/ChangeLog:

PR tree-optimization/101541
* match.pd (A CMP 0 ? A : -A): Improve
using bitwise_equal_p.

gcc/testsuite/ChangeLog:

PR tree-optimization/101541
* gcc.dg/tree-ssa/phi-opt-36.c: New test.
* gcc.dg/tree-ssa/phi-opt-37.c: New test.

21 months ago[PR31531] MATCH: Improve ~a < ~b and ~a < CST, allow a nop cast inbetween ~ and a/b
Andrew Pinski [Sun, 15 Oct 2023 22:18:42 +0000 (15:18 -0700)] 
[PR31531] MATCH: Improve ~a < ~b and ~a < CST, allow a nop cast inbetween ~ and a/b

Currently we able to simplify `~a CMP ~b` to `b CMP a` but we should allow a nop
conversion in between the `~` and the `a` which can show up. A similarly thing should
be done for `~a CMP CST`.

I had originally submitted the `~a CMP CST` case as
https://gcc.gnu.org/pipermail/gcc-patches/2021-November/585088.html;
I noticed we should do the same thing for the `~a CMP ~b` case and combined
it with that one here.

OK? Bootstrapped and tested on x86_64-linux-gnu with no regressions.

PR tree-optimization/31531

gcc/ChangeLog:

* match.pd (~X op ~Y): Allow for an optional nop convert.
(~X op C): Likewise.

gcc/testsuite/ChangeLog:

* gcc.dg/tree-ssa/pr31531-1.c: New test.
* gcc.dg/tree-ssa/pr31531-2.c: New test.

21 months agoc++: improve fold-expr location
Jason Merrill [Wed, 4 Oct 2023 21:20:40 +0000 (17:20 -0400)] 
c++: improve fold-expr location

I want to distinguish between constraint && and fold-expressions there of
written by the user and those implied by template parameter
type-constraints; to that end, let's improve our EXPR_LOCATION for an
explicit fold-expression.

The fold3.C change is needed because this moves the caret from the end of
the expression to the operator, which means the location of the error refers
to the macro invocation rather than the macro definition; both locations are
still printed, but which one is an error and which a note changes.

gcc/cp/ChangeLog:

* parser.cc (cp_parser_fold_expression): Track location range.
* semantics.cc (finish_unary_fold_expr)
(finish_left_unary_fold_expr, finish_right_unary_fold_expr)
(finish_binary_fold_expr): Add location parm.
* constraint.cc (finish_shorthand_constraint): Pass it.
* pt.cc (convert_generic_types_to_packs): Likewise.
* cp-tree.h: Adjust.

gcc/testsuite/ChangeLog:

* g++.dg/concepts/diagnostic3.C: Add expected column.
* g++.dg/cpp1z/fold3.C: Adjust diagnostic lines.

21 months agoc++: fix truncated diagnostic in C++23 [PR111272]
Marek Polacek [Fri, 13 Oct 2023 20:47:47 +0000 (16:47 -0400)] 
c++: fix truncated diagnostic in C++23 [PR111272]

In C++23, since P2448, a constexpr function F that calls a non-constexpr
function N is OK as long as we don't actually call F in a constexpr
context.  So instead of giving an error in maybe_save_constexpr_fundef,
we only give an error when evaluating the call.  Unfortunately, as shown
in this PR, the diagnostic can be truncated:

z.C:10:13: note: 'constexpr Jam::Jam()' is not usable as a 'constexpr' function because:
   10 |   constexpr Jam() { ft(); }
      |             ^~~

...because what?  With this patch, we say:

z.C:10:13: note: 'constexpr Jam::Jam()' is not usable as a 'constexpr' function because:
   10 |   constexpr Jam() { ft(); }
      |             ^~~
z.C:10:23: error: call to non-'constexpr' function 'int Jam::ft()'
   10 |   constexpr Jam() { ft(); }
      |                     ~~^~
z.C:8:7: note: 'int Jam::ft()' declared here
    8 |   int ft() { return 42; }
      |       ^~

Like maybe_save_constexpr_fundef, explain_invalid_constexpr_fn should
also check the body of a constructor, not just the mem-initializer.

PR c++/111272

gcc/cp/ChangeLog:

* constexpr.cc (explain_invalid_constexpr_fn): Also check the body of
a constructor in C++14 and up.

gcc/testsuite/ChangeLog:

* g++.dg/cpp1y/constexpr-diag1.C: New test.

21 months agoARC: Split asl dst,1,src into bset dst,0,src to implement 1<<x.
Roger Sayle [Mon, 16 Oct 2023 12:03:09 +0000 (13:03 +0100)] 
ARC: Split asl dst,1,src into bset dst,0,src to implement 1<<x.

This patch adds a pre-reload splitter to arc.md, to use the bset (set
specific bit instruction) to implement 1<<x (i.e. left shifts of one)
on ARC processors that don't have a barrel shifter.

Currently,

int foo(int x) {
  return 1 << x;
}

when compiled with -O2 -mcpu=em is compiled as a loop:

foo: mov_s   r2,1    ;3
        and.f lp_count,r0, 0x1f
        lpnz    2f
        add r2,r2,r2
        nop
2:      # end single insn loop
        j_s.d   [blink]
        mov_s   r0,r2   ;4

with this patch we instead generate a single instruction:

foo: bset    r0,0,r0
        j_s     [blink]

2023-10-16  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* config/arc/arc.md (*ashlsi3_1): New pre-reload splitter to
use bset dst,0,src to implement 1<<x on !TARGET_BARREL_SHIFTER.

21 months agos390: Fix expander popcountv8hi2_vx
Stefan Schulze Frielinghaus [Mon, 16 Oct 2023 11:39:04 +0000 (13:39 +0200)] 
s390: Fix expander popcountv8hi2_vx

The normal form of a CONST_INT which represents an integer of a mode
with fewer bits than in HOST_WIDE_INT is sign extended.  This even holds
for unsigned integers.

This fixes an ICE during cse1 where we bail out at rtl.h:2297 since
INTVAL (x.first) == sext_hwi (INTVAL (x.first), precision) does not hold.

gcc/ChangeLog:

* config/s390/vector.md (popcountv8hi2_vx): Sign extend each
unsigned vector element.

21 months agoRISC-V: Use VLS modes if the NITERS is known and smaller than VLS mode elements.
Juzhe-Zhong [Mon, 16 Oct 2023 08:25:04 +0000 (16:25 +0800)] 
RISC-V: Use VLS modes if the NITERS is known and smaller than VLS mode elements.

void
foo8 (int64_t *restrict a)
{
  for (int i = 0; i < 16; ++i)
    a[i] = a[i]-16;
}

We use VLS modes instead of VLA modes even it is specified by dynamic LMUL.

gcc/ChangeLog:

* config/riscv/riscv-vector-costs.cc (costs::preferred_new_lmul_p): Use VLS modes.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/costmodel/riscv/rvv/no-dynamic-lmul-1.c: New test.

21 months agouse more get_range_query
Jiufu Guo [Mon, 16 Oct 2023 07:28:52 +0000 (15:28 +0800)] 
use more get_range_query

For "get_global_range_query" SSA_NAME_RANGE_INFO can be queried.
For "get_range_query", it could get more context-aware range info.
And look at the implementation of "get_range_query",  it returns
global range if no local fun info.

So, if not quering for SSA_NAME and not chaning the IL, it would
be ok to use get_range_query to replace get_global_range_query.

gcc/ChangeLog:

* fold-const.cc (expr_not_equal_to): Replace get_global_range_query
by get_range_query.
* gimple-fold.cc (size_must_be_zero_p): Likewise.
* gimple-range-fold.cc (fur_source::fur_source): Likewise.
* gimple-ssa-warn-access.cc (check_nul_terminated_array): Likewise.
* tree-dfa.cc (get_ref_base_and_extent): Likewise.

21 months agoSupport 32/64-bit vectorization for conversion between _Float16 and integer/float.
liuhongt [Wed, 11 Oct 2023 09:14:47 +0000 (17:14 +0800)] 
Support 32/64-bit vectorization for conversion between _Float16 and integer/float.

gcc/ChangeLog:

* config/i386/mmx.md (V2FI_32): New mode iterator
(movd_v2hf_to_sse): Rename to ..
(movd_<mode>_to_sse): .. this.
(movd_v2hf_to_sse_reg): Rename to ..
(movd_<mode>_to_sse_reg): .. this.
(fix<fixunssuffix>_trunc<mode><mmxintvecmodelower>2): New
expander.
(fix<fixunssuffix>_truncv2hfv2si2): Ditto.
(float<floatunssuffix><mmxintvecmodelower><mode>2): Ditto.
(float<floatunssuffix>v2siv2hf2): Ditto.
(extendv2hfv2sf2): Ditto.
(truncv2sfv2hf2): Ditto.
* config/i386/sse.md (*vec_concatv8hf_movss): Rename to ..
(*vec_concat<mode>_movss): .. this.

gcc/testsuite/ChangeLog:

* gcc.target/i386/part-vect-hf-convert-1.c: New test.

21 months agoEnable vectorization for V2HF/V4HF rounding operations and sqrt.
liuhongt [Tue, 10 Oct 2023 06:52:33 +0000 (14:52 +0800)] 
Enable vectorization for V2HF/V4HF rounding operations and sqrt.

gcc/ChangeLog:

* config/i386/i386-expand.cc (ix86_sse_copysign_to_positive):
Handle HFmode.
(ix86_expand_round_sse4): Ditto.
* config/i386/i386.md (roundhf2): New expander.
(lroundhf<mode>2): Ditto.
(lrinthf<mode>2): Ditto.
(l<rounding_insn>hf<mode>2): Ditto.
* config/i386/mmx.md (sqrt<mode>2): Ditto.
(btrunc<mode>2): Ditto.
(nearbyint<mode>2): Ditto.
(rint<mode>2): Ditto.
(lrint<mode><mmxintvecmodelower>2): Ditto.
(floor<mode>2): Ditto.
(lfloor<mode><mmxintvecmodelower>2): Ditto.
(ceil<mode>2): Ditto.
(lceil<mode><mmxintvecmodelower>2): Ditto.
(round<mode>2): Ditto.
(lround<mode><mmxintvecmodelower>2): Ditto.
* config/i386/sse.md (lrint<mode><sseintvecmodelower>2): Ditto.
(lfloor<mode><sseintvecmodelower>2): Ditto.
(lceil<mode><sseintvecmodelower>2): Ditto.
(lround<mode><sseintvecmodelower>2): Ditto.
(sse4_1_round<ssescalarmodesuffix>): Extend to V8HF.
(round<mode>2): Extend to V8HF/V16HF/V32HF.

gcc/testsuite/ChangeLog:

* gcc.target/i386/part-vect-roundhf.c: New test.
* gcc.target/i386/part-vect-sqrtph-1.c: New test.

21 months agoDaily bump.
GCC Administrator [Mon, 16 Oct 2023 00:17:13 +0000 (00:17 +0000)] 
Daily bump.

21 months agolibgomp.texi: Update "Enabling OpenMP" + OpenACC / invoke.texi: -fopenacc/-fopenmp...
Tobias Burnus [Sun, 15 Oct 2023 21:10:17 +0000 (23:10 +0200)] 
libgomp.texi: Update "Enabling OpenMP" + OpenACC / invoke.texi: -fopenacc/-fopenmp update

The OpenACC specification does not mention the '!$ ' sentinel for conditional
compilation and the feature was removed in r11-5572-g1d6f6ac693a860
for PR fortran/98011; update libgomp.texi for this and update a leftover
comment. - Additionally, some other updates are done as well.

libgomp/
* libgomp.texi (Enabling OpenMP): Update for C/C++ attributes;
improve wording especially for Fortran; mention -fopenmp-simd.
(Enabling OpenACC): Minor cleanup; remove conditional compilation
sentinel.

gcc/
* doc/invoke.texi (-fopenacc, -fopenmp, -fopenmp-simd): Use @samp not
@code; document more completely the supported Fortran sentinels.

gcc/fortran
* scanner.cc (skip_free_comments, skip_fixed_comments): Remove
leftover 'OpenACC' from comments about OpenMP's conditional
compilation sentinel.

21 months agolibgomp.texi: Improve "OpenACC Environment Variables"
Tobias Burnus [Sun, 15 Oct 2023 20:36:18 +0000 (22:36 +0200)] 
libgomp.texi: Improve "OpenACC Environment Variables"

None of the ACC_* env vars was documented; in particular, the valid valids
for ACC_DEVICE_TYPE found to be lacking as those are not document in the
OpenACC spec.
GCC_ACC_NOTIFY was removed as I failed to find any traces of it but the
addition to the documentation in commit r6-6185-gcdf6119dad04dd
("libgomp.texi: Updates for OpenACC.").  It seems to be planned as GCC
version of the ACC_NOTIFY env var used by another compiler for offloading
debugging.

libgomp/
* libgomp.texi (ACC_DEVICE_TYPE, ACC_DEVICE_NUM, ACC_PROFLIB):
Actually document what the function does.
(GCC_ACC_NOTIFY): Remove unused env var.

21 months agolibgomp.texi: Use present not future tense
Tobias Burnus [Sun, 15 Oct 2023 20:22:36 +0000 (22:22 +0200)] 
libgomp.texi: Use present not future tense

libgomp/ChangeLog:

* libgomp.texi: Replace most future tense by present tense.

21 months agosim: add distclean dep for gnulib
Mike Frysinger [Sun, 15 Oct 2023 15:45:43 +0000 (21:30 +0545)] 
sim: add distclean dep for gnulib

ChangeLog:

* Makefile.def: Add distclean-sim dependency on distclean-gnulib.
* Makefile.in: Regenerate.

21 months agomiddle-end: Improved RTL expansion of 1LL << x.
Roger Sayle [Sun, 15 Oct 2023 13:53:22 +0000 (14:53 +0100)] 
middle-end: Improved RTL expansion of 1LL << x.

This patch improves the initial RTL expanded for double word shifts
on architectures with conditional moves, so that later passes don't
need to clean-up unnecessary and/or unused instructions.

Consider the general case, x << y, which is expanded well as:

t1 = y & 32;
t2 = 0;
t3 = x_lo >> 1;
t4 = y ^ ~0;
t5 = t3 >> t4;
tmp_hi = x_hi << y;
tmp_hi |= t5;
tmp_lo = x_lo << y;
out_hi = t1 ? tmp_lo : tmp_hi;
out_lo = t1 ? t2 : tmp_lo;

which is nearly optimal, the only thing that can be improved is
that using a unary NOT operation "t4 = ~y" is better than XOR
with -1, on targets that support it.  [Note the one_cmpl_optab
expander didn't fall back to XOR when this code was originally
written, but has been improved since].

Now consider the relatively common idiom of 1LL << y, which
currently produces the RTL equivalent of:

t1 = y & 32;
t2 = 0;
t3 = 1 >> 1;
t4 = y ^ ~0;
t5 = t3 >> t4;
tmp_hi = 0 << y;
tmp_hi |= t5;
tmp_lo = 1 << y;
out_hi = t1 ? tmp_lo : tmp_hi;
out_lo = t1 ? t2 : tmp_lo;

Notice here that t3 is always zero, so the assignment of t5
is a variable shift of zero, which expands to a loop on many
smaller targets, a similar shift by zero in the first tmp_hi
assignment (another loop), that the value of t4 is no longer
required (as t3 is zero), and that the ultimate value of tmp_hi
is always zero.

Fortunately, for many (but perhaps not all) targets this mess
gets cleaned up by later optimization passes.  However, this
patch avoids generating unnecessary RTL at expand time, by
calling simplify_expand_binop instead of expand_binop, and
avoiding generating dead or unnecessary code when intermediate
values are known to be zero.  For the 1LL << y test case above,
we now generate:

t1 = y & 32;
t2 = 0;
tmp_hi = 0;
tmp_lo = 1 << y;
out_hi = t1 ? tmp_lo : tmp_hi;
out_lo = t1 ? t2 : tmp_lo;

On arc-elf, for example, there are 18 RTL INSN_P instructions
generated by expand before this patch, but only 12 with this patch
(improving both compile-time and memory usage).

2023-10-15  Roger Sayle  <roger@nextmovesoftware.com>

gcc/ChangeLog
* optabs.cc (expand_subword_shift): Call simplify_expand_binop
instead of expand_binop.  Optimize cases (i.e. avoid generating
RTL) when CARRIES or INTO_INPUT is zero.  Use one_cmpl_optab
(i.e. NOT) instead of xor_optab with ~0 to calculate ~OP1.

21 months agomodula2: Add m2.etags rule to gcc/m2/Make-lang.in
Gaius Mulley [Sun, 15 Oct 2023 13:01:51 +0000 (14:01 +0100)] 
modula2: Add m2.etags rule to gcc/m2/Make-lang.in

This patch adds the m2.etags rule to gcc/m2/Make-lang.in which
generates etags for the .cc .c .h files within gcc/m2.

gcc/m2/ChangeLog:

* Make-lang.in (m2.tags): New rule.

Signed-off-by: Gaius Mulley <gaiusmod2@gmail.com>
21 months agowide-int: Fix estimation of buffer sizes for wide_int printing [PR111800]
Jakub Jelinek [Sun, 15 Oct 2023 12:23:14 +0000 (14:23 +0200)] 
wide-int: Fix estimation of buffer sizes for wide_int printing [PR111800]

As mentioned in the PR, my estimations on needed buffer size for wide_int
and especially widest_int printing were incorrect, I've used get_len ()
in the estimations, but that is true only for !wi::neg_p (x) values.
Under the hood, we have 3 ways to print numbers.
print_decs which if
  if ((wi.get_precision () <= HOST_BITS_PER_WIDE_INT)
      || (wi.get_len () == 1))
uses sprintf which always fits into WIDE_INT_PRINT_BUFFER_SIZE (positive or
negative) and otherwise uses print_hex,
print_decu which if
  if ((wi.get_precision () <= HOST_BITS_PER_WIDE_INT)
      || (wi.get_len () == 1 && !wi::neg_p (wi)))
uses sprintf which always fits into WIDE_INT_PRINT_BUFFER_SIZE (positive
only) and print_hex, which doesn't print most significant limbs which are
zero and the first limb which is non-zero prints such that redundant 0
hex digits aren't printed, while all limbs below that are printed with
"%016" PRIx64.  For wi::neg_p (x) values, the first limb of the precision
is always non-zero, so we print all the limbs for the precision.
So, the current estimations are accurate if !wi::neg_p (x), or when
print_decs will be used and x.get_len () == 1, otherwise we need to use
estimation based on get_precision () rather than get_len ().

I've introduced new inlines print_{dec{,s,u},hex}_buf_size which compute the
needed buffer length in bytes and return true if WIDE_INT_PRINT_BUFFER_SIZE
isn't sufficient and caller should XALLOCAVEC the buffer.

2023-10-15  Jakub Jelinek  <jakub@redhat.com>

PR tree-optimization/111800
gcc/
* wide-int-print.h (print_dec_buf_size, print_decs_buf_size,
print_decu_buf_size, print_hex_buf_size): New inline functions.
* wide-int.cc (assert_deceq): Use print_dec_buf_size.
(assert_hexeq): Use print_hex_buf_size.
* wide-int-print.cc (print_decs): Use print_decs_buf_size.
(print_decu): Use print_decu_buf_size.
(print_hex): Use print_hex_buf_size.
(pp_wide_int_large): Use print_dec_buf_size.
* value-range.cc (irange_bitmask::dump): Use print_hex_buf_size.
* value-range-pretty-print.cc (vrange_printer::print_irange_bitmasks):
Likewise.
* tree-ssa-loop-niter.cc (do_warn_aggressive_loop_optimizations): Use
print_dec_buf_size.  Use TYPE_SIGN macro in print_dec call argument.
gcc/c-family/
* c-warn.cc (match_case_to_enum_1): Assert w.get_precision ()
is smaller or equal to WIDE_INT_MAX_INL_PRECISION rather than
w.get_len () is smaller or equal to WIDE_INT_MAX_INL_ELTS.

21 months agod: Merge upstream dmd, druntime f9efc98fd7, phobos a3f22129d.
Iain Buclaw [Sun, 15 Oct 2023 10:05:10 +0000 (12:05 +0200)] 
d: Merge upstream dmd, druntime f9efc98fd7, phobos a3f22129d.

D front-end changes:

- Import dmd v2.105.2.
- A function with enum storage class is now deprecated.
- Global variables can now be initialized with Associative
  Arrays.
- Improvements for the C++ header generation of static variables
  used in a default argument context.

D runtime changes:

- Import druntime v2.105.2.
- The `core.memory.GC' functions `GC.enable', `GC.disable',
  `GC.collect', and `GC.minimize' `have been marked `@safe'.

Phobos changes:

- Import phobos v2.105.2.

gcc/d/ChangeLog:

* dmd/MERGE: Merge upstream dmd f9efc98fd7.
* dmd/VERSION: Bump version to v2.105.2.
* d-builtins.cc (build_frontend_type): Update for new front-end
interface.
* d-diagnostic.cc (verrorReport): Don't emit tips when error gagging
is turned on.
* d-lang.cc (d_handle_option): Remove obsolete parameter.
(d_post_options): Likewise.
(d_read_ddoc_files): New function.
(d_generate_ddoc_file): New function.
(d_parse_file): Update for new front-end interface.
* expr.cc (ExprVisitor::visit (AssocArrayLiteralExp *)): Check for new
front-end lowering of static associative arrays.

libphobos/ChangeLog:

* libdruntime/MERGE: Merge upstream druntime f9efc98fd7.
* libdruntime/Makefile.am (DRUNTIME_DSOURCES): Add
core/internal/newaa.d.
* libdruntime/Makefile.in: Regenerate.
* src/MERGE: Merge upstream phobos a3f22129d.
* testsuite/libphobos.hash/test_hash.d: Update test.
* testsuite/libphobos.phobos/phobos.exp: Add compiler flags
-Wno-deprecated.
* testsuite/libphobos.phobos_shared/phobos_shared.exp: Likewise.

gcc/testsuite/ChangeLog:

* lib/gdc-utils.exp (gdc-convert-args): Handle new compiler options.

21 months agocombine: Fix handling of unsigned constants
Stefan Schulze Frielinghaus [Sun, 15 Oct 2023 09:20:01 +0000 (11:20 +0200)] 
combine: Fix handling of unsigned constants

If a CONST_INT represents an integer of a mode with fewer bits than in
HOST_WIDE_INT, then the integer is sign extended.  For those two
optimizations touched by this patch, the integers of interest have only
the most significant bit set w.r.t their mode, therefore, they were sign
extended.  Thus in order to get the integer of interest, we have to chop
off the high bits.

gcc/ChangeLog:

* combine.cc (simplify_compare_const): Fix handling of unsigned
constants.

21 months agoRISC-V: Fix vsingle attribute
Juzhe-Zhong [Sun, 15 Oct 2023 03:08:31 +0000 (11:08 +0800)] 
RISC-V: Fix vsingle attribute

RVVM2x2QI should be rvvm2qi instead of rvvmq1i.

gcc/ChangeLog:

* config/riscv/vector-iterators.md: Fix vsingle incorrect attribute for RVVM2x2QI.

21 months agoDaily bump.
GCC Administrator [Sun, 15 Oct 2023 00:17:29 +0000 (00:17 +0000)] 
Daily bump.

21 months agolibgomp.fortran/allocate-6.f90: Run with -fdump-tree-gimple
Tobias Burnus [Sat, 14 Oct 2023 18:09:34 +0000 (20:09 +0200)] 
libgomp.fortran/allocate-6.f90: Run with -fdump-tree-gimple

libgomp/
* testsuite/libgomp.fortran/allocate-6.f90: Add missing
dg-additional-options "-fdump-tree-gimple"; fix scan.

21 months agoFix ICE in set_cell_span, at text-art/table.cc:148 with D front-end and -fanalyzer
Iain Buclaw [Sat, 14 Oct 2023 00:19:41 +0000 (02:19 +0200)] 
Fix ICE in set_cell_span, at text-art/table.cc:148 with D front-end and -fanalyzer

The internal error in analyzer turned out to be caused by a subtly
invalid tree representation of STRING_CSTs in the D front-end, fixed by
including the terminating NULL as part of the TREE_STRING_POINTER.

When adding a first analyzer test for D, it flagged up another subtle
mismatch in one assignment in the module support routines as well, fixed
by generating the correct field type for the compiler-generated struct.

PR d/111537

gcc/d/ChangeLog:

* expr.cc (ExprVisitor::visit (StringExp *)): Include null terminator
in STRING_CST string.
* modules.cc (get_compiler_dso_type): Generate ModuleInfo** type for
the minfo fields.

gcc/testsuite/ChangeLog:

* gdc.dg/analyzer/analyzer.exp: New test.
* gdc.dg/analyzer/pr111537.d: New test.

21 months agod: Reduce code duplication of writing generated files.
Iain Buclaw [Sun, 1 Oct 2023 22:12:25 +0000 (00:12 +0200)] 
d: Reduce code duplication of writing generated files.

Small refactoring ahead of the next merge from upstream, where a few
more front-end routines will stop doing the file handling themselves.

gcc/d/ChangeLog:

* d-lang.cc (d_write_file): New function.
(d_parse_file): Reduce code duplication.

21 months agolibgomp.texi: Note to 'Memory allocation' sect and missing mem-memory routines
Tobias Burnus [Sat, 14 Oct 2023 09:09:50 +0000 (11:09 +0200)] 
libgomp.texi: Note to 'Memory allocation' sect and missing mem-memory routines

This commit completes the documentation of the OpenMP memory-management
routines, except for the unimplemented TR11 additions.  It also makes clear
in the 'Memory allocation' section of the 'OpenMP-Implementation Specifics'
chapter under which condition OpenMP managed memory/allocators are used.

libgomp/ChangeLog:

* libgomp.texi: Fix some typos.
(Memory Management Routines): Document remaining 5.x routines.
(Memory allocation): Make clear when the section applies.

21 months agoFortran: Support OpenMP's 'allocate' directive for stack vars
Tobias Burnus [Sat, 14 Oct 2023 09:07:47 +0000 (11:07 +0200)] 
Fortran: Support OpenMP's 'allocate' directive for stack vars

gcc/fortran/ChangeLog:

* gfortran.h (ext_attr_t): Add omp_allocate flag.
* match.cc (gfc_free_omp_namelist): Void deleting same
u2.allocator multiple times now that a sequence can use
the same one.
* openmp.cc (gfc_match_omp_clauses, gfc_match_omp_allocate): Use
same allocator expr multiple times.
(is_predefined_allocator): Make static.
(gfc_resolve_omp_allocate): Update/extend restriction checks;
remove sorry message.
(resolve_omp_clauses): Reject corarrays in allocate/allocators
directive.
* parse.cc (check_omp_allocate_stmt): Permit procedure pointers
here (rejected later) for less misleading diagnostic.
* trans-array.cc (gfc_trans_auto_array_allocation): Propagate
size for GOMP_alloc and location to which it should be added to.
* trans-decl.cc (gfc_trans_deferred_vars): Handle 'omp allocate'
for stack variables; sorry for static variables/common blocks.
* trans-openmp.cc (gfc_trans_omp_clauses): Evaluate 'allocate'
clause's allocator only once; fix adding expressions to the
block.
(gfc_trans_omp_single): Pass a block to gfc_trans_omp_clauses.

gcc/ChangeLog:

* gimplify.cc (gimplify_bind_expr): Handle Fortran's
'omp allocate' for stack variables.

libgomp/ChangeLog:

* libgomp.texi (OpenMP Impl. Status): Mention that Fortran now
supports the allocate directive for stack variables.
* testsuite/libgomp.fortran/allocate-5.f90: New test.
* testsuite/libgomp.fortran/allocate-6.f90: New test.
* testsuite/libgomp.fortran/allocate-7.f90: New test.
* testsuite/libgomp.fortran/allocate-8.f90: New test.

gcc/testsuite/ChangeLog:

* c-c++-common/gomp/allocate-14.c: Fix directive name.
* c-c++-common/gomp/allocate-15.c: Likewise.
* c-c++-common/gomp/allocate-9.c: Fix comment typo.
* gfortran.dg/gomp/allocate-4.f90: Remove sorry dg-error.
* gfortran.dg/gomp/allocate-7.f90: Likewise.
* gfortran.dg/gomp/allocate-10.f90: New test.
* gfortran.dg/gomp/allocate-11.f90: New test.
* gfortran.dg/gomp/allocate-12.f90: New test.
* gfortran.dg/gomp/allocate-13.f90: New test.
* gfortran.dg/gomp/allocate-14.f90: New test.
* gfortran.dg/gomp/allocate-15.f90: New test.
* gfortran.dg/gomp/allocate-8.f90: New test.
* gfortran.dg/gomp/allocate-9.f90: New test.

21 months agomiddle-end: Allow _BitInt(65535) [PR102989]
Jakub Jelinek [Sat, 14 Oct 2023 07:35:44 +0000 (09:35 +0200)] 
middle-end: Allow _BitInt(65535) [PR102989]

The following patch lifts further restrictions which limited _BitInt to at
most 16319 bits up to 65535.
The problem was mainly in INTEGER_CST representation, which had 3
unsigned char members to describe lengths in number of 64-bit limbs, which
it wanted to fit into 32 bits.  This patch removes the third one which was
just a cache to save a few compile time cycles for wi::to_offset and
enlarges the other two members to unsigned short.
Furthermore, the same problem has been in some uses of trailing_wide_int*
(in value-range-storage*) and value-range-storage* itself, while other
uses of trailing_wide_int* have been fine (e.g. CONST_POLY_INT, where no
constants will be larger than 3/5/9/11 limbs depending on target, so 255
limit is plenty).  The patch turns all those length representations to be
unsigned short for consistency, so value-range-storage* can handle even
16320-65535 bits BITINT_TYPE ranges.  The cc1plus growth is about 16K,
so not really significant for 38M .text section.

Note, the reason for the new limit is
  unsigned int precision : 16;
TYPE_PRECISION limit, if we wanted to overcome that, TYPE_PRECISION would
need to use some other member for BITINT_TYPE from all the others and
we could reach that way 4194239 limit (65535 * 64 - 1, again implied by
INTEGER_CST and value-range-storage*).  Dunno if that is
worth it or if it is something we want to do for GCC 14 though.

2023-10-14  Jakub Jelinek  <jakub@redhat.com>

PR c/102989
gcc/
* tree-core.h (struct tree_base): Remove int_length.offset
member, change type of int_length.unextended and int_length.extended
from unsigned char to unsigned short.
* tree.h (TREE_INT_CST_OFFSET_NUNITS): Remove.
(wi::extended_tree <N>::get_len): Don't use TREE_INT_CST_OFFSET_NUNITS,
instead compute it at runtime from TREE_INT_CST_EXT_NUNITS and
TREE_INT_CST_NUNITS.
* tree.cc (wide_int_to_tree_1): Don't assert
TREE_INT_CST_OFFSET_NUNITS value.
(make_int_cst): Don't initialize TREE_INT_CST_OFFSET_NUNITS.
* wide-int.h (WIDE_INT_MAX_ELTS): Change from 255 to 1024.
(WIDEST_INT_MAX_ELTS): Change from 510 to 2048, adjust comment.
(trailing_wide_int_storage): Change m_len type from unsigned char *
to unsigned short *.
(trailing_wide_int_storage::trailing_wide_int_storage): Change second
argument from unsigned char * to unsigned short *.
(trailing_wide_ints): Change m_max_len type from unsigned char to
unsigned short.  Change m_len element type from
struct{unsigned char len;} to unsigned short.
(trailing_wide_ints <N>::operator []): Remove .len from m_len
accesses.
* value-range-storage.h (irange_storage::lengths_address): Change
return type from const unsigned char * to const unsigned short *.
(irange_storage::write_lengths_address): Change return type from
unsigned char * to unsigned short *.
* value-range-storage.cc (irange_storage::write_lengths_address):
Likewise.
(irange_storage::lengths_address): Change return type from
const unsigned char * to const unsigned short *.
(write_wide_int): Change len argument type from unsigned char *&
to unsigned short *&.
(irange_storage::set_irange): Change len variable type from
unsigned char * to unsigned short *.
(read_wide_int): Change len argument type from unsigned char to
unsigned short.  Use trailing_wide_int_storage <unsigned short>
instead of trailing_wide_int_storage and
trailing_wide_int <unsigned short> instead of trailing_wide_int.
(irange_storage::get_irange): Change len variable type from
unsigned char * to unsigned short *.
(irange_storage::size): Multiply n by sizeof (unsigned short)
in len_size variable initialization.
(irange_storage::dump): Change len variable type from
unsigned char * to unsigned short *.
gcc/cp/
* module.cc (trees_out::start, trees_in::start): Remove
TREE_INT_CST_OFFSET_NUNITS handling.
gcc/testsuite/
* gcc.dg/bitint-38.c: Change into dg-do run test, in addition
to checking the addition, division and right shift results at compile
time check it also at runtime.
* gcc.dg/bitint-39.c: New test.

21 months agoRISC-V: Remove redundant iterators.
Juzhe-Zhong [Sat, 14 Oct 2023 03:06:02 +0000 (11:06 +0800)] 
RISC-V: Remove redundant iterators.

These iterators are redundant, removed and commmitted.
gcc/ChangeLog:

* config/riscv/vector-iterators.md: Remove redundant iterators.

21 months agoDaily bump.
GCC Administrator [Sat, 14 Oct 2023 00:16:40 +0000 (00:16 +0000)] 
Daily bump.

21 months agoFortran: name conflict between internal procedure and derived type [PR104351]
Harald Anlauf [Wed, 11 Oct 2023 19:29:35 +0000 (21:29 +0200)] 
Fortran: name conflict between internal procedure and derived type [PR104351]

gcc/fortran/ChangeLog:

PR fortran/104351
* decl.cc (get_proc_name): Extend name conflict detection between
internal procedure and previous declaration also to derived type.

gcc/testsuite/ChangeLog:

PR fortran/104351
* gfortran.dg/derived_function_interface_1.f90: Adjust pattern.
* gfortran.dg/pr104351.f90: New test.

21 months agofortran: fix handling of options -ffpe-trap and -ffpe-summary [PR110957]
Harald Anlauf [Fri, 6 Oct 2023 20:21:56 +0000 (22:21 +0200)] 
fortran: fix handling of options -ffpe-trap and -ffpe-summary [PR110957]

gcc/fortran/ChangeLog:

PR fortran/110957
* invoke.texi: Update documentation to reflect '-ffpe-trap=none'.
* options.cc (gfc_handle_fpe_option): Fix mixup up of error messages
for options -ffpe-trap and -ffpe-summary.  Accept '-ffpe-trap=none'
to clear FPU traps previously set on command line.

21 months agoDo not add partial equivalences with no uses.
Andrew MacLeod [Thu, 12 Oct 2023 21:06:36 +0000 (17:06 -0400)] 
Do not add partial equivalences with no uses.

PR tree-optimization/111622
* value-relation.cc (equiv_oracle::add_partial_equiv): Do not
register a partial equivalence if an operand has no uses.

21 months agoOMP SIMD inbranch call vectorization for AVX512 style masks
Richard Biener [Fri, 13 Oct 2023 10:32:51 +0000 (12:32 +0200)] 
OMP SIMD inbranch call vectorization for AVX512 style masks

The following teaches vectorizable_simd_clone_call to handle
integer mode masks.  The tricky bit is to second-guess the
number of lanes represented by a single mask argument - the following
uses simdlen and the number of mask arguments to calculate that,
assuming ABIs have them uniform.

Similar to the VOIDmode handling there's a restriction on not
supporting splitting/merging of incoming vector masks to
more/less SIMD call arguments.

PR tree-optimization/111795
* tree-vect-stmts.cc (vectorizable_simd_clone_call): Handle
integer mode mask arguments.

* gcc.target/i386/vect-simd-clone-avx512-1.c: New testcase.
* gcc.target/i386/vect-simd-clone-avx512-2.c: Likewise.
* gcc.target/i386/vect-simd-clone-avx512-3.c: Likewise.

21 months agoAdd support for SLP vectorization of OpenMP SIMD clone calls
Richard Biener [Thu, 12 Oct 2023 12:25:07 +0000 (14:25 +0200)] 
Add support for SLP vectorization of OpenMP SIMD clone calls

This adds support for SLP vectorization of OpenMP SIMD clone calls.
There's a complication when vectorizing calls involving virtual
operands since this is now for the first time not only leafs (loads
or stores).  With SLP this runs into the issue that placement of
the vectorized stmts is not necessarily at one of the original
scalar stmts which leads to the magic updating virtual operands
in vect_finish_stmt_generation not working.  So we run into the
assert that updating virtual operands isn't necessary.  I've
papered over this similar to how we do for mismatched const/pure
attribution by setting vinfo->any_known_not_updated_vssa.

I've added two basic testcases with multi-lane SLP and verified
that with single-lane SLP enabled the rest of the existing testcases
pass.

* tree-vect-slp.cc (mask_call_maps): New.
(vect_get_operand_map): Handle IFN_MASK_CALL.
(vect_build_slp_tree_1): Likewise.
* tree-vect-stmts.cc (vectorizable_simd_clone_call): Handle
SLP.

* gcc.dg/vect/slp-simd-clone-1.c: New testcase.
* gcc.dg/vect/slp-simd-clone-2.c: Likewise.

21 months agoRISC-V Regression: Fix FAIL of bb-slp-68.c for RVV
Juzhe-Zhong [Fri, 13 Oct 2023 06:01:26 +0000 (14:01 +0800)] 
RISC-V Regression: Fix FAIL of bb-slp-68.c for RVV

Like comment said, this test failed on 64 bytes vector.
Both RVV and GCN has 64 bytes vector.

So it's more reasonable to use vect512.
gcc/testsuite/ChangeLog:

* gcc.dg/vect/bb-slp-68.c: Use vect512.

21 months agoRISC-V: Refine run test cases of math autovec
Pan Li [Wed, 11 Oct 2023 12:08:52 +0000 (20:08 +0800)] 
RISC-V: Refine run test cases of math autovec

For the run test cases of math autovec, we need a reference value to
check if the return value is expected or not.

The previous patch leverage hardcode for the reference value but we
can leverage the scalar math function instead. For example ceil after
autovec.

ASSERT (CEIL (Vector {1.2,...}) == Vector {2.0, ...});

But we can leverage the scalar math function to avoid potential mistakes.

ASSERT (CEIL (Vector {1.2,...}) == Vector {ceil (1.2), ...});

This patch remove some fflags check as it covered by check-body already.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/unop/math-ceil-run-1.c:
Use scalar func as reference instead of hardcode.
* gcc.target/riscv/rvv/autovec/unop/math-ceil-run-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-floor-run-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-floor-run-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-nearbyint-run-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-nearbyint-run-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-rint-run-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-rint-run-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-round-run-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-round-run-2.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-trunc-run-1.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/math-trunc-run-2.c: Ditto.

Signed-off-by: Pan Li <pan2.li@intel.com>
21 months agoRISC-V: Add test for FP llfloor auto vectorization
Pan Li [Fri, 13 Oct 2023 09:48:25 +0000 (17:48 +0800)] 
RISC-V: Add test for FP llfloor auto vectorization

The below FP API are supported already by sharing the same standard
name, as well as the machine mode.

long long llfloor (double);

This patch would like to add the test cases for ensuring the
correctness.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/unop/math-llfloor-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-llfloor-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/vls/math-llfloor-0.c: New test.

Signed-off-by: Pan Li <pan2.li@intel.com>
21 months agoRISC-V: Add test for FP ifloor auto vectorization
Pan Li [Fri, 13 Oct 2023 08:20:23 +0000 (16:20 +0800)] 
RISC-V: Add test for FP ifloor auto vectorization

The below FP API are supported already by sharing the same standard
name, as well as the machine mode.

int ifloor (float);

This patch would like to add the test cases for ensuring the
correctness.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/unop/math-ifloor-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-ifloor-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/vls/math-ifloor-0.c: New test.

Signed-off-by: Pan Li <pan2.li@intel.com>
21 months agoRISC-V: Add test for FP iceil auto vectorization
Pan Li [Fri, 13 Oct 2023 07:38:09 +0000 (15:38 +0800)] 
RISC-V: Add test for FP iceil auto vectorization

The below FP API are supported already by sharing the same standard
name, as well as the machine mode.

int iceil (float);

This patch would like to add the test cases for ensuring the
correctness.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/unop/math-iceil-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-iceil-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/vls/math-iceil-0.c: New test.

Signed-off-by: Pan Li <pan2.li@intel.com>
21 months agoRISC-V: Add test for FP llceil auto vectorization
Pan Li [Fri, 13 Oct 2023 07:16:27 +0000 (15:16 +0800)] 
RISC-V: Add test for FP llceil auto vectorization

The below FP API are supported already by sharing the same standard
name, as well as the machine mode.

long long llceil (double);

This patch would like to add the test cases for ensuring the
correctness.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/unop/math-llceil-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-llceil-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/vls/math-llceil-0.c: New test.

Signed-off-by: Pan Li <pan2.li@intel.com>
21 months agoC99 testsuite readiness: Some verified test case adjustments
Florian Weimer [Fri, 13 Oct 2023 07:34:55 +0000 (09:34 +0200)] 
C99 testsuite readiness: Some verified test case adjustments

The updated test cases still reproduce the bugs with old compilers.

gcc/testsuite/

* gcc.c-torture/compile/pc44485.c (func_21): Add missing cast.
* gcc.c-torture/compile/pr106101.c: Use builtins to avoid
calls to undeclared functions.  Change type of yyvsp to
char ** and introduce yyvsp1 to avoid type errors.
* gcc.c-torture/execute/pr111331-1.c: Add missing int.
* gcc.dg/pr100512.c: Unreduce test case and suppress only
-Wpointer-to-int-cast.
* gcc.dg/pr103003.c: Likewise.
* gcc.dg/pr103451.c: Add cast to long and suppress
-Wdiv-by-zero only.
* gcc.dg/pr68435.c: Avoid implicit int and missing
static function implementation warning.

21 months agoC99 test suite readiness: Some unverified test case adjustments
Florian Weimer [Fri, 13 Oct 2023 07:34:37 +0000 (09:34 +0200)] 
C99 test suite readiness: Some unverified test case adjustments

These changes are assumed not to interfere with the test objective,
but it was not possible to reproduce the historic test case failures
(with or without the modification here).

gcc/testsuite/

* gcc.c-torture/compile/20000105-1.c: Add missing int return type.
Call __builtin_exit instead of exit.
* gcc.c-torture/compile/20000105-2.c: Add missing void types.
* gcc.c-torture/compile/20000211-1.c (Lstream_fputc, Lstream_write)
(Lstream_flush_out, parse_doprnt_spec): Add missing function
declaration.
* gcc.c-torture/compile/20000224-1.c (call_critical_lisp_code):
Declare.
* gcc.c-torture/compile/20000314-2.c: Add missing void types.
* gcc.c-torture/compile/980816-1.c (XtVaCreateManagedWidget)
(XtAddCallback): Likewise.
* gcc.c-torture/compile/pr49474.c: Use struct
gfc_formal_arglist * instead of (implied) int type.
* gcc.c-torture/execute/20001111-1.c (foo): Add cast to
char *.
(main): Call __builtin_abort and __builtin_exit.

21 months agoC99 test suite readiness: Mark some C89 tests
Florian Weimer [Fri, 13 Oct 2023 07:34:37 +0000 (09:34 +0200)] 
C99 test suite readiness: Mark some C89 tests

Add -std=gnu89 to some tests which evidently target C89-only language
features.

gcc/testsuite/

* gcc.c-torture/compile/920501-11.c: Compile with -std=gnu89.
* gcc.c-torture/compile/920501-23.c: Likewise.
* gcc.c-torture/compile/920501-8.c: Likewise.
* gcc.c-torture/compile/920701-1.c: Likewise.
* gcc.c-torture/compile/930529-1.c: Likewise.

21 months agoor1k: Fix -Wincompatible-pointer-types warning during libgcc build
Florian Weimer [Fri, 13 Oct 2023 07:34:37 +0000 (09:34 +0200)] 
or1k: Fix -Wincompatible-pointer-types warning during libgcc build

libgcc/

* config/or1k/linux-unwind.h (or1k_fallback_frame_state): Add
missing cast.

21 months agoarc: Fix -Wincompatible-pointer-types warning during libgcc build
Florian Weimer [Fri, 13 Oct 2023 07:34:36 +0000 (09:34 +0200)] 
arc: Fix -Wincompatible-pointer-types warning during libgcc build

libgcc/

* config/arc/linux-unwind.h (arc_fallback_frame_state): Add
missing cast.

21 months agoriscv: Fix -Wincompatible-pointer-types warning during libgcc build
Florian Weimer [Fri, 13 Oct 2023 07:34:36 +0000 (09:34 +0200)] 
riscv: Fix -Wincompatible-pointer-types warning during libgcc build

libgcc/

* config/riscv/linux-unwind.h (riscv_fallback_frame_state): Add
missing cast.

21 months agocsky: Fix -Wincompatible-pointer-types warning during libgcc build
Florian Weimer [Fri, 13 Oct 2023 07:34:36 +0000 (09:34 +0200)] 
csky: Fix -Wincompatible-pointer-types warning during libgcc build

libgcc/

* config/csky/linux-unwind.h (csky_fallback_frame_state): Add
missing cast.

21 months agom68k: Avoid implicit function declaration in libgcc
Florian Weimer [Fri, 13 Oct 2023 07:34:36 +0000 (09:34 +0200)] 
m68k: Avoid implicit function declaration in libgcc

libgcc/

* config/m68k/fpgnulib.c (__cmpdf2): Declare.

21 months agolibstdc++: Fix tr1/8_c_compatibility/cstdio/functions.cc regression with recent glibc
Jakub Jelinek [Fri, 13 Oct 2023 07:09:32 +0000 (09:09 +0200)] 
libstdc++: Fix tr1/8_c_compatibility/cstdio/functions.cc regression with recent glibc

The following testcase started FAILing recently after the
https://sourceware.org/git/?p=glibc.git;a=commit;h=64b1a44183a3094672ed304532bedb9acc707554
glibc change which marked vfscanf with nonnull (1) attribute.
While vfwscanf hasn't been marked similarly (strangely), the patch changes
that too.  By using va_arg one hides the value of it from the compiler
(volatile keyword would do too, or making the FILE* stream a function
argument, but then it might need to be guarded by #if or something).

2023-10-13  Jakub Jelinek  <jakub@redhat.com>

* testsuite/tr1/8_c_compatibility/cstdio/functions.cc (test01):
Initialize stream to va_arg(ap, FILE*) rather than 0.
* testsuite/tr1/8_c_compatibility/cwchar/functions.cc (test01):
Likewise.

21 months agotree-optimization/111779 - Handle some BIT_FIELD_REFs in SRA
Richard Biener [Thu, 12 Oct 2023 09:34:57 +0000 (11:34 +0200)] 
tree-optimization/111779 - Handle some BIT_FIELD_REFs in SRA

The following handles byte-aligned, power-of-two and byte-multiple
sized BIT_FIELD_REF reads in SRA.  In particular this should cover
BIT_FIELD_REFs created by optimize_bit_field_compare.

For gcc.dg/tree-ssa/ssa-dse-26.c we now SRA the BIT_FIELD_REF
appearing there leading to more DSE, fully eliding the aggregates.

This results in the same false positive -Wuninitialized as the
older attempt to remove the folding from optimize_bit_field_compare,
fixed by initializing part of the aggregate unconditionally.

PR tree-optimization/111779
gcc/
* tree-sra.cc (sra_handled_bf_read_p): New function.
(build_access_from_expr_1): Handle some BIT_FIELD_REFs.
(sra_modify_expr): Likewise.
(make_fancy_name_1): Skip over BIT_FIELD_REF.

gcc/fortran/
* trans-expr.cc (gfc_trans_assignment_1): Initialize
lhs_caf_attr and rhs_caf_attr codimension flag to avoid
false positive -Wuninitialized.

gcc/testsuite/
* gcc.dg/tree-ssa/ssa-dse-26.c: Adjust for more DSE.
* gcc.dg/vect/vect-pr111779.c: New testcase.

21 months agotree-optimization/111773 - avoid CD-DCE of noreturn special calls
Richard Biener [Thu, 12 Oct 2023 08:13:58 +0000 (10:13 +0200)] 
tree-optimization/111773 - avoid CD-DCE of noreturn special calls

The support to elide calls to allocation functions in DCE runs into
the issue that when implementations are discovered noreturn we end
up DCEing the calls anyway, leaving blocks without termination and
without outgoing edges which is both invalid IL and wrong-code when
as in the example the noreturn call would throw.  The following
avoids taking advantage of both noreturn and the ability to elide
allocation at the same time.

For the testcase it's valid to throw or return 10 by eliding the
allocation.  But we have to do either where currently we'd run
off the function.

PR tree-optimization/111773
* tree-ssa-dce.cc (mark_stmt_if_obviously_necessary): Do
not elide noreturn calls that are reflected to the IL.

* g++.dg/torture/pr111773.C: New testcase.

21 months agoRISC-V: Add test for FP llround auto vectorization
Pan Li [Fri, 13 Oct 2023 06:13:26 +0000 (14:13 +0800)] 
RISC-V: Add test for FP llround auto vectorization

The below FP API are supported already by sharing the same standard
name, as well as the machine mode.

long long llround (double);

This patch would like to add the test cases for ensuring the correctness.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/unop/math-llround-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-llround-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/vls/math-llround-0.c: New test.

Signed-off-by: Pan Li <pan2.li@intel.com>
21 months agoRISC-V Regression: Fix FAIL of bb-slp-pr69907.c for RVV
Juzhe-Zhong [Fri, 13 Oct 2023 05:45:19 +0000 (13:45 +0800)] 
RISC-V Regression: Fix FAIL of bb-slp-pr69907.c for RVV

Like ARM SVE and GCN, add RVV.

gcc/testsuite/ChangeLog:

* gcc.dg/vect/bb-slp-pr69907.c: Add RVV.

21 months agoRISC-V: Add test for FP iroundf auto vectorization
Pan Li [Fri, 13 Oct 2023 04:11:56 +0000 (12:11 +0800)] 
RISC-V: Add test for FP iroundf auto vectorization

The below FP API are supported already by sharing the same standard
name, as well as the machine mode.

int iroundf (float);

This patch would like to add the test cases for ensuring the
correctness.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/unop/math-iround-0.c: New test.
* gcc.target/riscv/rvv/autovec/unop/math-iround-run-0.c: New test.
* gcc.target/riscv/rvv/autovec/vls/math-iround-0.c: New test.

Signed-off-by: Pan Li <pan2.li@intel.com>
21 months agoRISC-V: Fix the riscv_legitimize_poly_move issue on targets where the minimal VLEN...
Kito Cheng [Tue, 3 Oct 2023 02:27:24 +0000 (10:27 +0800)] 
RISC-V: Fix the riscv_legitimize_poly_move issue on targets where the minimal VLEN exceeds 512.

riscv_legitimize_poly_move was expected to ensure the poly value is at most 32
times smaller than the minimal VLEN (32 being derived from '4096 / 128').
This assumption held when our mode modeling was not so precisely defined.
However, now that we have modeled the mode size according to the correct minimal
VLEN info, the size difference between different RVV modes can be up to 64
times. For instance, comparing RVVMF64BI and RVVMF1BI, the sizes are [1, 1]
versus [64, 64] respectively.

gcc/ChangeLog:

* config/riscv/riscv.cc (riscv_legitimize_poly_move): Bump
max_power to 64.
* config/riscv/riscv.h (MAX_POLY_VARIANT): New.

gcc/testsuite/ChangeLog:

* g++.target/riscv/rvv/autovec/bug-01.C: New.
* g++.target/riscv/rvv/rvv.exp: Add autovec folder.

21 months agoRISC-V: Leverage stdint-gcc.h for RVV test cases
Pan Li [Fri, 13 Oct 2023 02:17:36 +0000 (10:17 +0800)] 
RISC-V: Leverage stdint-gcc.h for RVV test cases

Leverage stdint-gcc.h for the int64_t types instead of typedef.
Or we may have conflict with stdint-gcc.h in somewhere else.

gcc/testsuite/ChangeLog:

* gcc.target/riscv/rvv/autovec/unop/math-llrint-0.c: Include
stdint-gcc.h for int types.
* gcc.target/riscv/rvv/autovec/unop/math-llrint-run-0.c: Ditto.
* gcc.target/riscv/rvv/autovec/unop/test-math.h: Remove int64_t
typedef.

Signed-off-by: Pan Li <pan2.li@intel.com>