Will Schmidt [Mon, 13 Jul 2020 18:38:17 +0000 (13:38 -0500)]
[PATCH] rs6000 improve handling of built-in initialization [PR95952]
Handle a scenario with a combination of old hardware, gcc-8 and
binutils where gcc will ICE during it's selftest. This ICE was exposed
when the builtin processing for better #pragma support was added, where
we no longer skip builtin initialization based on the current mask.
Per the bug report and assorted debug, the ICE occurrs when building
the gcc-8 branch on a 970* based system with an old binutils. (gcc-9 and
newer is OK. binutils 2.34 is reported to allow success).
The attached patch adds a clause to the builtin initialization to skip
initialization of a builtin when the builtin mask is set but the icode
value is zero. The subsequent assert check remains in place.
2020-07-13 Will Schmidt <will_schmidt@vnet.ibm.com>
gcc/ChangeLog:
PR target/95952
* config/rs6000/rs6000.c (altivec_init_builtins): Add continue clause to
predicate builtin handling.
Andreas Krebbel [Wed, 21 Nov 2018 07:48:49 +0000 (07:48 +0000)]
S/390: Support vector load/store alignment hints
The IBM z14 POP adds an optional alignment operand to the vl, vst,
vlm, and vstm instruction (vector loads and stores). Vectors residing
on 8 or 16 byte boundaries might get loaded or stored faster on some
models given the instruction uses the proper hint operand. A wrong
hint will hurt performance though.
The attached testcase align-1 currently fails due to:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=88085
gcc/ChangeLog:
2018-11-21 Andreas Krebbel <krebbel@linux.ibm.com>
* configure.ac: Add check for Binutils to determine whether vector
load/store alignments hints are being supported.
* config.in: Regenerate.
* configure: Regenerate.
* config/s390/s390.c (print_operand): Support new output
modifier A.
* config/s390/s390.md ("movti"): Append alignment hint output
using the new output modifier 'A'.
* config/s390/vector.md ("mov<mode>", "*vec_tf_to_v1tf")
("*vec_ti_to_v1ti"): Likewise.
gcc/testsuite/ChangeLog:
2018-11-21 Andreas Krebbel <krebbel@linux.ibm.com>
* gcc.target/s390/vector/align-1.c: New test.
* gcc.target/s390/vector/align-2.c: New test.
Martin Liska [Thu, 2 Jul 2020 08:51:06 +0000 (10:51 +0200)]
gcc-changelog: sync from master.
contrib/ChangeLog:
* gcc-changelog/git_check_commit.py: New file.
* gcc-changelog/git_commit.py: New file.
* gcc-changelog/git_email.py: New file.
* gcc-changelog/git_repository.py: New file.
* gcc-changelog/git_update_version.py: New file.
* gcc-changelog/test_email.py: New file.
* gcc-changelog/test_patches.txt: New file.
Alex Coplan [Mon, 18 May 2020 15:29:04 +0000 (16:29 +0100)]
arm: Don't generate invalid LDRD insns
This fixes a bug in the arm backend where GCC generates invalid LDRD
instructions. The LDRD instruction requires the first transfer register to be
even, but GCC attempts to use odd registers here. For example, with the
following C code:
struct c {
double a;
} __attribute((aligned)) __attribute((packed));
struct c d;
struct c f(struct c);
void e() { f(d); }
The struct d is passed in registers r1 and r2 to the function f, and GCC
attempted to do this with a LDRD instruction when compiling with -march=armv7-a
on a soft float toolchain.
The fix is analogous to the corresponding one for STRD in the same function:
https://gcc.gnu.org/git/?p=gcc.git;a=commit;h=52057dc4ac5295caebf83147f688d769c93cbc8d
gcc/:
* config/arm/arm.c (output_move_double): Fix codegen when loading into
a register pair with an odd base register.
gcc/testsuite/:
* gcc.c-torture/compile/packed-aligned-1.c: New test.
* gcc.c-torture/execute/packed-aligned.c: New test.
Harald Anlauf [Wed, 24 Jun 2020 20:33:11 +0000 (22:33 +0200)]
Revert "PR fortran/95689 - ICE in check_sym_interfaces, at fortran/interface.c:2015"
With submodules, name mangling of interfaces may result in long internal
symbols overflowing an internal buffer. We now check that we do not
exceed the enlarged buffer size.
gcc/fortran/
PR fortran/95689
* interface.c (check_sym_interfaces): Enlarge temporary buffer,
and add check on length on mangled name to prevent overflow.
gcc/testsuite/
PR fortran/95689
* gfortran.dg/pr95689.f90: New test.
Harald Anlauf [Sat, 20 Jun 2020 14:09:45 +0000 (16:09 +0200)]
PR fortran/95689 - ICE in check_sym_interfaces, at fortran/interface.c:2015
With submodules, name mangling of interfaces may result in long internal
symbols overflowing an internal buffer. We now check that we do not
exceed the enlarged buffer size.
gcc/fortran/
PR fortran/95689
* interface.c (check_sym_interfaces): Enlarge temporary buffer,
and add check on length on mangled name to prevent overflow.
Eric Botcazou [Tue, 23 Jun 2020 16:33:28 +0000 (18:33 +0200)]
Fix memory corruption with vector and variant record
The problem is that Has_Constrained_Partial_View must be tested on the
base type of the designated type of an allocator.
gcc/ada/ChangeLog:
* gcc-interface/trans.c (gnat_to_gnu) <N_Allocator>: Minor tweaks.
Call Has_Constrained_Partial_View on base type of designated type.
Bin Cheng [Sat, 20 Jun 2020 08:56:21 +0000 (16:56 +0800)]
Add missing unit dependence vector in data dependence analysis
Current data dependence analysis misses unit distant vector if DRs in
DDR have the same invariant access functions. This adds the vector as
the constant access function case.
+FAIL: libgomp.c/target-39.c (internal compiler error)
+FAIL: libgomp.c/target-39.c (test for excess errors)
+UNRESOLVED: libgomp.c/target-39.c compilation failed to produce executable
This is in a '--enable-offload-targets=[...],hsa' build, with '-foffload=hsa'
enabled (by default).
during GIMPLE pass: hsagen
source-gcc/libgomp/testsuite/libgomp.c/target-39.c: In function ‘main._omp_fn.0.hsa.0’:
source-gcc/libgomp/testsuite/libgomp.c/target-39.c:23:11: internal compiler error: Segmentation fault
23 | #pragma omp target map(from:err)
| ^~~
[...]
GDB:
Program received signal SIGSEGV, Segmentation fault.
fndecl_built_in_p (node=0x0, name=BUILT_IN_PREFETCH) at [...]/source-gcc/gcc/tree.h:6267
6267 return (fndecl_built_in_p (node, BUILT_IN_NORMAL)
(gdb) bt
#0 fndecl_built_in_p (node=0x0, name=BUILT_IN_PREFETCH) at [...]/source-gcc/gcc/tree.h:6267
#1 0x0000000000b19739 in gen_hsa_insns_for_call (stmt=stmt@entry=0x7ffff693b200, hbb=hbb@entry=0x2b152c0) at [...]/source-gcc/gcc/hsa-gen.c:5304
#2 0x0000000000b1aca7 in gen_hsa_insns_for_gimple_stmt (stmt=0x7ffff693b200, hbb=hbb@entry=0x2b152c0) at [...]/source-gcc/gcc/hsa-gen.c:5770
#3 0x0000000000b1bd21 in gen_body_from_gimple () at [...]/source-gcc/gcc/hsa-gen.c:5999
#4 0x0000000000b1dbd2 in generate_hsa (kernel=<optimized out>) at [...]/source-gcc/gcc/hsa-gen.c:6596
#5 0x0000000000b1de66 in (anonymous namespace)::pass_gen_hsail::execute (this=0x2a2aac0) at [...]/source-gcc/gcc/hsa-gen.c:6680
#6 0x0000000000d06f90 in execute_one_pass (pass=pass@entry=0x2a2aac0) at [...]/source-gcc/gcc/passes.c:2502
[...]
(gdb) up
#1 0x0000000000b19739 in gen_hsa_insns_for_call (stmt=stmt@entry=0x7ffff693b200, hbb=hbb@entry=0x2b152c0) at /home/thomas/tmp/source/gcc/build/track-slim-omp/source-gcc/gcc/hsa-gen.c:5304
5304 if (fndecl_built_in_p (function_decl, BUILT_IN_PREFETCH))
(gdb) print function_decl
$1 = (tree) 0x0
(gdb) list
5299 if (!gimple_call_builtin_p (stmt, BUILT_IN_NORMAL))
5300 {
5301 tree function_decl = gimple_call_fndecl (stmt);
5302 /* Prefetch pass can create type-mismatching prefetch builtin calls which
5303 fail the gimple_call_builtin_p test above. Handle them here. */
5304 if (fndecl_built_in_p (function_decl, BUILT_IN_PREFETCH))
5305 return;
5306
5307 if (function_decl == NULL_TREE)
5308 {
The problem is present already since 2016-11-23 commit 56b1c60e412fcf1245b4780871553cbdebb956a3 (r242761) "Merge from HSA branch to
trunk", and the fix obvious enough.
Thomas Koenig [Sun, 14 Jun 2020 11:54:19 +0000 (13:54 +0200)]
When avoiding double deallocation, look at namespace, expression and component.
Our finalization handling is a mess. Really, we should get to try and get
this fixed for gcc 11.
In the meantime, here is a patch which fixes a regression I introduced
when fixing a regression with a memory leak. The important thing
here is to realize that we do not need to finalize (and deallocate)
multiple times for the same expression and the same component
in the same namespace. It might cause code size regressions, but
better big code than wrong code...
PR fortran/94109
* class.c (finalize_component): Return early if finalization has
already happened for expression and component within namespace.
* gfortran.h (gfc_was_finalized): New type.
(gfc_namespace): Add member was_finalzed.
(gfc_expr): Remove finalized.
* symbol.c (gfc_free_namespace): Free was_finalized.
Carl Love [Wed, 10 Jun 2020 21:12:08 +0000 (16:12 -0500)]
pr94833, fix vec_first_match_index for nulls
Backported patch from mainline. Updated ChangeLog format
commit 24f68831d25bad739a6fe167a58b5b4c0c3cbf9a
Author: Carl Love <cel@us.ibm.com>
Date: Wed Apr 29 10:23:11 2020 -0500
Note the mainlin patch adds tests to the test case in mainline. The test
case does not exist in GCC 8. Only the functional change in
gcc/config/rs6000/vsx.md was backported.
Thomas Schwinge [Wed, 6 May 2020 08:05:49 +0000 (10:05 +0200)]
'libgomp.oacc-fortran/{error_,}stop-{1,2,3}.f': initialize before the checkpoint
If, for example, GCC is configured such that 'libgomp-plugin-nvptx.so.1'
dynamically links against 'libcuda.so.1', but testing is run on a system where
there is no 'libcuda.so.1', this produces output such as:
PASS: libgomp.oacc-fortran/error_stop-1.f -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable -O0 (test for excess errors)
PASS: libgomp.oacc-fortran/error_stop-1.f -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable -O0 execution test
FAIL: libgomp.oacc-fortran/error_stop-1.f -DACC_DEVICE_TYPE_host=1 -DACC_MEM_SHARED=1 -foffload=disable -O0 output pattern test, is CheCKpOInT
libgomp: while loading libgomp-plugin-nvptx.so.1: libcuda.so.1: cannot open shared object file: No such file or directory
ERROR STOP
Error termination. Backtrace: [...]
, should match CheCKpOInT(
|
|^M)+ERROR STOP (
|
|^M)+Error termination.*
..., where after 'CheCKpOInT' we got 'libgomp: while loading [...]' injected
before the expected 'ERROR STOP'.
This fixes PR94591. The problem was the function
aarch64_evpc_rev_local() matching vector permutations that were not
reversals. In particular, prior to this patch, this function matched the
identity permutation which led to generating bogus REV64 insns which
were rejected by the assembler.
gcc/
PR target/94591
* config/aarch64/aarch64.c (aarch64_evpc_rev_local): Don't match
identity permutation.
gcc/testsuite/
PR target/94591
* gcc.c-torture/execute/pr94591.c: New test.
Kyrylo Tkachov [Wed, 6 May 2020 15:20:38 +0000 (16:20 +0100)]
[AArch64] Use __getauxval instead of getauxval in LSE detection code in libgcc
This version of the fix uses __getauxval instead of getauxval.
The whole thing is guarded simply on __gnu_linux__.
__getauxval was introduced in 2.16 but the aarch64 port was added in 2.17 so in practice I expect all aarch64 glibcs to support __getauxval.
Bootstrapped and tested on aarch64-none-linux-gnu.
Also tested on aarch64-none-elf.
Martin Liska [Fri, 29 May 2020 09:29:25 +0000 (11:29 +0200)]
Fix various limitations of git-backport.py.
I've just tested the script and I'm going to install the patch
to all active branches.
contrib/ChangeLog:
* git-backport.py: The script did 'git co HEAD~' when
there was no modified ChangeLog file in a successful
git cherry pick.
Run cherry-pick --continue without editor.
Mark Eggleston [Thu, 7 May 2020 07:15:26 +0000 (08:15 +0100)]
Fortran : ICE in gfc_trans_label_assign PR50392
A function may contain an assigned goto. If the the return variable
is an integer a statement can be assigned to it. Prior to this fix
this resulted in an ICE.
PR fortran/50392
* trans-decl.c (gfc_get_symbol_decl): Remove unnecessary block
delimiters. Add auxiliary variables if a label is assigned to
a return variable. (gfc_gat_fake_result): If the symbol has an
assign attribute set declaration from the symbol's backend
declaration.
2020-05-28 Mark Eggleston <markeggleston@gnu.gcc.org>
gcc/testsuite/
PR fortran/50392
* gfortran.dg/pr50392.f: New test.