James Cowgill [Thu, 8 Jan 2015 16:05:57 +0000 (16:05 +0000)]
Remove inclusion of acquire_release_volatile.h on mips
(Apply commit 6fefba1 from 'release-7_4' branch.)
I'm not entirely sure how this ended up here. It might have been the case
that MIPS processors used to do this, but the ISA manuals don't say anything
about volatile loads / stores having acquire / release semantics, so just
remove it to be safe.
Ivan Maidanski [Sat, 17 Aug 2013 11:05:36 +0000 (15:05 +0400)]
Fix: Do not use LDREXD/STREXD for Clang3.2/arm (and earlier)
(Apply commit 2daf685 from 'master' branch.)
* src/atomic_ops/sysdeps/gcc/arm.h
(AO_compare_double_and_swap_double): Do not define for
pre-Clang3.3 (since the latter does not allocate register pairs for
LDREXD/STREXD instructions properly); add comment.
* src/atomic_ops/sysdeps/gcc/arm.h
(AO_double_compare_and_swap): Specify that LDREXD and STREXD use 2
adjacent registers (thus preventing Clang3.3 from register allocation
failures leading to "registers may not be the same" or
"even register required" GAS errors).
Ivan Maidanski [Fri, 29 Mar 2013 04:50:51 +0000 (08:50 +0400)]
Fix generalized fetch_and_add function
* src/atomic_ops/generalize-small.template (AO_XSIZE_fetch_and_add):
Do not use AO_EXPECT_FALSE (since defined only in v7.3+).
* src/atomic_ops/generalize-small.h: Regenerate.
Ivan Maidanski [Fri, 29 Mar 2013 03:31:06 +0000 (07:31 +0400)]
Fix asm constraints of primitives in sunc/x86.h
(Apply commit 31df7fa from 'master' branch.)
* src/atomic_ops/sysdeps/sunc/x86.h (AO_fetch_and_add_full,
AO_char_fetch_and_add_full, AO_short_fetch_and_add_full, AO_or_full,
AO_test_and_set_full,
AO_compare_and_swap_full): Use "+m" asm constraint for *addr instead
of "=m" (because the value pointed by addr is read and written by the
code).
Ivan Maidanski [Wed, 13 Mar 2013 18:46:24 +0000 (22:46 +0400)]
Fix asm constraint of CAS memory operand for gcc/alpha
* src/atomic_ops/sysdeps/gcc/alpha.h (AO_compare_and_swap): Use "+m"
asm constraint for *addr instead of "=m" (because the value pointed by
addr is read and written by the code).
Ivan Maidanski [Tue, 12 Mar 2013 18:29:33 +0000 (22:29 +0400)]
Fix asm constraint of compare_and_swap for clang-3.1/mips
(Apply commit 3e4c369 from 'master' branch.)
* src/atomic_ops/sysdeps/gcc/mips.h (AO_compare_and_swap): Use "+m"
asm constraint for *addr instead of "+R" (the letter is unsupported
by clang3.1 resulting in "invalid output constraint in asm" error).
Ivan Maidanski [Thu, 14 Feb 2013 10:08:47 +0000 (14:08 +0400)]
Fix (remove) invalid include of read_ordered.h for ARM
(fix commit 2df9c1e partially)
* src/atomic_ops/sysdeps/gcc/arm.h: Do not include read_ordered.h
(because load_acquire should contain a DMB instruction in a multi-core
case, new load_acquire/read primitives implementation is generalized
using nop_full/read which either contain a DMB instruction, or is just
a compiler barrier for uniprocessor).
* src/atomic_ops/sysdeps/armcc/arm_v6.h: Likewise.
* src/atomic_ops/sysdeps/msftc/arm.h: Likewise.
Ivan Maidanski [Mon, 14 Jan 2013 20:56:35 +0000 (00:56 +0400)]
Fix README regarding _acquire_read barrier
* doc/README.txt (_acquire_read): Add information about memory
barrier (similar as in atomic_ops.h).
* doc/README.txt (_release_read): Remove information about
non-existing barrier.
Ivan Maidanski [Sun, 13 Jan 2013 18:42:52 +0000 (22:42 +0400)]
Fix return type of AO_int_X primitives defined in ao_t_is_int header
(Apply commit 3e9fc4f from 'add-dbl-load-store' branch.)
* src/atomic_ops/sysdeps/ao_t_is_int.h
(AO_int_load_acquire, AO_int_fetch_and_add_full,
AO_int_fetch_and_add1_acquire, AO_int_fetch_and_add1_release,
AO_int_fetch_and_sub1_acquire, AO_int_fetch_and_sub1_release): Cast
result to unsigned int instead of signed int.
Ivan Maidanski [Sun, 13 Jan 2013 16:19:51 +0000 (20:19 +0400)]
Fix ordered_except_wr header inclusion for s390
* src/atomic_ops/sysdeps/gcc/s390.h: Move include of
ordered_except_wr.h down to be after all_aligned_atomic_load_store.h
inclusion (since the latter defines AO_X_load/store primitives used by
the former header).
Ivan Maidanski [Sun, 13 Jan 2013 16:04:36 +0000 (20:04 +0400)]
Fix read_ordered.h inclusion for ARM
* src/atomic_ops/sysdeps/armcc/arm_v6.h: Move include of read_ordered.h
down to the end of the file to be after AO_load/store definition
(since read_ordered.h depends on these primitives).
* src/atomic_ops/sysdeps/gcc/arm.h: Likewise.
* src/atomic_ops/sysdeps/msftc/arm.h: Likewise.
Ivan Maidanski [Mon, 7 Jan 2013 15:24:02 +0000 (19:24 +0400)]
Fix return type of AO_char/short/int_load_read() in read_ordered.h
* src/atomic_ops/sysdeps/read_ordered.h (AO_char_load_read,
AO_short_load_read, AO_int_load_read): Change return type (and the
type of "result" local variable) from AO_t to the type matching the
name of the function (i.e., unsigned char/short/int, respectively).
Ivan Maidanski [Fri, 4 Jan 2013 19:28:42 +0000 (23:28 +0400)]
Fix template-based headers regeneration order in src/Makefile
* src/atomic_ops/Makefile.am (BUILT_SOURCES): Define to force
regeneration of generalize-small.h (if the corresponding
source template is changed) before compilation of .c files (that use
atomic_ops.h).
Ivan Maidanski [Wed, 2 Jan 2013 08:31:08 +0000 (12:31 +0400)]
Fix AO_XSIZE_load/store definition order in generalize-small template
* src/atomic_ops/generalize-small.template (AO_XSIZE_load): Move
definition (which is based on AO_XSIZE_load_acquire) down to be after
AO_XSIZE_load_acquire definition.
* src/atomic_ops/generalize-small.template (AO_XSIZE_store): Move
definition (which is based on AO_XSIZE_store_release) down to be after
AO_XSIZE_store_release definition.
* src/atomic_ops/generalize-small.h : Regenerate.
Ivan Maidanski [Mon, 8 Oct 2012 20:06:19 +0000 (00:06 +0400)]
Fix variable type to match printf format specifier in test_stack
* tests/test_stack.c (run_one_test): Change type of "index" local
variable from long to int to match printf format specifier (when
VERBOSE defined); cast from "arg" pointer to integer via size_t (to
avoid 64-bit compiler warning).
Ivan Maidanski [Mon, 8 Oct 2012 10:51:30 +0000 (14:51 +0400)]
Properly fix compare_double_and_swap_double for gcc/x32 (use built-in CAS)
(Apply commit cbe517e from 'master' branch.)
* src/atomic_ops/sysdeps/gcc/x86.h (AO_double_compare_and_swap_full):
New function (only if __x86_64__) implemented using GCC built-in
__sync CAS primitive available for x32.
* src/atomic_ops/sysdeps/gcc/x86.h
(AO_HAVE_double_compare_and_swap_full): New macro (for x32 only).
* src/atomic_ops/sysdeps/gcc/x86.h
(AO_compare_double_and_swap_double_full): Implement using
double_compare_and_swap_full instead of cmpxchg8b for x32 (since
x86_64 has cmpxchg and cmpxchg16b but not cmpxchg8b).
Ivan Maidanski [Mon, 1 Oct 2012 10:48:43 +0000 (14:48 +0400)]
Fix double_ptr_storage definition for gcc/x32
(Apply commit 9957645 from 'master' branch.)
* src/atomic_ops/sysdeps/standard_ao_double_t.h (double_ptr_storage):
Define as "unsigned long long" instead of __m128 for gcc/x32 (i.e.,
ILP32 on x86_64); update and reformat comment.
Ivan Maidanski [Tue, 18 Sep 2012 16:48:08 +0000 (20:48 +0400)]
Fix compare_double_and_swap_double_full for gcc/x32
* src/atomic_ops.h: Include gcc/x86.h instead of gcc/x86_64.h if
gcc/x32 (i.e., GCC (or Intel compiler), __x86_64__ is defined, and
__ILP32__ is defined) to have AO_compare_double_and_swap_double_full
defined properly (based on cmpxchg8b); explicitly define
AO_USE_PENTIUM4_INSTRS for gcc/x32.
* src/atomic_ops/sysdeps/gcc/x86_64.h: Remove ILP32-specific code
(AO_T_IS_INT definition).
Ivan Maidanski [Wed, 3 Oct 2012 14:02:03 +0000 (18:02 +0400)]
Fix AO_pause sleep delay for particular argument values (Win32)
* src/atomic_ops.c (AO_pause): Fix millis value (passed to Win32
Sleep) for the case of 'n' argument in range between 12 and 21,
inclusive if AO_USE_WIN32_PTHREADS (sleep for 1 ms in this case).
* src/atomic_ops_stack.c (AO_pause): Evaluate 'msecs' value using the
same algorithm as in atomic_ops.c (Win32 only).
Ivan Maidanski [Wed, 3 Oct 2012 07:59:33 +0000 (11:59 +0400)]
Fix visibility and initial value of 'dummy' variable in atomic_ops_stack
* src/atomic_ops_stack.c (dummy): Define as static (to make symbol
without AO_ prefix visible only within this file); initialize to 1
(same as in atomic_ops.c, so that AO_spin would never really alter
its value).
Ivan Maidanski [Tue, 18 Sep 2012 15:03:19 +0000 (19:03 +0400)]
Fix cmpxchg16b-based compare_double_and_swap_double for SunCC/x86_64
* src/atomic_ops/sysdeps/sunc/x86_64.h
(AO_compare_double_and_swap_double_full): Fix inline assembly
arguments commenting out "m"(*addr) one (similar to that in
sunc/x86.h) if AO_CMPXCHG16B_AVAILABLE defined.
Ivan Maidanski [Wed, 5 Sep 2012 04:25:24 +0000 (08:25 +0400)]
Fix compare_double_and_swap_double for clang/x86 in PIC mode
* src/atomic_ops/sysdeps/gcc/x86.h
(AO_compare_double_and_swap_double_full): Re-implement for PIC mode
saving EBX to a local variable (instead of pushing it to stack) and
saving memory operand address to a register (edi which is manually
preserved), so that the whole code could also work even if EBX points
to memory operand (e.g., in Clang); test __PIC__ using ifdef (instead
of "if"); update comments; reformat code.
Ivan Maidanski [Wed, 4 Jul 2012 04:32:03 +0000 (08:32 +0400)]
Fix AO_compare_double_and_swap_double_full for gcc-4.2.1/x86 in PIC mode
* src/atomic_ops/sysdeps/gcc/x86.h
(AO_compare_double_and_swap_double_full): Use EDI register for
"new_val1" argument in PIC mode only for GCC 4.3+ to workaround
a problem with older compiler versions (e.g., GCC 4.2.1 [FreeBSD])
that do not recognize 'D' as a valid register specification; update
comment.
* src/atomic_ops/generalize-small.template
(AO_XSIZE_fetch_and_add_full,
AO_XSIZE_fetch_and_add_acquire, AO_XSIZE_fetch_and_add_release):
Fix type of function arguments (or local variables) by adding
'unsigned' to XCTYPE.
* src/atomic_ops/generalize-small.h: Regenerate.
H.J. Lu [Tue, 1 May 2012 10:31:30 +0000 (14:31 +0400)]
Fix for x32 by removing 'q' suffix in x86-64 instructions
(Apply commit 711f7fe from 'master' branch)
We do not need the 'q' suffix on x86_64 atomic instructions for AO_t
which is defined as "unsigned long"; the later is 32-bit for x32
and 64-bit for x86-64; the register operand in x86-64 atomic
instructions is sufficient to properly determine the register size.
Ivan Maidanski [Wed, 28 Mar 2012 04:07:31 +0000 (08:07 +0400)]
Fix configure to define __PIC__ macro explicitly only if not done by GCC
(Apply commit 058dd11 from 'master' branch)
* configure.ac (PICFLAG): Adjust AC_MSG_CHECKING message; report
"none" result if no -fPIC is needed (for Cygwin/MinGW); determine
whether GCC -fPIC option causes __PIC__ macro definition; pass
-D __PIC__ to PICFLAG (instead of CFLAGS) only if not defined
automatically in case -fPIC specified; update comment.
Ivan Maidanski [Mon, 26 Mar 2012 04:34:27 +0000 (08:34 +0400)]
Fix double_ptr_storage definition for GCC pre-v4 (x86_64)
* src/atomic_ops/sysdeps/standard_ao_double_t.h: Do not use __m128
if GCC pre-v4 (at least, xmmintrin.h is missing in gcc-3.4.3-x86_64
included in Solaris 10 distribution).
Ivan Maidanski [Wed, 7 Mar 2012 15:12:49 +0000 (19:12 +0400)]
Fix AO_compare_double_and_swap_double_full for gcc/x86 (PIC mode)
* src/atomic_ops/sysdeps/gcc/x86.h
(AO_compare_double_and_swap_double_full): Use EDI register for
"new_val1" argument instead of a memory operand and use XCHG assembler
instruction instead of push/pop in case of PIC mode (to workaround
a bug in GCC 4.6.1); update the comment.
Ivan Maidanski [Wed, 7 Mar 2012 03:31:55 +0000 (07:31 +0400)]
Fix a typo in comment
* src/atomic_ops/sysdeps/gcc/x86.h
(AO_compare_double_and_swap_double_full): Fix argument name in the
comment.
* src/atomic_ops/sysdeps/sunc/x86.h
(AO_compare_double_and_swap_double_full): Likewise.
projects / libatomic_ops / log