go/sqlite3
1
0
Fork 0
mirror of https://github.com/ncruces/go-sqlite3.git synced 2026-01-12 05:59:14 +00:00
Code Issues Packages Projects Releases Wiki Activity

Optimize strlen and strchr on ARM (#262)

Browse Source
This commit is contained in:
Nuno Cruces
2025-04-15 00:44:31 +01:00
committed by GitHub
parent 9706fa9607
commit ad16d329ea
14 changed files with 326 additions and 268 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
sqlite3/libc/build.sh
View File

@@ -18,6 +18,8 @@ trap 'rm -f libc.tmp' EXIT
-mbulk-memory -mreference-types \
-mnontrapping-fptoint -msign-ext \
-fno-stack-protector -fno-stack-clash-protection \
-Wl,--stack-first \
-Wl,--import-undefined \
-Wl,--initial-memory=16777216 \
-Wl,--export=memchr \
-Wl,--export=memcmp \

BIN
sqlite3/libc/libc.wasm
View File

Binary file not shown.

557
sqlite3/libc/libc.wat
View File

@@ -3,7 +3,7 @@
(type $1 (func (param i32 i32) (result i32)))
(type $2 (func (param i32) (result i32)))
(memory $0 256)
(data $0 (i32.const 1024) "\01")
(data $0 (i32.const 65536) "\01")
(export "memory" (memory $0))
(export "memset" (func $memset))
(export "memcpy" (func $memcpy))
@@ -229,62 +229,70 @@
(local $1 i32)
(local $2 i32)
(local $3 v128)
(local $scratch v128)
(if
(i32.eqz
(local.tee $2
(i32.and
(i8x16.bitmask
(i8x16.eq
(v128.load
(local.tee $1
(i32.and
(local.get $0)
(i32.const -16)
)
(block $block1
(block $block
(br_if $block
(i8x16.all_true
(local.tee $3
(v128.load
(local.tee $1
(i32.and
(local.get $0)
(i32.const -16)
)
)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
)
(i32.shl
(i32.const -1)
(i32.and
(local.get $0)
(i32.const 15)
)
)
)
)
)
(then
(loop $label
(br_if $label
(i32.eqz
(local.tee $2
(br_if $block
(i32.eqz
(local.tee $2
(i32.and
(i8x16.bitmask
(i8x16.eq
(block (result v128)
(local.set $scratch
(v128.load offset=16
(local.get $1)
)
)
(local.set $1
(i32.add
(local.get $1)
(i32.const 16)
)
)
(local.get $scratch)
)
(local.get $3)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
)
(i32.shl
(i32.const -1)
(i32.and
(local.get $0)
(i32.const 15)
)
)
)
)
)
)
(br $block1)
)
(loop $label
(local.set $3
(v128.load offset=16
(local.get $1)
)
)
(local.set $1
(i32.add
(local.get $1)
(i32.const 16)
)
)
(br_if $label
(i8x16.all_true
(local.get $3)
)
)
)
(local.set $2
(i8x16.bitmask
(i8x16.eq
(local.get $3)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
)
)
)
(i32.add
@@ -624,74 +632,81 @@
(i32.const 0)
)
(func $strchrnul (param $0 i32) (param $1 i32) (result i32)
(local $2 i32)
(local $2 v128)
(local $3 v128)
(local $4 v128)
(if
(i32.eqz
(local.tee $0
(i32.and
(i8x16.bitmask
(v128.or
(i8x16.eq
(local.tee $3
(v128.load
(local.tee $2
(i32.and
(local.get $0)
(i32.const -16)
)
(local $4 i32)
(block $block
(if
(v128.any_true
(local.tee $2
(v128.or
(i8x16.eq
(local.tee $2
(v128.load
(local.tee $4
(i32.and
(local.get $0)
(i32.const -16)
)
)
)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $3)
(local.tee $4
(i8x16.splat
(local.get $1)
)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $2)
(local.tee $3
(i8x16.splat
(local.get $1)
)
)
)
)
(i32.shl
(i32.const -1)
)
)
(then
(br_if $block
(local.tee $0
(i32.and
(local.get $0)
(i32.const 15)
(i8x16.bitmask
(local.get $2)
)
(i32.shl
(i32.const -1)
(i32.and
(local.get $0)
(i32.const 15)
)
)
)
)
)
)
)
(then
(loop $label
(local.set $3
(v128.load offset=16
(local.get $2)
)
(loop $label
(local.set $2
(v128.load offset=16
(local.get $4)
)
(local.set $2
(i32.add
(local.get $2)
(i32.const 16)
)
)
(local.set $4
(i32.add
(local.get $4)
(i32.const 16)
)
(br_if $label
(i32.eqz
(local.tee $0
(i8x16.bitmask
(v128.or
(i8x16.eq
(local.get $3)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $3)
(local.get $4)
)
)
(br_if $label
(i32.eqz
(v128.any_true
(local.tee $2
(v128.or
(i8x16.eq
(local.get $2)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $2)
(local.get $3)
)
)
)
@@ -699,83 +714,95 @@
)
)
)
(local.set $0
(i8x16.bitmask
(local.get $2)
)
)
)
(i32.add
(local.get $2)
(local.get $4)
(i32.ctz
(local.get $0)
)
)
)
(func $strchr (param $0 i32) (param $1 i32) (result i32)
(local $2 i32)
(local $2 v128)
(local $3 v128)
(local $4 v128)
(if
(i32.eqz
(local.tee $0
(i32.and
(i8x16.bitmask
(v128.or
(i8x16.eq
(local.tee $3
(v128.load
(local.tee $2
(i32.and
(local.get $0)
(i32.const -16)
)
(local $4 i32)
(block $block
(if
(v128.any_true
(local.tee $2
(v128.or
(i8x16.eq
(local.tee $2
(v128.load
(local.tee $4
(i32.and
(local.get $0)
(i32.const -16)
)
)
)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $3)
(local.tee $4
(i8x16.splat
(local.get $1)
)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $2)
(local.tee $3
(i8x16.splat
(local.get $1)
)
)
)
)
(i32.shl
(i32.const -1)
)
)
(then
(br_if $block
(local.tee $0
(i32.and
(local.get $0)
(i32.const 15)
(i8x16.bitmask
(local.get $2)
)
(i32.shl
(i32.const -1)
(i32.and
(local.get $0)
(i32.const 15)
)
)
)
)
)
)
)
(then
(loop $label
(local.set $3
(v128.load offset=16
(local.get $2)
)
(loop $label
(local.set $2
(v128.load offset=16
(local.get $4)
)
(local.set $2
(i32.add
(local.get $2)
(i32.const 16)
)
)
(local.set $4
(i32.add
(local.get $4)
(i32.const 16)
)
(br_if $label
(i32.eqz
(local.tee $0
(i8x16.bitmask
(v128.or
(i8x16.eq
(local.get $3)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $3)
(local.get $4)
)
)
(br_if $label
(i32.eqz
(v128.any_true
(local.tee $2
(v128.or
(i8x16.eq
(local.get $2)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $2)
(local.get $3)
)
)
)
@@ -783,11 +810,16 @@
)
)
)
(local.set $0
(i8x16.bitmask
(local.get $2)
)
)
)
(select
(local.tee $0
(i32.add
(local.get $2)
(local.get $4)
(i32.ctz
(local.get $0)
)
@@ -808,67 +840,72 @@
(func $strspn (param $0 i32) (param $1 i32) (result i32)
(local $2 i32)
(local $3 i32)
(v128.store
(i32.const 66560)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(v128.store
(i32.const 66544)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(if
(local.tee $2
(i32.eqz
(local.tee $2
(i32.load8_u
(local.get $1)
)
)
)
(then
(return
(i32.const 0)
)
)
)
(v128.store
(i32.const 65520)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(v128.store
(i32.const 65504)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(local.set $1
(i32.add
(local.get $1)
(i32.const 1)
)
)
(loop $label
(i32.store
(local.tee $3
(i32.add
(i32.and
(i32.shr_u
(local.get $2)
(i32.const 3)
)
(i32.const 28)
)
(i32.const 65504)
)
)
(i32.or
(i32.load
(local.get $3)
)
(i32.shl
(i32.const 1)
(local.get $2)
)
)
)
(local.set $2
(i32.load8_u
(local.get $1)
)
)
(then
(local.set $1
(i32.add
(local.get $1)
(i32.const 1)
)
)
(loop $label
(i32.store
(local.tee $3
(i32.add
(i32.and
(i32.shr_u
(local.get $2)
(i32.const 3)
)
(i32.const 28)
)
(i32.const 66544)
)
)
(i32.or
(i32.load
(local.get $3)
)
(i32.shl
(i32.const 1)
(local.get $2)
)
)
)
(local.set $2
(i32.load8_u
(local.get $1)
)
)
(local.set $1
(i32.add
(local.get $1)
(i32.const 1)
)
)
(br_if $label
(local.get $2)
)
(local.set $1
(i32.add
(local.get $1)
(i32.const 1)
)
)
(br_if $label
(local.get $2)
)
)
(if
(local.tee $2
@@ -893,7 +930,7 @@
)
(i32.const 28)
)
(i32.const 66544)
(i32.const 65504)
)
)
(local.get $2)
@@ -952,71 +989,78 @@
)
)
)
(if
(i32.eqz
(local.tee $2
(i32.and
(i8x16.bitmask
(v128.or
(i8x16.eq
(local.tee $4
(v128.load
(local.tee $1
(i32.and
(local.get $0)
(i32.const -16)
)
(block $block1
(if
(v128.any_true
(local.tee $4
(v128.or
(i8x16.eq
(local.tee $4
(v128.load
(local.tee $1
(i32.and
(local.get $0)
(i32.const -16)
)
)
)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $4)
(local.tee $5
(i8x16.splat
(local.get $2)
)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $4)
(local.tee $5
(i8x16.splat
(local.get $2)
)
)
)
)
(i32.shl
(i32.const -1)
)
)
(then
(br_if $block1
(local.tee $2
(i32.and
(local.get $0)
(i32.const 15)
(i8x16.bitmask
(local.get $4)
)
(i32.shl
(i32.const -1)
(i32.and
(local.get $0)
(i32.const 15)
)
)
)
)
)
)
)
(then
(loop $label
(local.set $4
(v128.load offset=16
(local.get $1)
)
(loop $label
(local.set $4
(v128.load offset=16
(local.get $1)
)
(local.set $1
(i32.add
(local.get $1)
(i32.const 16)
)
)
(local.set $1
(i32.add
(local.get $1)
(i32.const 16)
)
(br_if $label
(i32.eqz
(local.tee $2
(i8x16.bitmask
(v128.or
(i8x16.eq
(local.get $4)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $4)
(local.get $5)
)
)
(br_if $label
(i32.eqz
(v128.any_true
(local.tee $4
(v128.or
(i8x16.eq
(local.get $4)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(i8x16.eq
(local.get $4)
(local.get $5)
)
)
)
@@ -1024,6 +1068,11 @@
)
)
)
(local.set $2
(i8x16.bitmask
(local.get $4)
)
)
)
(return
(i32.sub
@@ -1038,11 +1087,11 @@
)
)
(v128.store
(i32.const 66560)
(i32.const 65520)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(v128.store
(i32.const 66544)
(i32.const 65504)
(v128.const i32x4 0x00000000 0x00000000 0x00000000 0x00000000)
)
(local.set $1
@@ -1062,7 +1111,7 @@
)
(i32.const 28)
)
(i32.const 66544)
(i32.const 65504)
)
)
(i32.or
@@ -1112,7 +1161,7 @@
)
(i32.const 28)
)
(i32.const 66544)
(i32.const 65504)
)
)
(local.get $2)

25
sqlite3/strings.c
View File

@@ -66,10 +66,13 @@ size_t strlen(const char *s) {
const v128_t *w = (void *)(s - align);
while (true) {
int mask =
wasm_i8x16_bitmask(wasm_i8x16_eq(*w, (v128_t){})) >> align << align;
if (mask) {
return (char *)w - s + __builtin_ctz(mask);
if (!wasm_i8x16_all_true(*w)) {
const v128_t cmp = wasm_i8x16_eq(*w, (v128_t){});
int mask = wasm_i8x16_bitmask(cmp) >> align << align;
__builtin_assume(mask || align);
if (mask) {
return (char *)w - s + __builtin_ctz(mask);
}
}
align = 0;
w++;
@@ -136,11 +139,13 @@ char *strchrnul(const char *s, int c) {
const v128_t wc = wasm_i8x16_splat(c);
while (true) {
int mask = wasm_i8x16_bitmask(wasm_i8x16_eq(*w, (v128_t){}) |
wasm_i8x16_eq(*w, wc)) >>
align << align;
if (mask) {
return (char *)w + __builtin_ctz(mask);
const v128_t cmp = wasm_i8x16_eq(*w, (v128_t){}) | wasm_i8x16_eq(*w, wc);
if (wasm_v128_any_true(cmp)) {
int mask = wasm_i8x16_bitmask(cmp) >> align << align;
__builtin_assume(mask || align);
if (mask) {
return (char *)w + __builtin_ctz(mask);
}
}
align = 0;
w++;
@@ -159,6 +164,8 @@ char *strchr(const char *s, int c) {
<< ((b) % (8 * sizeof(size_t))))
size_t strspn(const char *s, const char *c) {
if (!c[0]) return 0;
const char *const a = s;
size_t byteset[32 / sizeof(size_t)] = {0};