This patch enables SSE2 memset for AMD's upcoming Orochi processor.
This patch also fixes the following bug:
For misaligned blocks larger than > 144 Bytes, memset branches into
the integer code path depending on the value of misalignment even if
the startup code chooses the SSE2 code path upfront, when multiarch
is enabled.
32bit memset-sse2.S assumes cache size is multiple of 128 bytes. If
it isn't true, memset-sse2.S will fail. For example, a processor can
have 24576 KB L3 cache and 20 cores. That is 2516582 byte per core. Half
of it is 1258291, which isn't helpful for vector instructions. This
patch rounds cache sizes to multiple of 256 bytes and adds "raw" cache
sizes.
This patch includes optimized 64bit memcpy/memmove for Atom, Core 2 and
Core i7. It improves memcpy by up to 3X on Atom, up to 4X on Core 2 and
up to 1X on Core i7. It also improves memmove by up to 3X on Atom, up to
4X on Core 2 and up to 2X on Core i7.
If a signal arrived during a symbol lookup and the signal handler also
required a symbol lookup, the end of the lookup in the signal handler reset
the flag whether restoring AVX/SSE registers is needed. Resetting means
in this case that the tail part of the outer lookup code will try to
restore the registers and this can fail miserably. We now restore to the
previous value which makes nesting calls possible.
On 64-bit machines we should not split doubles into two 32 bit
integer and handle the words separately. We have wide registers.
This patch implements a 64-bit ceil version. Ideally all other
functions will be converted over time.
This patch fixes mixed SSE/AVX audit and checks AVX only once in
_dl_runtime_profile. When an AVX or SSE register value in pltenter is
modified, we have to make sure that the SSE part value is the same in both
lr_xmm and lr_vector fields so that pltexit will get the correct value
from either lr_xmm or lr_vector fields. AVX-enabled pltenter should
update both lr_xmm and lr_vector fields to support stacked AVX/SSE
pltenter functions.
The meaning of the 25-14 bits in EAX returned from cpuid with EAX = 4
has been changed from "the maximum number of threads sharing the cache"
to "the maximum number of addressable IDs for logical processors sharing
the cache" if cpuid takes EAX = 11. We need to use results from both
EAX = 4 and EAX = 11 to get the number of threads sharing the cache.
The 25-14 bits in EAX on Core i7 is 15 although the number of logical
processors is 8. Here is a white paper on this:
http://software.intel.com/en-us/articles/intel-64-architecture-processor-topology-enumeration/
This patch correctly counts number of logical processors on Intel CPUs
with EAX = 11 support on cpuid. Tested on Dinnington, Core i7 and
Nehalem EX/EP.
It also fixed Pentium Ds workaround since EBX may not have the right
value returned from cpuid with EAX = 1.
This patch adds 32bit SSE4.2 string functions. It uses -16L instead of
0xfffffffffffffff0L, which works for both 32bit and 64bit long. Tested
on 32bit Core i7 and Core 2.
This patch adds multiarch support when configured for i686. I modified
some x86-64 functions to support 32bit. I will contribute 32bit SSE string
and memory functions later.
We use sigaltstack internally which on some systems is a syscall
and should be used as such. Move the x86-64 version to the Linux
specific directory and create in its place a file which always
causes compile errors.
We use a callback function into libc.so to get access to the data
structure with the information and have special versions of the test
macros which automatically use this function.
SSE registers are used for passing parameters and must be preserved
in runtime relocations. This is inside ld.so enforced through the
tests in tst-xmmymm.sh. But the malloc routines used after startup
come from libc.so and can be arbitrarily complex. It's overkill
to save the SSE registers all the time because of that. These calls
are rare. Instead we save them on demand. The new infrastructure
put in place in this patch makes this possible and efficient.
The test now takes the callgraph into account. Only code called
during runtime relocation is affected by the limitation. We now
determine the affected object files as closely as possible from
the outside. This allowed to remove some the specializations
for some of the string functions as they are only used in other
code paths.
This patch introduces a test to make sure no function modifies the
xmm/ymm registers. With the exception of the auditing functions.
The test is probably too pessimistic. All code linked into ld.so
is checked. Perhaps at some point the callgraph starting from
_dl_fixup and _dl_profile_fixup is checked and we can start using
faster SSE-using functions in parts of ld.so.
There will be more than one function which, in multiarch mode, wants
to use SSSE3. We should not test in each of them for Atoms with
slow SSSE3. Instead, disable the SSSE3 bit in the startup code for
such machines.
The original AVX patch used a function pointer to handle the difference
between machines with and without AVX support. This is insecure. A
well-placed memory exploit could lead to redirection of the execution.
Using a variable and several tests is a bit slower but cannot be
exploited in this way.
Some of the new multi-arch string functions for x86-64 were
not aligned to 16 byte boundarie,s possibly creating unnecessary
cache line misses and delays.
This patch adds SSSE3 strcpy/stpcpy. I got up to 4X speed up on Core 2
and Core i7. I disabled it on Atom since SSSE3 version is slower for
shorter (<64byte) data.
The test to call the indirect function now includes a subtest to
checked whether the symbol is defined. When coming to that point
this is almost always the case. The test for STT_GNU_IFUNC on the
other hand rarely is true. Move it to the front means we don't have
to perform the second test unless really necessary.
SO far Intel and AMD use exactly the same bits meaning the same
things in CPUID index 1. Simplify the code. Should an architecture
come along which doesn't use the same semantics then it must use a
different index value than COMMON_CPUID_INDEX_1.
If longjmp restores the stack frame to an address which is beyond
the stack frame at the time of the longjmp call it would install
an uninitialized stack frame. If compiled with _FORTIFY_SOURCE
defined, longjmp will now bail out in this situation.
