[libc] memmove optimizations #70043
Merged
[libc] memmove optimizations #70043
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-libc Author: Dmitry Vyukov (dvyukov) ChangesSee individual commits for description. Full diff: https://github.com/llvm/llvm-project/pull/70043.diff 8 Files Affected:
diff --git a/libc/benchmarks/LibcMemoryBenchmarkMain.cpp b/libc/benchmarks/LibcMemoryBenchmarkMain.cpp
index acd7c30717597a1..bc6fd8b38cb6ddc 100644
--- a/libc/benchmarks/LibcMemoryBenchmarkMain.cpp
+++ b/libc/benchmarks/LibcMemoryBenchmarkMain.cpp
@@ -42,9 +42,15 @@ static cl::opt<std::string>
SizeDistributionName("size-distribution-name",
cl::desc("The name of the distribution to use"));
-static cl::opt<bool>
- SweepMode("sweep-mode",
- cl::desc("If set, benchmark all sizes from 0 to sweep-max-size"));
+static cl::opt<bool> SweepMode(
+ "sweep-mode",
+ cl::desc(
+ "If set, benchmark all sizes from sweep-min-size to sweep-max-size"));
+
+static cl::opt<uint32_t>
+ SweepMinSize("sweep-min-size",
+ cl::desc("The minimum size to use in sweep-mode"),
+ cl::init(0));
static cl::opt<uint32_t>
SweepMaxSize("sweep-max-size",
@@ -185,7 +191,7 @@ struct MemfunctionBenchmarkSweep final : public MemfunctionBenchmarkBase {
BO.InitialIterations = 100;
auto &Measurements = Study.Measurements;
Measurements.reserve(NumTrials * SweepMaxSize);
- for (size_t Size = 0; Size <= SweepMaxSize; ++Size) {
+ for (size_t Size = SweepMinSize; Size <= SweepMaxSize; ++Size) {
CurrentSweepSize = Size;
runTrials(BO, Measurements);
}
diff --git a/libc/src/string/memmove.cpp b/libc/src/string/memmove.cpp
index 7d473afc0b42ee7..a6478629d514027 100644
--- a/libc/src/string/memmove.cpp
+++ b/libc/src/string/memmove.cpp
@@ -15,10 +15,21 @@ namespace LIBC_NAMESPACE {
LLVM_LIBC_FUNCTION(void *, memmove,
(void *dst, const void *src, size_t count)) {
- if (is_disjoint(dst, src, count))
- inline_memcpy(dst, src, count);
- else
- inline_memmove(dst, src, count);
+ // inline_memmove may handle some small sizes as efficiently
+ // as inline_memcpy. For these sizes we may not do is_disjoint check.
+ // This both avoids additional code for the most frequent smaller sizes
+ // and removes code bloat (we don't need the memcpy logic for small sizes).
+ // Here we heavily rely on inlining and dead code elimination: from the first
+ // inline_memmove we should get only handling of small sizes, and from
+ // the second inline_memmove and inline_memcpy we should get only handling
+ // of larger sizes.
+ inline_memmove(dst, src, count, true);
+ if (count >= LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE_SLOW_SIZE) {
+ if (is_disjoint(dst, src, count))
+ inline_memcpy(dst, src, count);
+ else
+ inline_memmove(dst, src, count);
+ }
return dst;
}
diff --git a/libc/src/string/memory_utils/aarch64/inline_memmove.h b/libc/src/string/memory_utils/aarch64/inline_memmove.h
index ca28655c916820c..5e5c23e34be62d1 100644
--- a/libc/src/string/memory_utils/aarch64/inline_memmove.h
+++ b/libc/src/string/memory_utils/aarch64/inline_memmove.h
@@ -18,7 +18,8 @@
namespace LIBC_NAMESPACE {
-LIBC_INLINE void inline_memmove_aarch64(Ptr dst, CPtr src, size_t count) {
+LIBC_INLINE void inline_memmove_aarch64(Ptr dst, CPtr src, size_t count,
+ bool fast_only) {
static_assert(aarch64::kNeon, "aarch64 supports vector types");
using uint128_t = generic_v128;
using uint256_t = generic_v256;
@@ -39,6 +40,8 @@ LIBC_INLINE void inline_memmove_aarch64(Ptr dst, CPtr src, size_t count) {
return generic::Memmove<uint256_t>::head_tail(dst, src, count);
if (count <= 128)
return generic::Memmove<uint512_t>::head_tail(dst, src, count);
+ if (fast_only)
+ return;
if (dst < src) {
generic::Memmove<uint256_t>::align_forward<Arg::Src>(dst, src, count);
return generic::Memmove<uint512_t>::loop_and_tail_forward(dst, src, count);
diff --git a/libc/src/string/memory_utils/generic/builtin.h b/libc/src/string/memory_utils/generic/builtin.h
index 5239329f653b341..1dabc856053d191 100644
--- a/libc/src/string/memory_utils/generic/builtin.h
+++ b/libc/src/string/memory_utils/generic/builtin.h
@@ -26,8 +26,8 @@ inline_memcpy_builtin(Ptr dst, CPtr src, size_t count, size_t offset = 0) {
__builtin_memcpy(dst + offset, src + offset, count);
}
-[[maybe_unused]] LIBC_INLINE void inline_memmove_builtin(Ptr dst, CPtr src,
- size_t count) {
+[[maybe_unused]] LIBC_INLINE void
+inline_memmove_builtin(Ptr dst, CPtr src, size_t count, bool fast_only) {
__builtin_memmove(dst, src, count);
}
diff --git a/libc/src/string/memory_utils/generic/byte_per_byte.h b/libc/src/string/memory_utils/generic/byte_per_byte.h
index a666c5da3136041..89497382aede338 100644
--- a/libc/src/string/memory_utils/generic/byte_per_byte.h
+++ b/libc/src/string/memory_utils/generic/byte_per_byte.h
@@ -29,7 +29,7 @@ inline_memcpy_byte_per_byte(Ptr dst, CPtr src, size_t count,
}
[[maybe_unused]] LIBC_INLINE void
-inline_memmove_byte_per_byte(Ptr dst, CPtr src, size_t count) {
+inline_memmove_byte_per_byte(Ptr dst, CPtr src, size_t count, bool fast_only) {
if (count == 0 || dst == src)
return;
if (dst < src) {
diff --git a/libc/src/string/memory_utils/inline_memmove.h b/libc/src/string/memory_utils/inline_memmove.h
index f72ea24ab538d69..0440bbe94d542e9 100644
--- a/libc/src/string/memory_utils/inline_memmove.h
+++ b/libc/src/string/memory_utils/inline_memmove.h
@@ -14,27 +14,36 @@
#if defined(LIBC_TARGET_ARCH_IS_X86)
#include "src/string/memory_utils/x86_64/inline_memmove.h"
#define LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE inline_memmove_x86
+#define LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE_SLOW_SIZE 129
#elif defined(LIBC_TARGET_ARCH_IS_AARCH64)
#include "src/string/memory_utils/aarch64/inline_memmove.h"
#define LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE inline_memmove_aarch64
+#define LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE_SLOW_SIZE 129
#elif defined(LIBC_TARGET_ARCH_IS_ANY_RISCV)
#include "src/string/memory_utils/riscv/inline_memmove.h"
#define LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE inline_memmove_riscv
+#define LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE_SLOW_SIZE 0
#elif defined(LIBC_TARGET_ARCH_IS_ARM)
#include "src/string/memory_utils/generic/byte_per_byte.h"
#define LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE inline_memmove_byte_per_byte
+#define LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE_SLOW_SIZE 0
#elif defined(LIBC_TARGET_ARCH_IS_GPU)
#include "src/string/memory_utils/generic/builtin.h"
#define LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE inline_memmove_builtin
+#define LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE_SLOW_SIZE 0
#else
#error "Unsupported architecture"
#endif
namespace LIBC_NAMESPACE {
-LIBC_INLINE void inline_memmove(void *dst, const void *src, size_t count) {
+LIBC_INLINE void inline_memmove(void *dst, const void *src, size_t count,
+ bool fast_only = false) {
+ if (LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE_SLOW_SIZE == 0 && fast_only)
+ return;
LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE(reinterpret_cast<Ptr>(dst),
- reinterpret_cast<CPtr>(src), count);
+ reinterpret_cast<CPtr>(src), count,
+ fast_only);
}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/string/memory_utils/riscv/inline_memmove.h b/libc/src/string/memory_utils/riscv/inline_memmove.h
index 1c26917a96d9d18..5e34b2817729972 100644
--- a/libc/src/string/memory_utils/riscv/inline_memmove.h
+++ b/libc/src/string/memory_utils/riscv/inline_memmove.h
@@ -17,9 +17,11 @@
namespace LIBC_NAMESPACE {
-[[maybe_unused]] LIBC_INLINE void
-inline_memmove_riscv(Ptr __restrict dst, CPtr __restrict src, size_t count) {
- return inline_memmove_byte_per_byte(dst, src, count);
+[[maybe_unused]] LIBC_INLINE void inline_memmove_riscv(Ptr __restrict dst,
+ CPtr __restrict src,
+ size_t count,
+ bool fast_only) {
+ return inline_memmove_byte_per_byte(dst, src, count, fast_only);
}
} // namespace LIBC_NAMESPACE
diff --git a/libc/src/string/memory_utils/x86_64/inline_memmove.h b/libc/src/string/memory_utils/x86_64/inline_memmove.h
index 95ad07f75219581..ee397c63471f1ad 100644
--- a/libc/src/string/memory_utils/x86_64/inline_memmove.h
+++ b/libc/src/string/memory_utils/x86_64/inline_memmove.h
@@ -18,40 +18,61 @@
namespace LIBC_NAMESPACE {
-LIBC_INLINE void inline_memmove_x86(Ptr dst, CPtr src, size_t count) {
+LIBC_INLINE void inline_memmove_x86(Ptr dst, CPtr src, size_t count,
+ bool fast_only) {
#if defined(__AVX512F__)
+ constexpr size_t vector_size = 64;
using uint128_t = generic_v128;
using uint256_t = generic_v256;
using uint512_t = generic_v512;
#elif defined(__AVX__)
+ constexpr size_t vector_size = 32;
using uint128_t = generic_v128;
using uint256_t = generic_v256;
using uint512_t = cpp::array<generic_v256, 2>;
#elif defined(__SSE2__)
+ constexpr size_t vector_size = 16;
using uint128_t = generic_v128;
using uint256_t = cpp::array<generic_v128, 2>;
using uint512_t = cpp::array<generic_v128, 4>;
#else
+ constexpr size_t vector_size = 8;
using uint128_t = cpp::array<uint64_t, 2>;
using uint256_t = cpp::array<uint64_t, 4>;
using uint512_t = cpp::array<uint64_t, 8>;
#endif
+ (void)vector_size;
if (count == 0)
return;
if (count == 1)
return generic::Memmove<uint8_t>::block(dst, src);
- if (count <= 4)
- return generic::Memmove<uint16_t>::head_tail(dst, src, count);
- if (count <= 8)
+ if (count == 2)
+ return generic::Memmove<uint16_t>::block(dst, src);
+ if (count == 3)
+ return generic::Memmove<cpp::array<uint8_t, 3>>::block(dst, src);
+ if (count == 4)
+ return generic::Memmove<uint32_t>::block(dst, src);
+ if (count < 8)
return generic::Memmove<uint32_t>::head_tail(dst, src, count);
- if (count <= 16)
+ // If count is equal to a power of 2, we can handle it as head-tail
+ // of both smaller size and larger size (head-tail are either
+ // non-overlapping for smaller size, or completely collapsed
+ // for larger size). It seems to be more profitable to do the copy
+ // with the larger size, if it's natively supported (e.g. doing
+ // 2 collapsed 32-byte moves for count=64 if AVX2 is supported).
+ // But it's not profitable to use larger size if it's not natively
+ // supported: we will both use more instructions and handle fewer
+ // sizes in earlier branches.
+ if (count < 16 + (vector_size <= sizeof(uint64_t)))
return generic::Memmove<uint64_t>::head_tail(dst, src, count);
- if (count <= 32)
+ if (count < 32 + (vector_size <= sizeof(uint128_t)))
return generic::Memmove<uint128_t>::head_tail(dst, src, count);
- if (count <= 64)
+ if (count < 64 + (vector_size <= sizeof(uint256_t)))
return generic::Memmove<uint256_t>::head_tail(dst, src, count);
if (count <= 128)
return generic::Memmove<uint512_t>::head_tail(dst, src, count);
+ if (fast_only)
+ return;
if (dst < src) {
generic::Memmove<uint256_t>::align_forward<Arg::Src>(dst, src, count);
return generic::Memmove<uint512_t>::loop_and_tail_forward(dst, src, count);
|
dvyukov
force-pushed
the
dvyukov-memmove-small-size
branch
from
9914fbe to
9e2694f
Compare
|
Oh, "Rebase and merge" is not enabled, that's sad. |
Yeah we'll need to push individual commits as separate PRs. The first one -concerning the benchmark- LGTM can go in independently. A word of caution though on the sweeping benchmark. It exercises the same size many times in a row to accumulate measurement precision, as a consequence it will benefit from branch prediction and gives idealized results that may not be realized in production. It gives an upper bound though. For the second one I think it's a good idea to factor in small sizes. The code duplication is suboptimal. But I'm not sure if introducing the The third patch looks good on the intent but I don't get the boolean addition in the This is either |
A slightly cleaner way to do it may be to return true/false from inline_memmove instead of the separate LIBC_SRC_STRING_MEMORY_UTILS_MEMMOVE_SLOW_SIZE. I did not do it only because it breaks all of the nice returns of voids :) |
Does it look better?
Yes. Does the following look better? |
|
And how about splitting
We'd have the following code Each arch would be responsible for providing the implementations side by side so we can visually check they complement each other. For arch where we don't have a fast path Inlining would remove the branch altogether. WDTY? |
|
I missed your reply before sending mine...
It's more explicit yes. |
|
Moved the benchmarks change to #70302. |
dvyukov
force-pushed
the
dvyukov-memmove-small-size
branch
from
9e2694f to
65dfb6d
Compare
|
All done. PTAL. |
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
1. Remove is_disjoint check for smaller sizes and reduce code bloat.
inline_memmove may handle some small sizes as efficiently
as inline_memcpy. For these sizes we may not do is_disjoint check.
This both avoids additional code for the most frequent smaller sizes
and removes code bloat (we don't need the memcpy logic for small sizes).
Here we heavily rely on inlining and dead code elimination: from the first
inline_memmove we should get only handling of small sizes, and from
the second inline_memmove and inline_memcpy we should get only handling
of larger sizes.
2. Use the memcpy thresholds for memmove.
Memcpy thresholds were more carefully tuned.
This becomes more important since we use memmove
for all small sizes always now.
3. Fix boundary conditions for sizes = 16/32/64.
See the added comment for explanations.
Memmove function size drops from 885 to 715 bytes
due to removed duplication.
│ baseline │ small-size │
│ sec/op │ sec/op vs base │
memmove/Google_A 3.208n ± 0% 2.911n ± 0% -9.25% (n=100)
memmove/Google_B 4.113n ± 1% 3.428n ± 0% -16.65% (n=100)
memmove/Google_D 5.838n ± 0% 4.158n ± 0% -28.78% (n=100)
memmove/Google_S 4.712n ± 1% 3.899n ± 0% -17.25% (n=100)
memmove/Google_U 3.609n ± 0% 3.247n ± 1% -10.02% (n=100)
memmove/0 2.982n ± 0% 2.169n ± 0% -27.26% (n=50)
memmove/1 3.253n ± 0% 2.168n ± 0% -33.34% (n=50)
memmove/2 3.255n ± 0% 2.169n ± 0% -33.38% (n=50)
memmove/3 3.259n ± 2% 2.175n ± 0% -33.27% (p=0.000 n=50)
memmove/4 3.259n ± 0% 2.168n ± 5% -33.46% (p=0.000 n=50)
memmove/5 2.488n ± 0% 1.926n ± 0% -22.57% (p=0.000 n=50)
memmove/6 2.490n ± 0% 1.928n ± 0% -22.59% (p=0.000 n=50)
memmove/7 2.492n ± 0% 1.927n ± 0% -22.65% (p=0.000 n=50)
memmove/8 2.737n ± 0% 2.711n ± 0% -0.97% (p=0.000 n=50)
memmove/9 2.736n ± 0% 2.711n ± 0% -0.94% (p=0.000 n=50)
memmove/10 2.739n ± 0% 2.711n ± 0% -1.04% (p=0.000 n=50)
memmove/11 2.740n ± 0% 2.711n ± 0% -1.07% (p=0.000 n=50)
memmove/12 2.740n ± 0% 2.711n ± 0% -1.09% (p=0.000 n=50)
memmove/13 2.744n ± 0% 2.711n ± 0% -1.22% (p=0.000 n=50)
memmove/14 2.742n ± 0% 2.711n ± 0% -1.14% (p=0.000 n=50)
memmove/15 2.742n ± 0% 2.711n ± 0% -1.15% (p=0.000 n=50)
memmove/16 2.997n ± 0% 2.981n ± 0% -0.52% (p=0.000 n=50)
memmove/17 2.998n ± 0% 2.981n ± 0% -0.55% (p=0.000 n=50)
memmove/18 2.998n ± 0% 2.981n ± 0% -0.55% (p=0.000 n=50)
memmove/19 2.999n ± 0% 2.982n ± 0% -0.59% (p=0.000 n=50)
memmove/20 2.998n ± 0% 2.981n ± 0% -0.55% (p=0.000 n=50)
memmove/21 3.000n ± 0% 2.981n ± 0% -0.61% (p=0.000 n=50)
memmove/22 3.002n ± 0% 2.981n ± 0% -0.68% (p=0.000 n=50)
memmove/23 3.002n ± 0% 2.981n ± 0% -0.67% (p=0.000 n=50)
memmove/24 3.002n ± 0% 2.981n ± 0% -0.70% (n=50)
memmove/25 3.002n ± 0% 2.981n ± 0% -0.68% (p=0.000 n=50)
memmove/26 3.004n ± 0% 2.982n ± 0% -0.74% (p=0.000 n=50)
memmove/27 3.005n ± 0% 2.981n ± 0% -0.79% (n=50)
memmove/28 3.005n ± 0% 2.982n ± 0% -0.77% (n=50)
memmove/29 3.009n ± 0% 2.981n ± 0% -0.92% (n=50)
memmove/30 3.008n ± 0% 2.981n ± 0% -0.89% (n=50)
memmove/31 3.007n ± 0% 2.982n ± 0% -0.86% (n=50)
memmove/32 3.540n ± 0% 2.998n ± 0% -15.31% (p=0.000 n=50)
memmove/33 3.544n ± 0% 2.997n ± 0% -15.44% (p=0.000 n=50)
memmove/34 3.546n ± 0% 2.999n ± 0% -15.42% (n=50)
memmove/35 3.545n ± 0% 2.999n ± 0% -15.40% (n=50)
memmove/36 3.548n ± 0% 2.998n ± 0% -15.52% (p=0.000 n=50)
memmove/37 3.546n ± 0% 3.000n ± 0% -15.41% (n=50)
memmove/38 3.549n ± 0% 2.999n ± 0% -15.49% (p=0.000 n=50)
memmove/39 3.549n ± 0% 2.999n ± 0% -15.48% (p=0.000 n=50)
memmove/40 3.549n ± 0% 3.000n ± 0% -15.46% (p=0.000 n=50)
memmove/41 3.550n ± 0% 3.001n ± 0% -15.47% (n=50)
memmove/42 3.549n ± 0% 3.001n ± 0% -15.43% (n=50)
memmove/43 3.552n ± 0% 3.001n ± 0% -15.52% (p=0.000 n=50)
memmove/44 3.552n ± 0% 3.001n ± 0% -15.51% (n=50)
memmove/45 3.552n ± 0% 3.002n ± 0% -15.48% (n=50)
memmove/46 3.554n ± 0% 3.001n ± 0% -15.55% (p=0.000 n=50)
memmove/47 3.556n ± 0% 3.002n ± 0% -15.58% (p=0.000 n=50)
memmove/48 3.555n ± 0% 3.003n ± 0% -15.54% (n=50)
memmove/49 3.557n ± 0% 3.002n ± 0% -15.59% (p=0.000 n=50)
memmove/50 3.557n ± 0% 3.004n ± 0% -15.55% (p=0.000 n=50)
memmove/51 3.556n ± 0% 3.004n ± 0% -15.53% (p=0.000 n=50)
memmove/52 3.561n ± 0% 3.004n ± 0% -15.65% (p=0.000 n=50)
memmove/53 3.558n ± 0% 3.004n ± 0% -15.57% (p=0.000 n=50)
memmove/54 3.561n ± 0% 3.005n ± 0% -15.62% (n=50)
memmove/55 3.560n ± 0% 3.006n ± 0% -15.57% (n=50)
memmove/56 3.562n ± 0% 3.006n ± 0% -15.60% (p=0.000 n=50)
memmove/57 3.563n ± 0% 3.006n ± 0% -15.64% (n=50)
memmove/58 3.565n ± 0% 3.007n ± 0% -15.64% (p=0.000 n=50)
memmove/59 3.564n ± 0% 3.006n ± 0% -15.66% (p=0.000 n=50)
memmove/60 3.570n ± 0% 3.008n ± 0% -15.74% (p=0.000 n=50)
memmove/61 3.566n ± 0% 3.009n ± 0% -15.63% (p=0.000 n=50)
memmove/62 3.567n ± 0% 3.007n ± 0% -15.70% (p=0.000 n=50)
memmove/63 3.568n ± 0% 3.008n ± 0% -15.71% (p=0.000 n=50)
memmove/64 4.104n ± 0% 3.008n ± 0% -26.70% (p=0.000 n=50)
memmove/65 4.126n ± 0% 3.662n ± 0% -11.26% (p=0.000 n=50)
memmove/66 4.128n ± 0% 3.662n ± 0% -11.29% (n=50)
memmove/67 4.129n ± 0% 3.662n ± 0% -11.31% (n=50)
memmove/68 4.129n ± 0% 3.661n ± 0% -11.33% (p=0.000 n=50)
memmove/69 4.130n ± 0% 3.662n ± 0% -11.34% (p=0.000 n=50)
memmove/70 4.130n ± 0% 3.662n ± 0% -11.33% (n=50)
memmove/71 4.132n ± 0% 3.662n ± 0% -11.38% (p=0.000 n=50)
memmove/72 4.131n ± 0% 3.661n ± 0% -11.39% (n=50)
memmove/73 4.135n ± 0% 3.661n ± 0% -11.45% (p=0.000 n=50)
memmove/74 4.137n ± 0% 3.662n ± 0% -11.49% (n=50)
memmove/75 4.138n ± 0% 3.662n ± 0% -11.51% (p=0.000 n=50)
memmove/76 4.139n ± 0% 3.661n ± 0% -11.56% (p=0.000 n=50)
memmove/77 4.136n ± 0% 3.662n ± 0% -11.47% (p=0.000 n=50)
memmove/78 4.143n ± 0% 3.661n ± 0% -11.62% (p=0.000 n=50)
memmove/79 4.142n ± 0% 3.661n ± 0% -11.60% (n=50)
memmove/80 4.142n ± 0% 3.661n ± 0% -11.62% (p=0.000 n=50)
memmove/81 4.140n ± 0% 3.661n ± 0% -11.57% (n=50)
memmove/82 4.146n ± 0% 3.661n ± 0% -11.69% (n=50)
memmove/83 4.143n ± 0% 3.661n ± 0% -11.63% (p=0.000 n=50)
memmove/84 4.143n ± 0% 3.661n ± 0% -11.63% (n=50)
memmove/85 4.147n ± 0% 3.661n ± 0% -11.73% (p=0.000 n=50)
memmove/86 4.142n ± 0% 3.661n ± 0% -11.62% (p=0.000 n=50)
memmove/87 4.147n ± 0% 3.661n ± 0% -11.72% (p=0.000 n=50)
memmove/88 4.148n ± 0% 3.661n ± 0% -11.74% (n=50)
memmove/89 4.152n ± 0% 3.661n ± 0% -11.84% (n=50)
memmove/90 4.151n ± 0% 3.661n ± 0% -11.81% (n=50)
memmove/91 4.150n ± 0% 3.661n ± 0% -11.78% (n=50)
memmove/92 4.153n ± 0% 3.661n ± 0% -11.86% (n=50)
memmove/93 4.158n ± 0% 3.661n ± 0% -11.95% (n=50)
memmove/94 4.157n ± 0% 3.661n ± 0% -11.95% (p=0.000 n=50)
memmove/95 4.155n ± 0% 3.661n ± 0% -11.90% (p=0.000 n=50)
memmove/96 4.149n ± 0% 3.660n ± 0% -11.79% (n=50)
memmove/97 4.157n ± 0% 3.661n ± 0% -11.94% (n=50)
memmove/98 4.157n ± 0% 3.661n ± 0% -11.94% (n=50)
memmove/99 4.168n ± 0% 3.661n ± 0% -12.17% (p=0.000 n=50)
memmove/100 4.159n ± 0% 3.660n ± 0% -12.00% (p=0.000 n=50)
memmove/101 4.161n ± 0% 3.660n ± 0% -12.03% (p=0.000 n=50)
memmove/102 4.165n ± 0% 3.660n ± 0% -12.12% (p=0.000 n=50)
memmove/103 4.164n ± 0% 3.661n ± 0% -12.08% (n=50)
memmove/104 4.164n ± 0% 3.660n ± 0% -12.11% (n=50)
memmove/105 4.165n ± 0% 3.660n ± 0% -12.12% (p=0.000 n=50)
memmove/106 4.166n ± 0% 3.660n ± 0% -12.15% (n=50)
memmove/107 4.171n ± 0% 3.660n ± 1% -12.26% (p=0.000 n=50)
memmove/108 4.173n ± 0% 3.660n ± 0% -12.30% (p=0.000 n=50)
memmove/109 4.170n ± 0% 3.660n ± 0% -12.24% (n=50)
memmove/110 4.174n ± 0% 3.660n ± 0% -12.31% (n=50)
memmove/111 4.176n ± 0% 3.660n ± 0% -12.35% (p=0.000 n=50)
memmove/112 4.174n ± 0% 3.659n ± 0% -12.34% (p=0.000 n=50)
memmove/113 4.176n ± 0% 3.660n ± 0% -12.35% (n=50)
memmove/114 4.182n ± 0% 3.660n ± 0% -12.49% (n=50)
memmove/115 4.185n ± 0% 3.660n ± 0% -12.55% (n=50)
memmove/116 4.184n ± 0% 3.659n ± 0% -12.54% (n=50)
memmove/117 4.182n ± 0% 3.660n ± 0% -12.50% (n=50)
memmove/118 4.188n ± 0% 3.660n ± 0% -12.61% (n=50)
memmove/119 4.186n ± 0% 3.660n ± 0% -12.57% (p=0.000 n=50)
memmove/120 4.189n ± 0% 3.659n ± 0% -12.63% (n=50)
memmove/121 4.187n ± 0% 3.660n ± 0% -12.60% (n=50)
memmove/122 4.186n ± 0% 3.660n ± 0% -12.58% (n=50)
memmove/123 4.187n ± 0% 3.660n ± 0% -12.60% (n=50)
memmove/124 4.189n ± 0% 3.659n ± 0% -12.65% (n=50)
memmove/125 4.195n ± 0% 3.659n ± 0% -12.78% (n=50)
memmove/126 4.197n ± 0% 3.659n ± 0% -12.81% (n=50)
memmove/127 4.194n ± 0% 3.659n ± 0% -12.75% (n=50)
memmove/128 5.035n ± 0% 3.659n ± 0% -27.32% (n=50)
memmove/129 5.127n ± 0% 5.164n ± 0% +0.73% (p=0.000 n=50)
memmove/130 5.130n ± 0% 5.176n ± 0% +0.88% (p=0.000 n=50)
memmove/131 5.127n ± 0% 5.180n ± 0% +1.05% (p=0.000 n=50)
memmove/132 5.131n ± 0% 5.169n ± 0% +0.75% (p=0.000 n=50)
memmove/133 5.137n ± 0% 5.179n ± 0% +0.81% (p=0.000 n=50)
memmove/134 5.140n ± 0% 5.178n ± 0% +0.74% (p=0.000 n=50)
memmove/135 5.141n ± 0% 5.187n ± 0% +0.88% (p=0.000 n=50)
memmove/136 5.133n ± 0% 5.184n ± 0% +0.99% (p=0.000 n=50)
memmove/137 5.148n ± 0% 5.186n ± 0% +0.73% (p=0.000 n=50)
memmove/138 5.143n ± 0% 5.189n ± 0% +0.88% (p=0.000 n=50)
memmove/139 5.142n ± 0% 5.192n ± 0% +0.97% (p=0.000 n=50)
memmove/140 5.141n ± 0% 5.192n ± 0% +1.01% (p=0.000 n=50)
memmove/141 5.155n ± 0% 5.188n ± 0% +0.64% (p=0.000 n=50)
memmove/142 5.146n ± 0% 5.192n ± 0% +0.90% (p=0.000 n=50)
memmove/143 5.142n ± 0% 5.203n ± 0% +1.19% (p=0.000 n=50)
memmove/144 5.146n ± 0% 5.197n ± 0% +0.99% (p=0.000 n=50)
memmove/145 5.146n ± 0% 5.196n ± 0% +0.97% (p=0.000 n=50)
memmove/146 5.151n ± 0% 5.207n ± 0% +1.10% (p=0.000 n=50)
memmove/147 5.151n ± 0% 5.205n ± 0% +1.06% (p=0.000 n=50)
memmove/148 5.156n ± 0% 5.190n ± 0% +0.66% (p=0.000 n=50)
memmove/149 5.158n ± 0% 5.212n ± 0% +1.04% (p=0.000 n=50)
memmove/150 5.160n ± 0% 5.203n ± 0% +0.84% (p=0.000 n=50)
memmove/151 5.167n ± 0% 5.210n ± 0% +0.83% (p=0.000 n=50)
memmove/152 5.157n ± 0% 5.206n ± 0% +0.94% (p=0.000 n=50)
memmove/153 5.170n ± 0% 5.211n ± 0% +0.80% (p=0.000 n=50)
memmove/154 5.169n ± 0% 5.222n ± 0% +1.02% (p=0.000 n=50)
memmove/155 5.171n ± 0% 5.215n ± 0% +0.87% (p=0.000 n=50)
memmove/156 5.174n ± 0% 5.214n ± 0% +0.78% (p=0.000 n=50)
memmove/157 5.171n ± 0% 5.218n ± 0% +0.92% (p=0.000 n=50)
memmove/158 5.168n ± 0% 5.224n ± 0% +1.09% (p=0.000 n=50)
memmove/159 5.179n ± 0% 5.218n ± 0% +0.76% (p=0.000 n=50)
memmove/160 5.170n ± 0% 5.219n ± 0% +0.95% (p=0.000 n=50)
memmove/161 5.187n ± 0% 5.220n ± 0% +0.64% (p=0.000 n=50)
memmove/162 5.189n ± 0% 5.234n ± 0% +0.86% (p=0.000 n=50)
memmove/163 5.199n ± 0% 5.250n ± 0% +0.99% (p=0.000 n=50)
memmove/164 5.205n ± 0% 5.260n ± 0% +1.04% (p=0.000 n=50)
memmove/165 5.208n ± 0% 5.261n ± 0% +1.01% (p=0.000 n=50)
memmove/166 5.227n ± 0% 5.275n ± 0% +0.91% (p=0.000 n=50)
memmove/167 5.233n ± 0% 5.281n ± 0% +0.92% (p=0.000 n=50)
memmove/168 5.236n ± 0% 5.295n ± 0% +1.12% (p=0.000 n=50)
memmove/169 5.256n ± 0% 5.297n ± 0% +0.79% (p=0.000 n=50)
memmove/170 5.259n ± 0% 5.302n ± 0% +0.80% (p=0.000 n=50)
memmove/171 5.269n ± 0% 5.321n ± 0% +0.97% (p=0.000 n=50)
memmove/172 5.266n ± 0% 5.318n ± 0% +0.98% (p=0.000 n=50)
memmove/173 5.272n ± 0% 5.330n ± 0% +1.09% (p=0.000 n=50)
memmove/174 5.284n ± 0% 5.331n ± 0% +0.89% (p=0.000 n=50)
memmove/175 5.284n ± 0% 5.322n ± 0% +0.72% (p=0.000 n=50)
memmove/176 5.298n ± 0% 5.337n ± 0% +0.74% (p=0.000 n=50)
memmove/177 5.282n ± 0% 5.338n ± 0% +1.04% (p=0.000 n=50)
memmove/178 5.299n ± 0% 5.337n ± 0% +0.71% (p=0.000 n=50)
memmove/179 5.296n ± 0% 5.343n ± 0% +0.88% (p=0.000 n=50)
memmove/180 5.292n ± 0% 5.343n ± 0% +0.97% (p=0.000 n=50)
memmove/181 5.303n ± 0% 5.335n ± 0% +0.60% (p=0.000 n=50)
memmove/182 5.305n ± 0% 5.338n ± 0% +0.62% (p=0.000 n=50)
memmove/183 5.298n ± 0% 5.329n ± 0% +0.59% (p=0.000 n=50)
memmove/184 5.299n ± 0% 5.333n ± 0% +0.64% (p=0.000 n=50)
memmove/185 5.291n ± 0% 5.330n ± 0% +0.73% (p=0.000 n=50)
memmove/186 5.296n ± 0% 5.332n ± 0% +0.68% (p=0.000 n=50)
memmove/187 5.297n ± 0% 5.320n ± 0% +0.44% (p=0.000 n=50)
memmove/188 5.286n ± 0% 5.314n ± 0% +0.53% (p=0.000 n=50)
memmove/189 5.293n ± 0% 5.318n ± 0% +0.46% (p=0.000 n=50)
memmove/190 5.294n ± 0% 5.318n ± 0% +0.45% (p=0.000 n=50)
memmove/191 5.292n ± 0% 5.314n ± 0% +0.40% (p=0.032 n=50)
memmove/192 5.272n ± 0% 5.304n ± 0% +0.60% (p=0.000 n=50)
memmove/193 5.279n ± 0% 5.310n ± 0% +0.57% (p=0.000 n=50)
memmove/194 5.294n ± 0% 5.308n ± 0% +0.26% (p=0.018 n=50)
memmove/195 5.302n ± 0% 5.311n ± 0% +0.18% (p=0.010 n=50)
memmove/196 5.301n ± 0% 5.316n ± 0% +0.28% (p=0.023 n=50)
memmove/197 5.302n ± 0% 5.327n ± 0% +0.47% (p=0.000 n=50)
memmove/198 5.310n ± 0% 5.326n ± 0% +0.30% (p=0.003 n=50)
memmove/199 5.303n ± 0% 5.319n ± 0% +0.30% (p=0.009 n=50)
memmove/200 5.312n ± 0% 5.330n ± 0% +0.35% (p=0.001 n=50)
memmove/201 5.307n ± 0% 5.333n ± 0% +0.50% (p=0.000 n=50)
memmove/202 5.311n ± 0% 5.334n ± 0% +0.44% (p=0.000 n=50)
memmove/203 5.313n ± 0% 5.335n ± 0% +0.41% (p=0.006 n=50)
memmove/204 5.312n ± 0% 5.332n ± 0% +0.36% (p=0.002 n=50)
memmove/205 5.318n ± 0% 5.345n ± 0% +0.50% (p=0.000 n=50)
memmove/206 5.311n ± 0% 5.333n ± 0% +0.42% (p=0.002 n=50)
memmove/207 5.310n ± 0% 5.338n ± 0% +0.52% (p=0.000 n=50)
memmove/208 5.319n ± 0% 5.341n ± 0% +0.40% (p=0.004 n=50)
memmove/209 5.330n ± 0% 5.346n ± 0% +0.30% (p=0.004 n=50)
memmove/210 5.329n ± 0% 5.349n ± 0% +0.38% (p=0.002 n=50)
memmove/211 5.318n ± 0% 5.340n ± 0% +0.41% (p=0.000 n=50)
memmove/212 5.339n ± 0% 5.343n ± 0% ~ (p=0.396 n=50)
memmove/213 5.329n ± 0% 5.343n ± 0% +0.25% (p=0.017 n=50)
memmove/214 5.339n ± 0% 5.358n ± 0% +0.35% (p=0.035 n=50)
memmove/215 5.342n ± 0% 5.346n ± 0% ~ (p=0.063 n=50)
memmove/216 5.338n ± 0% 5.359n ± 0% +0.39% (p=0.002 n=50)
memmove/217 5.341n ± 0% 5.362n ± 0% +0.39% (p=0.015 n=50)
memmove/218 5.354n ± 0% 5.373n ± 0% +0.36% (p=0.041 n=50)
memmove/219 5.352n ± 0% 5.362n ± 0% ~ (p=0.143 n=50)
memmove/220 5.344n ± 0% 5.370n ± 0% +0.50% (p=0.001 n=50)
memmove/221 5.345n ± 0% 5.373n ± 0% +0.53% (p=0.000 n=50)
memmove/222 5.348n ± 0% 5.360n ± 0% +0.23% (p=0.014 n=50)
memmove/223 5.354n ± 0% 5.377n ± 0% +0.43% (p=0.024 n=50)
memmove/224 5.352n ± 0% 5.363n ± 0% ~ (p=0.052 n=50)
memmove/225 5.372n ± 0% 5.380n ± 0% ~ (p=0.481 n=50)
memmove/226 5.368n ± 0% 5.386n ± 0% +0.34% (p=0.004 n=50)
memmove/227 5.386n ± 0% 5.402n ± 0% +0.29% (p=0.028 n=50)
memmove/228 5.400n ± 0% 5.408n ± 0% ~ (p=0.174 n=50)
memmove/229 5.423n ± 0% 5.427n ± 0% ~ (p=0.444 n=50)
memmove/230 5.411n ± 0% 5.429n ± 0% +0.33% (p=0.020 n=50)
memmove/231 5.420n ± 0% 5.433n ± 0% +0.24% (p=0.034 n=50)
memmove/232 5.435n ± 0% 5.441n ± 0% ~ (p=0.235 n=50)
memmove/233 5.446n ± 0% 5.462n ± 0% ~ (p=0.590 n=50)
memmove/234 5.467n ± 0% 5.461n ± 0% ~ (p=0.921 n=50)
memmove/235 5.472n ± 0% 5.478n ± 0% ~ (p=0.883 n=50)
memmove/236 5.466n ± 0% 5.478n ± 0% ~ (p=0.324 n=50)
memmove/237 5.471n ± 0% 5.489n ± 0% ~ (p=0.132 n=50)
memmove/238 5.485n ± 0% 5.489n ± 0% ~ (p=0.460 n=50)
memmove/239 5.484n ± 0% 5.488n ± 0% ~ (p=0.833 n=50)
memmove/240 5.483n ± 0% 5.495n ± 0% ~ (p=0.095 n=50)
memmove/241 5.498n ± 0% 5.514n ± 0% ~ (p=0.077 n=50)
memmove/242 5.518n ± 0% 5.517n ± 0% ~ (p=0.481 n=50)
memmove/243 5.514n ± 0% 5.511n ± 0% ~ (p=0.503 n=50)
memmove/244 5.510n ± 0% 5.497n ± 0% -0.24% (p=0.038 n=50)
memmove/245 5.516n ± 0% 5.505n ± 0% ~ (p=0.317 n=50)
memmove/246 5.513n ± 1% 5.494n ± 0% ~ (p=0.147 n=50)
memmove/247 5.518n ± 0% 5.499n ± 0% -0.36% (p=0.011 n=50)
memmove/248 5.503n ± 0% 5.492n ± 0% ~ (p=0.267 n=50)
memmove/249 5.498n ± 0% 5.497n ± 0% ~ (p=0.765 n=50)
memmove/250 5.485n ± 0% 5.493n ± 0% ~ (p=0.348 n=50)
memmove/251 5.503n ± 0% 5.482n ± 0% -0.37% (p=0.013 n=50)
memmove/252 5.497n ± 0% 5.485n ± 0% ~ (p=0.077 n=50)
memmove/253 5.489n ± 0% 5.496n ± 0% ~ (p=0.850 n=50)
memmove/254 5.497n ± 0% 5.491n ± 0% ~ (p=0.548 n=50)
memmove/255 5.484n ± 1% 5.494n ± 0% ~ (p=0.888 n=50)
memmove/256 6.952n ± 0% 7.676n ± 0% +10.41% (p=0.000 n=50)
geomean 4.406n 4.127n -6.33%
dvyukov
force-pushed
the
dvyukov-memmove-small-size
branch
from
65dfb6d to
7a694d9
Compare
zahiraam
pushed a commit
to zahiraam/llvm-project
that referenced
this pull request
Oct 26, 2023
1. Remove is_disjoint check for smaller sizes and reduce code bloat.
inline_memmove may handle some small sizes as efficiently
as inline_memcpy. For these sizes we may not do is_disjoint check.
This both avoids additional code for the most frequent smaller sizes
and removes code bloat (we don't need the memcpy logic for small sizes).
Here we heavily rely on inlining and dead code elimination: from the
first
inline_memmove we should get only handling of small sizes, and from
the second inline_memmove and inline_memcpy we should get only handling
of larger sizes.
2. Use the memcpy thresholds for memmove.
Memcpy thresholds were more carefully tuned.
This becomes more important since we use memmove
for all small sizes always now.
3. Fix boundary conditions for sizes = 16/32/64.
See the added comment for explanations.
Memmove function size drops from 885 to 715 bytes
due to removed duplication.
```
│ baseline │ small-size │
│ sec/op │ sec/op vs base │
memmove/Google_A 3.208n ± 0% 2.911n ± 0% -9.25% (n=100)
memmove/Google_B 4.113n ± 1% 3.428n ± 0% -16.65% (n=100)
memmove/Google_D 5.838n ± 0% 4.158n ± 0% -28.78% (n=100)
memmove/Google_S 4.712n ± 1% 3.899n ± 0% -17.25% (n=100)
memmove/Google_U 3.609n ± 0% 3.247n ± 1% -10.02% (n=100)
memmove/0 2.982n ± 0% 2.169n ± 0% -27.26% (n=50)
memmove/1 3.253n ± 0% 2.168n ± 0% -33.34% (n=50)
memmove/2 3.255n ± 0% 2.169n ± 0% -33.38% (n=50)
memmove/3 3.259n ± 2% 2.175n ± 0% -33.27% (p=0.000 n=50)
memmove/4 3.259n ± 0% 2.168n ± 5% -33.46% (p=0.000 n=50)
memmove/5 2.488n ± 0% 1.926n ± 0% -22.57% (p=0.000 n=50)
memmove/6 2.490n ± 0% 1.928n ± 0% -22.59% (p=0.000 n=50)
memmove/7 2.492n ± 0% 1.927n ± 0% -22.65% (p=0.000 n=50)
memmove/8 2.737n ± 0% 2.711n ± 0% -0.97% (p=0.000 n=50)
memmove/9 2.736n ± 0% 2.711n ± 0% -0.94% (p=0.000 n=50)
memmove/10 2.739n ± 0% 2.711n ± 0% -1.04% (p=0.000 n=50)
memmove/11 2.740n ± 0% 2.711n ± 0% -1.07% (p=0.000 n=50)
memmove/12 2.740n ± 0% 2.711n ± 0% -1.09% (p=0.000 n=50)
memmove/13 2.744n ± 0% 2.711n ± 0% -1.22% (p=0.000 n=50)
memmove/14 2.742n ± 0% 2.711n ± 0% -1.14% (p=0.000 n=50)
memmove/15 2.742n ± 0% 2.711n ± 0% -1.15% (p=0.000 n=50)
memmove/16 2.997n ± 0% 2.981n ± 0% -0.52% (p=0.000 n=50)
memmove/17 2.998n ± 0% 2.981n ± 0% -0.55% (p=0.000 n=50)
memmove/18 2.998n ± 0% 2.981n ± 0% -0.55% (p=0.000 n=50)
memmove/19 2.999n ± 0% 2.982n ± 0% -0.59% (p=0.000 n=50)
memmove/20 2.998n ± 0% 2.981n ± 0% -0.55% (p=0.000 n=50)
memmove/21 3.000n ± 0% 2.981n ± 0% -0.61% (p=0.000 n=50)
memmove/22 3.002n ± 0% 2.981n ± 0% -0.68% (p=0.000 n=50)
memmove/23 3.002n ± 0% 2.981n ± 0% -0.67% (p=0.000 n=50)
memmove/24 3.002n ± 0% 2.981n ± 0% -0.70% (n=50)
memmove/25 3.002n ± 0% 2.981n ± 0% -0.68% (p=0.000 n=50)
memmove/26 3.004n ± 0% 2.982n ± 0% -0.74% (p=0.000 n=50)
memmove/27 3.005n ± 0% 2.981n ± 0% -0.79% (n=50)
memmove/28 3.005n ± 0% 2.982n ± 0% -0.77% (n=50)
memmove/29 3.009n ± 0% 2.981n ± 0% -0.92% (n=50)
memmove/30 3.008n ± 0% 2.981n ± 0% -0.89% (n=50)
memmove/31 3.007n ± 0% 2.982n ± 0% -0.86% (n=50)
memmove/32 3.540n ± 0% 2.998n ± 0% -15.31% (p=0.000 n=50)
memmove/33 3.544n ± 0% 2.997n ± 0% -15.44% (p=0.000 n=50)
memmove/34 3.546n ± 0% 2.999n ± 0% -15.42% (n=50)
memmove/35 3.545n ± 0% 2.999n ± 0% -15.40% (n=50)
memmove/36 3.548n ± 0% 2.998n ± 0% -15.52% (p=0.000 n=50)
memmove/37 3.546n ± 0% 3.000n ± 0% -15.41% (n=50)
memmove/38 3.549n ± 0% 2.999n ± 0% -15.49% (p=0.000 n=50)
memmove/39 3.549n ± 0% 2.999n ± 0% -15.48% (p=0.000 n=50)
memmove/40 3.549n ± 0% 3.000n ± 0% -15.46% (p=0.000 n=50)
memmove/41 3.550n ± 0% 3.001n ± 0% -15.47% (n=50)
memmove/42 3.549n ± 0% 3.001n ± 0% -15.43% (n=50)
memmove/43 3.552n ± 0% 3.001n ± 0% -15.52% (p=0.000 n=50)
memmove/44 3.552n ± 0% 3.001n ± 0% -15.51% (n=50)
memmove/45 3.552n ± 0% 3.002n ± 0% -15.48% (n=50)
memmove/46 3.554n ± 0% 3.001n ± 0% -15.55% (p=0.000 n=50)
memmove/47 3.556n ± 0% 3.002n ± 0% -15.58% (p=0.000 n=50)
memmove/48 3.555n ± 0% 3.003n ± 0% -15.54% (n=50)
memmove/49 3.557n ± 0% 3.002n ± 0% -15.59% (p=0.000 n=50)
memmove/50 3.557n ± 0% 3.004n ± 0% -15.55% (p=0.000 n=50)
memmove/51 3.556n ± 0% 3.004n ± 0% -15.53% (p=0.000 n=50)
memmove/52 3.561n ± 0% 3.004n ± 0% -15.65% (p=0.000 n=50)
memmove/53 3.558n ± 0% 3.004n ± 0% -15.57% (p=0.000 n=50)
memmove/54 3.561n ± 0% 3.005n ± 0% -15.62% (n=50)
memmove/55 3.560n ± 0% 3.006n ± 0% -15.57% (n=50)
memmove/56 3.562n ± 0% 3.006n ± 0% -15.60% (p=0.000 n=50)
memmove/57 3.563n ± 0% 3.006n ± 0% -15.64% (n=50)
memmove/58 3.565n ± 0% 3.007n ± 0% -15.64% (p=0.000 n=50)
memmove/59 3.564n ± 0% 3.006n ± 0% -15.66% (p=0.000 n=50)
memmove/60 3.570n ± 0% 3.008n ± 0% -15.74% (p=0.000 n=50)
memmove/61 3.566n ± 0% 3.009n ± 0% -15.63% (p=0.000 n=50)
memmove/62 3.567n ± 0% 3.007n ± 0% -15.70% (p=0.000 n=50)
memmove/63 3.568n ± 0% 3.008n ± 0% -15.71% (p=0.000 n=50)
memmove/64 4.104n ± 0% 3.008n ± 0% -26.70% (p=0.000 n=50)
memmove/65 4.126n ± 0% 3.662n ± 0% -11.26% (p=0.000 n=50)
memmove/66 4.128n ± 0% 3.662n ± 0% -11.29% (n=50)
memmove/67 4.129n ± 0% 3.662n ± 0% -11.31% (n=50)
memmove/68 4.129n ± 0% 3.661n ± 0% -11.33% (p=0.000 n=50)
memmove/69 4.130n ± 0% 3.662n ± 0% -11.34% (p=0.000 n=50)
memmove/70 4.130n ± 0% 3.662n ± 0% -11.33% (n=50)
memmove/71 4.132n ± 0% 3.662n ± 0% -11.38% (p=0.000 n=50)
memmove/72 4.131n ± 0% 3.661n ± 0% -11.39% (n=50)
memmove/73 4.135n ± 0% 3.661n ± 0% -11.45% (p=0.000 n=50)
memmove/74 4.137n ± 0% 3.662n ± 0% -11.49% (n=50)
memmove/75 4.138n ± 0% 3.662n ± 0% -11.51% (p=0.000 n=50)
memmove/76 4.139n ± 0% 3.661n ± 0% -11.56% (p=0.000 n=50)
memmove/77 4.136n ± 0% 3.662n ± 0% -11.47% (p=0.000 n=50)
memmove/78 4.143n ± 0% 3.661n ± 0% -11.62% (p=0.000 n=50)
memmove/79 4.142n ± 0% 3.661n ± 0% -11.60% (n=50)
memmove/80 4.142n ± 0% 3.661n ± 0% -11.62% (p=0.000 n=50)
memmove/81 4.140n ± 0% 3.661n ± 0% -11.57% (n=50)
memmove/82 4.146n ± 0% 3.661n ± 0% -11.69% (n=50)
memmove/83 4.143n ± 0% 3.661n ± 0% -11.63% (p=0.000 n=50)
memmove/84 4.143n ± 0% 3.661n ± 0% -11.63% (n=50)
memmove/85 4.147n ± 0% 3.661n ± 0% -11.73% (p=0.000 n=50)
memmove/86 4.142n ± 0% 3.661n ± 0% -11.62% (p=0.000 n=50)
memmove/87 4.147n ± 0% 3.661n ± 0% -11.72% (p=0.000 n=50)
memmove/88 4.148n ± 0% 3.661n ± 0% -11.74% (n=50)
memmove/89 4.152n ± 0% 3.661n ± 0% -11.84% (n=50)
memmove/90 4.151n ± 0% 3.661n ± 0% -11.81% (n=50)
memmove/91 4.150n ± 0% 3.661n ± 0% -11.78% (n=50)
memmove/92 4.153n ± 0% 3.661n ± 0% -11.86% (n=50)
memmove/93 4.158n ± 0% 3.661n ± 0% -11.95% (n=50)
memmove/94 4.157n ± 0% 3.661n ± 0% -11.95% (p=0.000 n=50)
memmove/95 4.155n ± 0% 3.661n ± 0% -11.90% (p=0.000 n=50)
memmove/96 4.149n ± 0% 3.660n ± 0% -11.79% (n=50)
memmove/97 4.157n ± 0% 3.661n ± 0% -11.94% (n=50)
memmove/98 4.157n ± 0% 3.661n ± 0% -11.94% (n=50)
memmove/99 4.168n ± 0% 3.661n ± 0% -12.17% (p=0.000 n=50)
memmove/100 4.159n ± 0% 3.660n ± 0% -12.00% (p=0.000 n=50)
memmove/101 4.161n ± 0% 3.660n ± 0% -12.03% (p=0.000 n=50)
memmove/102 4.165n ± 0% 3.660n ± 0% -12.12% (p=0.000 n=50)
memmove/103 4.164n ± 0% 3.661n ± 0% -12.08% (n=50)
memmove/104 4.164n ± 0% 3.660n ± 0% -12.11% (n=50)
memmove/105 4.165n ± 0% 3.660n ± 0% -12.12% (p=0.000 n=50)
memmove/106 4.166n ± 0% 3.660n ± 0% -12.15% (n=50)
memmove/107 4.171n ± 0% 3.660n ± 1% -12.26% (p=0.000 n=50)
memmove/108 4.173n ± 0% 3.660n ± 0% -12.30% (p=0.000 n=50)
memmove/109 4.170n ± 0% 3.660n ± 0% -12.24% (n=50)
memmove/110 4.174n ± 0% 3.660n ± 0% -12.31% (n=50)
memmove/111 4.176n ± 0% 3.660n ± 0% -12.35% (p=0.000 n=50)
memmove/112 4.174n ± 0% 3.659n ± 0% -12.34% (p=0.000 n=50)
memmove/113 4.176n ± 0% 3.660n ± 0% -12.35% (n=50)
memmove/114 4.182n ± 0% 3.660n ± 0% -12.49% (n=50)
memmove/115 4.185n ± 0% 3.660n ± 0% -12.55% (n=50)
memmove/116 4.184n ± 0% 3.659n ± 0% -12.54% (n=50)
memmove/117 4.182n ± 0% 3.660n ± 0% -12.50% (n=50)
memmove/118 4.188n ± 0% 3.660n ± 0% -12.61% (n=50)
memmove/119 4.186n ± 0% 3.660n ± 0% -12.57% (p=0.000 n=50)
memmove/120 4.189n ± 0% 3.659n ± 0% -12.63% (n=50)
memmove/121 4.187n ± 0% 3.660n ± 0% -12.60% (n=50)
memmove/122 4.186n ± 0% 3.660n ± 0% -12.58% (n=50)
memmove/123 4.187n ± 0% 3.660n ± 0% -12.60% (n=50)
memmove/124 4.189n ± 0% 3.659n ± 0% -12.65% (n=50)
memmove/125 4.195n ± 0% 3.659n ± 0% -12.78% (n=50)
memmove/126 4.197n ± 0% 3.659n ± 0% -12.81% (n=50)
memmove/127 4.194n ± 0% 3.659n ± 0% -12.75% (n=50)
memmove/128 5.035n ± 0% 3.659n ± 0% -27.32% (n=50)
memmove/129 5.127n ± 0% 5.164n ± 0% +0.73% (p=0.000 n=50)
memmove/130 5.130n ± 0% 5.176n ± 0% +0.88% (p=0.000 n=50)
memmove/131 5.127n ± 0% 5.180n ± 0% +1.05% (p=0.000 n=50)
memmove/132 5.131n ± 0% 5.169n ± 0% +0.75% (p=0.000 n=50)
memmove/133 5.137n ± 0% 5.179n ± 0% +0.81% (p=0.000 n=50)
memmove/134 5.140n ± 0% 5.178n ± 0% +0.74% (p=0.000 n=50)
memmove/135 5.141n ± 0% 5.187n ± 0% +0.88% (p=0.000 n=50)
memmove/136 5.133n ± 0% 5.184n ± 0% +0.99% (p=0.000 n=50)
memmove/137 5.148n ± 0% 5.186n ± 0% +0.73% (p=0.000 n=50)
memmove/138 5.143n ± 0% 5.189n ± 0% +0.88% (p=0.000 n=50)
memmove/139 5.142n ± 0% 5.192n ± 0% +0.97% (p=0.000 n=50)
memmove/140 5.141n ± 0% 5.192n ± 0% +1.01% (p=0.000 n=50)
memmove/141 5.155n ± 0% 5.188n ± 0% +0.64% (p=0.000 n=50)
memmove/142 5.146n ± 0% 5.192n ± 0% +0.90% (p=0.000 n=50)
memmove/143 5.142n ± 0% 5.203n ± 0% +1.19% (p=0.000 n=50)
memmove/144 5.146n ± 0% 5.197n ± 0% +0.99% (p=0.000 n=50)
memmove/145 5.146n ± 0% 5.196n ± 0% +0.97% (p=0.000 n=50)
memmove/146 5.151n ± 0% 5.207n ± 0% +1.10% (p=0.000 n=50)
memmove/147 5.151n ± 0% 5.205n ± 0% +1.06% (p=0.000 n=50)
memmove/148 5.156n ± 0% 5.190n ± 0% +0.66% (p=0.000 n=50)
memmove/149 5.158n ± 0% 5.212n ± 0% +1.04% (p=0.000 n=50)
memmove/150 5.160n ± 0% 5.203n ± 0% +0.84% (p=0.000 n=50)
memmove/151 5.167n ± 0% 5.210n ± 0% +0.83% (p=0.000 n=50)
memmove/152 5.157n ± 0% 5.206n ± 0% +0.94% (p=0.000 n=50)
memmove/153 5.170n ± 0% 5.211n ± 0% +0.80% (p=0.000 n=50)
memmove/154 5.169n ± 0% 5.222n ± 0% +1.02% (p=0.000 n=50)
memmove/155 5.171n ± 0% 5.215n ± 0% +0.87% (p=0.000 n=50)
memmove/156 5.174n ± 0% 5.214n ± 0% +0.78% (p=0.000 n=50)
memmove/157 5.171n ± 0% 5.218n ± 0% +0.92% (p=0.000 n=50)
memmove/158 5.168n ± 0% 5.224n ± 0% +1.09% (p=0.000 n=50)
memmove/159 5.179n ± 0% 5.218n ± 0% +0.76% (p=0.000 n=50)
memmove/160 5.170n ± 0% 5.219n ± 0% +0.95% (p=0.000 n=50)
memmove/161 5.187n ± 0% 5.220n ± 0% +0.64% (p=0.000 n=50)
memmove/162 5.189n ± 0% 5.234n ± 0% +0.86% (p=0.000 n=50)
memmove/163 5.199n ± 0% 5.250n ± 0% +0.99% (p=0.000 n=50)
memmove/164 5.205n ± 0% 5.260n ± 0% +1.04% (p=0.000 n=50)
memmove/165 5.208n ± 0% 5.261n ± 0% +1.01% (p=0.000 n=50)
memmove/166 5.227n ± 0% 5.275n ± 0% +0.91% (p=0.000 n=50)
memmove/167 5.233n ± 0% 5.281n ± 0% +0.92% (p=0.000 n=50)
memmove/168 5.236n ± 0% 5.295n ± 0% +1.12% (p=0.000 n=50)
memmove/169 5.256n ± 0% 5.297n ± 0% +0.79% (p=0.000 n=50)
memmove/170 5.259n ± 0% 5.302n ± 0% +0.80% (p=0.000 n=50)
memmove/171 5.269n ± 0% 5.321n ± 0% +0.97% (p=0.000 n=50)
memmove/172 5.266n ± 0% 5.318n ± 0% +0.98% (p=0.000 n=50)
memmove/173 5.272n ± 0% 5.330n ± 0% +1.09% (p=0.000 n=50)
memmove/174 5.284n ± 0% 5.331n ± 0% +0.89% (p=0.000 n=50)
memmove/175 5.284n ± 0% 5.322n ± 0% +0.72% (p=0.000 n=50)
memmove/176 5.298n ± 0% 5.337n ± 0% +0.74% (p=0.000 n=50)
memmove/177 5.282n ± 0% 5.338n ± 0% +1.04% (p=0.000 n=50)
memmove/178 5.299n ± 0% 5.337n ± 0% +0.71% (p=0.000 n=50)
memmove/179 5.296n ± 0% 5.343n ± 0% +0.88% (p=0.000 n=50)
memmove/180 5.292n ± 0% 5.343n ± 0% +0.97% (p=0.000 n=50)
memmove/181 5.303n ± 0% 5.335n ± 0% +0.60% (p=0.000 n=50)
memmove/182 5.305n ± 0% 5.338n ± 0% +0.62% (p=0.000 n=50)
memmove/183 5.298n ± 0% 5.329n ± 0% +0.59% (p=0.000 n=50)
memmove/184 5.299n ± 0% 5.333n ± 0% +0.64% (p=0.000 n=50)
memmove/185 5.291n ± 0% 5.330n ± 0% +0.73% (p=0.000 n=50)
memmove/186 5.296n ± 0% 5.332n ± 0% +0.68% (p=0.000 n=50)
memmove/187 5.297n ± 0% 5.320n ± 0% +0.44% (p=0.000 n=50)
memmove/188 5.286n ± 0% 5.314n ± 0% +0.53% (p=0.000 n=50)
memmove/189 5.293n ± 0% 5.318n ± 0% +0.46% (p=0.000 n=50)
memmove/190 5.294n ± 0% 5.318n ± 0% +0.45% (p=0.000 n=50)
memmove/191 5.292n ± 0% 5.314n ± 0% +0.40% (p=0.032 n=50)
memmove/192 5.272n ± 0% 5.304n ± 0% +0.60% (p=0.000 n=50)
memmove/193 5.279n ± 0% 5.310n ± 0% +0.57% (p=0.000 n=50)
memmove/194 5.294n ± 0% 5.308n ± 0% +0.26% (p=0.018 n=50)
memmove/195 5.302n ± 0% 5.311n ± 0% +0.18% (p=0.010 n=50)
memmove/196 5.301n ± 0% 5.316n ± 0% +0.28% (p=0.023 n=50)
memmove/197 5.302n ± 0% 5.327n ± 0% +0.47% (p=0.000 n=50)
memmove/198 5.310n ± 0% 5.326n ± 0% +0.30% (p=0.003 n=50)
memmove/199 5.303n ± 0% 5.319n ± 0% +0.30% (p=0.009 n=50)
memmove/200 5.312n ± 0% 5.330n ± 0% +0.35% (p=0.001 n=50)
memmove/201 5.307n ± 0% 5.333n ± 0% +0.50% (p=0.000 n=50)
memmove/202 5.311n ± 0% 5.334n ± 0% +0.44% (p=0.000 n=50)
memmove/203 5.313n ± 0% 5.335n ± 0% +0.41% (p=0.006 n=50)
memmove/204 5.312n ± 0% 5.332n ± 0% +0.36% (p=0.002 n=50)
memmove/205 5.318n ± 0% 5.345n ± 0% +0.50% (p=0.000 n=50)
memmove/206 5.311n ± 0% 5.333n ± 0% +0.42% (p=0.002 n=50)
memmove/207 5.310n ± 0% 5.338n ± 0% +0.52% (p=0.000 n=50)
memmove/208 5.319n ± 0% 5.341n ± 0% +0.40% (p=0.004 n=50)
memmove/209 5.330n ± 0% 5.346n ± 0% +0.30% (p=0.004 n=50)
memmove/210 5.329n ± 0% 5.349n ± 0% +0.38% (p=0.002 n=50)
memmove/211 5.318n ± 0% 5.340n ± 0% +0.41% (p=0.000 n=50)
memmove/212 5.339n ± 0% 5.343n ± 0% ~ (p=0.396 n=50)
memmove/213 5.329n ± 0% 5.343n ± 0% +0.25% (p=0.017 n=50)
memmove/214 5.339n ± 0% 5.358n ± 0% +0.35% (p=0.035 n=50)
memmove/215 5.342n ± 0% 5.346n ± 0% ~ (p=0.063 n=50)
memmove/216 5.338n ± 0% 5.359n ± 0% +0.39% (p=0.002 n=50)
memmove/217 5.341n ± 0% 5.362n ± 0% +0.39% (p=0.015 n=50)
memmove/218 5.354n ± 0% 5.373n ± 0% +0.36% (p=0.041 n=50)
memmove/219 5.352n ± 0% 5.362n ± 0% ~ (p=0.143 n=50)
memmove/220 5.344n ± 0% 5.370n ± 0% +0.50% (p=0.001 n=50)
memmove/221 5.345n ± 0% 5.373n ± 0% +0.53% (p=0.000 n=50)
memmove/222 5.348n ± 0% 5.360n ± 0% +0.23% (p=0.014 n=50)
memmove/223 5.354n ± 0% 5.377n ± 0% +0.43% (p=0.024 n=50)
memmove/224 5.352n ± 0% 5.363n ± 0% ~ (p=0.052 n=50)
memmove/225 5.372n ± 0% 5.380n ± 0% ~ (p=0.481 n=50)
memmove/226 5.368n ± 0% 5.386n ± 0% +0.34% (p=0.004 n=50)
memmove/227 5.386n ± 0% 5.402n ± 0% +0.29% (p=0.028 n=50)
memmove/228 5.400n ± 0% 5.408n ± 0% ~ (p=0.174 n=50)
memmove/229 5.423n ± 0% 5.427n ± 0% ~ (p=0.444 n=50)
memmove/230 5.411n ± 0% 5.429n ± 0% +0.33% (p=0.020 n=50)
memmove/231 5.420n ± 0% 5.433n ± 0% +0.24% (p=0.034 n=50)
memmove/232 5.435n ± 0% 5.441n ± 0% ~ (p=0.235 n=50)
memmove/233 5.446n ± 0% 5.462n ± 0% ~ (p=0.590 n=50)
memmove/234 5.467n ± 0% 5.461n ± 0% ~ (p=0.921 n=50)
memmove/235 5.472n ± 0% 5.478n ± 0% ~ (p=0.883 n=50)
memmove/236 5.466n ± 0% 5.478n ± 0% ~ (p=0.324 n=50)
memmove/237 5.471n ± 0% 5.489n ± 0% ~ (p=0.132 n=50)
memmove/238 5.485n ± 0% 5.489n ± 0% ~ (p=0.460 n=50)
memmove/239 5.484n ± 0% 5.488n ± 0% ~ (p=0.833 n=50)
memmove/240 5.483n ± 0% 5.495n ± 0% ~ (p=0.095 n=50)
memmove/241 5.498n ± 0% 5.514n ± 0% ~ (p=0.077 n=50)
memmove/242 5.518n ± 0% 5.517n ± 0% ~ (p=0.481 n=50)
memmove/243 5.514n ± 0% 5.511n ± 0% ~ (p=0.503 n=50)
memmove/244 5.510n ± 0% 5.497n ± 0% -0.24% (p=0.038 n=50)
memmove/245 5.516n ± 0% 5.505n ± 0% ~ (p=0.317 n=50)
memmove/246 5.513n ± 1% 5.494n ± 0% ~ (p=0.147 n=50)
memmove/247 5.518n ± 0% 5.499n ± 0% -0.36% (p=0.011 n=50)
memmove/248 5.503n ± 0% 5.492n ± 0% ~ (p=0.267 n=50)
memmove/249 5.498n ± 0% 5.497n ± 0% ~ (p=0.765 n=50)
memmove/250 5.485n ± 0% 5.493n ± 0% ~ (p=0.348 n=50)
memmove/251 5.503n ± 0% 5.482n ± 0% -0.37% (p=0.013 n=50)
memmove/252 5.497n ± 0% 5.485n ± 0% ~ (p=0.077 n=50)
memmove/253 5.489n ± 0% 5.496n ± 0% ~ (p=0.850 n=50)
memmove/254 5.497n ± 0% 5.491n ± 0% ~ (p=0.548 n=50)
memmove/255 5.484n ± 1% 5.494n ± 0% ~ (p=0.888 n=50)
memmove/256 6.952n ± 0% 7.676n ± 0% +10.41% (p=0.000 n=50)
geomean 4.406n 4.127n -6.33%
```
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
inline_memmove may handle some small sizes as efficiently
as inline_memcpy. For these sizes we may not do is_disjoint check.
This both avoids additional code for the most frequent smaller sizes
and removes code bloat (we don't need the memcpy logic for small sizes).
Here we heavily rely on inlining and dead code elimination: from the first
inline_memmove we should get only handling of small sizes, and from
the second inline_memmove and inline_memcpy we should get only handling
of larger sizes.
Use the memcpy thresholds for memmove.
Memcpy thresholds were more carefully tuned.
This becomes more important since we use memmove
for all small sizes always now.
Fix boundary conditions for sizes = 16/32/64.
See the added comment for explanations.
Memmove function size drops from 885 to 715 bytes
due to removed duplication.