After gcc commit a459ee44c0a74b0df0485ed7a56683816c02aae9 Author: Kyrylo Tkachov kyrylo.tkachov@arm.com
aarch64: Improve size heuristic for cpymem expansion
the following benchmarks grew in size by more than 1%: - 458.sjeng grew in size by 4% from 105780 to 109944 bytes - 459.GemsFDTD grew in size by 2% from 247504 to 251468 bytes
Below reproducer instructions can be used to re-build both "first_bad" and "last_good" cross-toolchains used in this bisection. Naturally, the scripts will fail when triggerring benchmarking jobs if you don't have access to Linaro TCWG CI.
For your convenience, we have uploaded tarballs with pre-processed source and assembly files at: - First_bad save-temps: https://ci.linaro.org/job/tcwg_bmk_ci_gnu-bisect-tcwg_bmk_apm-gnu-master-aar... - Last_good save-temps: https://ci.linaro.org/job/tcwg_bmk_ci_gnu-bisect-tcwg_bmk_apm-gnu-master-aar... - Baseline save-temps: https://ci.linaro.org/job/tcwg_bmk_ci_gnu-bisect-tcwg_bmk_apm-gnu-master-aar...
Configuration: - Benchmark: SPEC CPU2006 - Toolchain: GCC + Glibc + GNU Linker - Version: all components were built from their tip of trunk - Target: aarch64-linux-gnu - Compiler flags: -Os -flto - Hardware: APM Mustang 8x X-Gene1
This benchmarking CI is work-in-progress, and we welcome feedback and suggestions at linaro-toolchain@lists.linaro.org . In our improvement plans is to add support for SPEC CPU2017 benchmarks and provide "perf report/annotate" data behind these reports.
THIS IS THE END OF INTERESTING STUFF. BELOW ARE LINKS TO BUILDS, REPRODUCTION INSTRUCTIONS, AND THE RAW COMMIT.
This commit has regressed these CI configurations: - tcwg_bmk_gnu_apm/gnu-master-aarch64-spec2k6-Os_LTO
First_bad build: https://ci.linaro.org/job/tcwg_bmk_ci_gnu-bisect-tcwg_bmk_apm-gnu-master-aar... Last_good build: https://ci.linaro.org/job/tcwg_bmk_ci_gnu-bisect-tcwg_bmk_apm-gnu-master-aar... Baseline build: https://ci.linaro.org/job/tcwg_bmk_ci_gnu-bisect-tcwg_bmk_apm-gnu-master-aar... Even more details: https://ci.linaro.org/job/tcwg_bmk_ci_gnu-bisect-tcwg_bmk_apm-gnu-master-aar...
Reproduce builds: <cut> mkdir investigate-gcc-a459ee44c0a74b0df0485ed7a56683816c02aae9 cd investigate-gcc-a459ee44c0a74b0df0485ed7a56683816c02aae9
# Fetch scripts git clone https://git.linaro.org/toolchain/jenkins-scripts
# Fetch manifests and test.sh script mkdir -p artifacts/manifests curl -o artifacts/manifests/build-baseline.sh https://ci.linaro.org/job/tcwg_bmk_ci_gnu-bisect-tcwg_bmk_apm-gnu-master-aar... --fail curl -o artifacts/manifests/build-parameters.sh https://ci.linaro.org/job/tcwg_bmk_ci_gnu-bisect-tcwg_bmk_apm-gnu-master-aar... --fail curl -o artifacts/test.sh https://ci.linaro.org/job/tcwg_bmk_ci_gnu-bisect-tcwg_bmk_apm-gnu-master-aar... --fail chmod +x artifacts/test.sh
# Reproduce the baseline build (build all pre-requisites) ./jenkins-scripts/tcwg_bmk-build.sh @@ artifacts/manifests/build-baseline.sh
# Save baseline build state (which is then restored in artifacts/test.sh) mkdir -p ./bisect rsync -a --del --delete-excluded --exclude /bisect/ --exclude /artifacts/ --exclude /gcc/ ./ ./bisect/baseline/
cd gcc
# Reproduce first_bad build git checkout --detach a459ee44c0a74b0df0485ed7a56683816c02aae9 ../artifacts/test.sh
# Reproduce last_good build git checkout --detach 8f95e3c04d659d541ca4937b3df2f1175a1c5f05 ../artifacts/test.sh
cd .. </cut>
Full commit (up to 1000 lines): <cut> commit a459ee44c0a74b0df0485ed7a56683816c02aae9 Author: Kyrylo Tkachov kyrylo.tkachov@arm.com Date: Wed Sep 29 11:21:45 2021 +0100
aarch64: Improve size heuristic for cpymem expansion
Similar to my previous patch for setmem this one does the same for the cpymem expansion. We count the number of ops emitted and compare it against the alternative of just calling the library function when optimising for size. For the code: void cpy_127 (char *out, char *in) { __builtin_memcpy (out, in, 127); }
void cpy_128 (char *out, char *in) { __builtin_memcpy (out, in, 128); }
we now emit a call to memcpy (with an extra MOV-immediate instruction for the size) instead of: cpy_127(char*, char*): ldp q0, q1, [x1] stp q0, q1, [x0] ldp q0, q1, [x1, 32] stp q0, q1, [x0, 32] ldp q0, q1, [x1, 64] stp q0, q1, [x0, 64] ldr q0, [x1, 96] str q0, [x0, 96] ldr q0, [x1, 111] str q0, [x0, 111] ret cpy_128(char*, char*): ldp q0, q1, [x1] stp q0, q1, [x0] ldp q0, q1, [x1, 32] stp q0, q1, [x0, 32] ldp q0, q1, [x1, 64] stp q0, q1, [x0, 64] ldp q0, q1, [x1, 96] stp q0, q1, [x0, 96] ret
which is a clear code size win. Speed optimisation heuristics remain unchanged.
2021-09-29 Kyrylo Tkachov kyrylo.tkachov@arm.com
* config/aarch64/aarch64.c (aarch64_expand_cpymem): Count number of emitted operations and adjust heuristic for code size.
2021-09-29 Kyrylo Tkachov kyrylo.tkachov@arm.com
* gcc.target/aarch64/cpymem-size.c: New test. --- gcc/config/aarch64/aarch64.c | 36 ++++++++++++++++++-------- gcc/testsuite/gcc.target/aarch64/cpymem-size.c | 29 +++++++++++++++++++++ 2 files changed, 54 insertions(+), 11 deletions(-)
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c index ac17c1c88fb..a9a1800af53 100644 --- a/gcc/config/aarch64/aarch64.c +++ b/gcc/config/aarch64/aarch64.c @@ -23390,7 +23390,8 @@ aarch64_copy_one_block_and_progress_pointers (rtx *src, rtx *dst, }
/* Expand cpymem, as if from a __builtin_memcpy. Return true if - we succeed, otherwise return false. */ + we succeed, otherwise return false, indicating that a libcall to + memcpy should be emitted. */
bool aarch64_expand_cpymem (rtx *operands) @@ -23407,11 +23408,13 @@ aarch64_expand_cpymem (rtx *operands)
unsigned HOST_WIDE_INT size = INTVAL (operands[2]);
- /* Inline up to 256 bytes when optimizing for speed. */ + /* Try to inline up to 256 bytes. */ unsigned HOST_WIDE_INT max_copy_size = 256;
- if (optimize_function_for_size_p (cfun)) - max_copy_size = 128; + bool size_p = optimize_function_for_size_p (cfun); + + if (size > max_copy_size) + return false;
int copy_bits = 256;
@@ -23421,13 +23424,14 @@ aarch64_expand_cpymem (rtx *operands) || !TARGET_SIMD || (aarch64_tune_params.extra_tuning_flags & AARCH64_EXTRA_TUNE_NO_LDP_STP_QREGS)) - { - copy_bits = 128; - max_copy_size = max_copy_size / 2; - } + copy_bits = 128;
- if (size > max_copy_size) - return false; + /* Emit an inline load+store sequence and count the number of operations + involved. We use a simple count of just the loads and stores emitted + rather than rtx_insn count as all the pointer adjustments and reg copying + in this function will get optimized away later in the pipeline. */ + start_sequence (); + unsigned nops = 0;
base = copy_to_mode_reg (Pmode, XEXP (dst, 0)); dst = adjust_automodify_address (dst, VOIDmode, base, 0); @@ -23456,7 +23460,8 @@ aarch64_expand_cpymem (rtx *operands) cur_mode = V4SImode;
aarch64_copy_one_block_and_progress_pointers (&src, &dst, cur_mode); - + /* A single block copy is 1 load + 1 store. */ + nops += 2; n -= mode_bits;
/* Emit trailing copies using overlapping unaligned accesses - this is @@ -23471,7 +23476,16 @@ aarch64_expand_cpymem (rtx *operands) n = n_bits; } } + rtx_insn *seq = get_insns (); + end_sequence (); + + /* A memcpy libcall in the worst case takes 3 instructions to prepare the + arguments + 1 for the call. */ + unsigned libcall_cost = 4; + if (size_p && libcall_cost < nops) + return false;
+ emit_insn (seq); return true; }
diff --git a/gcc/testsuite/gcc.target/aarch64/cpymem-size.c b/gcc/testsuite/gcc.target/aarch64/cpymem-size.c new file mode 100644 index 00000000000..4d488b74301 --- /dev/null +++ b/gcc/testsuite/gcc.target/aarch64/cpymem-size.c @@ -0,0 +1,29 @@ +/* { dg-do compile } */ +/* { dg-options "-Os" } */ + +#include <stdlib.h> + +/* +** cpy_127: +** mov x2, 127 +** b memcpy +*/ +void +cpy_127 (char *out, char *in) +{ + __builtin_memcpy (out, in, 127); +} + +/* +** cpy_128: +** mov x2, 128 +** b memcpy +*/ +void +cpy_128 (char *out, char *in) +{ + __builtin_memcpy (out, in, 128); +} + +/* { dg-final { check-function-bodies "**" "" "" } } */ + </cut>