This patch was based on the __memcmp_power8 and the recent
__strlen_power10.
Improvements from __memcmp_power8:
1. Don't need alignment code.
On POWER10 lxvp and lxvl do not generate alignment interrupts, so
they are safe for use on caching-inhibited memory. Notice that the
comparison on the main loop will wait for both VSR to be ready.
Therefore aligning one of the input address does not improve
performance. In order to align both registers a vperm is necessary
which add too much overhead.
2. Uses new POWER10 instructions
This code uses lxvp to decrease contention on load by loading 32 bytes
per instruction.
The vextractbm is used to have a smaller tail code for calculating the
return value.
3. Performance improvement
This version has around 35% better performance on average. I saw no
performance regressions for any length or alignment.
Thanks Matheus for helping me out with some details.
Co-authored-by: Matheus Castanho <msc@linux.ibm.com>
Reviewed-by: Raphael M Zinsly <rzinsly@linux.ibm.com>
Reuse code for optimized strlen to implement a faster version of rawmemchr.
This takes advantage of the same benefits provided by the strlen implementation,
but needs some extra steps. __strlen_power10 code should be unchanged after this
change.
rawmemchr returns a pointer to the char found, while strlen returns only the
length, so we have to take that into account when preparing the return value.
To quickly check 64B, the loop on __strlen_power10 merges the whole block into
16B by using unsigned minimum vector operations (vminub) and checks if there are
any \0 on the resulting vector. The same code is used by rawmemchr if the char c
is 0. However, this approach does not work when c != 0. We first need to
subtract each byte by c, so that the value we are looking for is converted to a
0, then taking the minimum and checking for nulls works again.
The new code branches after it has compared ~256 bytes and chooses which of the
two strategies above will be used in the main loop, based on the char c. This
extra branch adds some overhead (~5%) for length ~256, but is quickly amortized
by the faster loop for larger sizes.
Compared to __rawmemchr_power9, this version is ~20% faster for length < 256.
Because of the optimized main loop, the improvement becomes ~35% for c != 0
and ~50% for c = 0 for strings longer than 256.
Reviewed-by: Lucas A. M. Magalhaes <lamm@linux.ibm.com>
Reviewed-by: Raphael M Zinsly <rzinsly@linux.ibm.com>
This implementation is based on __memset_power8 and integrates a lot
of suggestions from Anton Blanchard.
The biggest difference is that it makes extensive use of stxvl to
alignment and tail code to avoid branches and small stores. It has
three main execution paths:
a) "Short lengths" for lengths up to 64 bytes, avoiding as many
branches as possible.
b) "General case" for larger lengths, it has an alignment section
using stxvl to avoid branches, a 128 bytes loop and then a tail
code, again using stxvl with few branches.
c) "Zeroing cache blocks" for lengths from 256 bytes upwards and set
value being zero. It is mostly the __memset_power8 code but the
alignment phase was simplified because, at this point, address is
already 16-bytes aligned and also changed to use vector stores.
The tail code was also simplified to reuse the general case tail.
All unaligned stores use stxvl instructions that do not generate
alignment interrupts on POWER10, making it safe to use on
caching-inhibited memory.
On average, this implementation provides something around 30%
improvement when compared to __memset_power8.
Reviewed-by: Matheus Castanho <msc@linux.ibm.com>
Reviewed-by: Tulio Magno Quites Machado Filho <tuliom@linux.ibm.com>
This implementation is based on __memcpy_power8_cached and integrates
suggestions from Anton Blanchard.
It benefits from loads and stores with length for short lengths and for
tail code, simplifying the code.
All unaligned memory accesses use instructions that do not generate
alignment interrupts on POWER10, making it safe to use on
caching-inhibited memory.
The main loop has also been modified in order to increase instruction
throughput by reducing the dependency on updates from previous iterations.
On average, this implementation provides around 30% improvement when
compared to __memcpy_power7 and 10% improvement in comparison to
__memcpy_power8_cached.
This patch was initially based on the __memmove_power7 with some ideas
from strncpy implementation for Power 9.
Improvements from __memmove_power7:
1. Use lxvl/stxvl for alignment code.
The code for Power 7 uses branches when the input is not naturally
aligned to the width of a vector. The new implementation uses
lxvl/stxvl instead which reduces pressure on GPRs. It also allows
the removal of branch instructions, implicitly removing branch stalls
and mispredictions.
2. Use of lxv/stxv and lxvl/stxvl pair is safe to use on Cache Inhibited
memory.
On Power 10 vector load and stores are safe to use on CI memory for
addresses unaligned to 16B. This code takes advantage of this to
do unaligned loads.
The unaligned loads don't have a significant performance impact by
themselves. However doing so decreases register pressure on GPRs
and interdependence stalls on load/store pairs. This also improved
readability as there are now less code paths for different alignments.
Finally this reduces the overall code size.
3. Improved performance.
This version runs on average about 30% better than memmove_power7
for lengths larger than 8KB. For input lengths shorter than 8KB
the improvement is smaller, it has on average about 17% better
performance.
This version has a degradation of about 50% for input lengths
in the 0 to 31 bytes range when dest is unaligned.
Reviewed-by: Tulio Magno Quites Machado Filho <tuliom@linux.ibm.com>
Improvements compared to POWER9 version:
1. Take into account first 16B comparison for aligned strings
The previous version compares the first 16B and increments r4 by the number
of bytes until the address is 16B-aligned, then starts doing aligned loads at
that address. For aligned strings, this causes the first 16B to be compared
twice, because the increment is 0. Here we calculate the next 16B-aligned
address differently, which avoids that issue.
2. Use simple comparisons for the first ~192 bytes
The main loop is good for big strings, but comparing 16B each time is better
for smaller strings. So after aligning the address to 16 Bytes, we check
more 176B in 16B chunks. There may be some overlaps with the main loop for
unaligned strings, but we avoid using the more aggressive strategy too soon,
and also allow the loop to start at a 64B-aligned address. This greatly
benefits smaller strings and avoids overlapping checks if the string is
already aligned at a 64B boundary.
3. Reduce dependencies between load blocks caused by address calculation on loop
Doing a precise time tracing on the code showed many loads in the loop were
stalled waiting for updates to r4 from previous code blocks. This
implementation avoids that as much as possible by using 2 registers (r4 and
r5) to hold addresses to be used by different parts of the code.
Also, the previous code aligned the address to 16B, then to 64B by doing a
few 48B loops (if needed) until the address was aligned. The main loop could
not start until that 48B loop had finished and r4 was updated with the
current address. Here we calculate the address used by the loop very early,
so it can start sooner.
The main loop now uses 2 pointers 128B apart to make pointer updates less
frequent, and also unrolls 1 iteration to guarantee there is enough time
between iterations to update the pointers, reducing stalled cycles.
4. Use new P10 instructions
lxvp is used to load 32B with a single instruction, reducing contention in
the load queue.
vextractbm allows simplifying the tail code for the loop, replacing
vbpermq and avoiding having to generate a permute control vector.
Reviewed-by: Paul E Murphy <murphyp@linux.ibm.com>
Reviewed-by: Raphael M Zinsly <rzinsly@linux.ibm.com>
Reviewed-by: Lucas A. M. Magalhaes <lamm@linux.ibm.com>
1. Add the directories to hold POWER10 files.
2. Add support to select POWER10 libraries based on AT_PLATFORM.
3. Let submachine=power10 be set automatically.
