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

Fix bcmp.

Browse Source
This commit is contained in:
Nuno Cruces
2025-05-13 14:27:26 +01:00
parent c1bed07e3a
commit fc5ced209c
11 changed files with 212 additions and 218 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
sqlite3/libc/build.sh
View File

@@ -19,9 +19,9 @@ EOF
-Wall -Wextra -Wno-unused-parameter -Wno-unused-function \
-o libc.wasm -I. "$SRCS" \
-mexec-model=reactor \
-msimd128 -mmutable-globals -mmultivalue \
-mbulk-memory -mreference-types \
-mnontrapping-fptoint -msign-ext \
-mmutable-globals -mnontrapping-fptoint \
-msimd128 -mbulk-memory -msign-ext \
-mreference-types -mmultivalue \
-fno-stack-protector -fno-stack-clash-protection \
-Wl,-z,stack-size=4096 \
-Wl,--stack-first \
@@ -54,8 +54,8 @@ EOF
"$BINARYEN/wasm-ctor-eval" -g -c _initialize libc.wasm -o libc.tmp
"$BINARYEN/wasm-opt" -g --strip --strip-producers -c -O3 \
libc.tmp -o libc.wasm \
--enable-simd --enable-mutable-globals --enable-multivalue \
--enable-bulk-memory --enable-reference-types \
--enable-nontrapping-float-to-int --enable-sign-ext
--enable-mutable-globals --enable-nontrapping-float-to-int \
--enable-simd --enable-bulk-memory --enable-sign-ext \
--enable-reference-types --enable-multivalue
"$BINARYEN/wasm-dis" -o libc.wat libc.wasm

BIN
sqlite3/libc/libc.wasm
View File

Binary file not shown.

270
sqlite3/libc/libc.wat
View File

@@ -2164,13 +2164,13 @@
(local $18 v128)
(local $19 v128)
(local $20 v128)
(local.set $5
(local.set $10
(i32.load8_u
(local.get $2)
)
)
(local.set $6
(local.tee $10
(local.tee $8
(i32.sub
(local.get $3)
(i32.const 1)
@@ -2181,7 +2181,7 @@
(loop $label
(br_if $block
(i32.ne
(local.get $5)
(local.get $10)
(local.tee $7
(i32.load8_u
(i32.add
@@ -2205,15 +2205,15 @@
(i32.load8_u
(i32.add
(local.get $2)
(local.get $10)
(local.get $8)
)
)
)
(local.set $6
(local.get $10)
(local.get $8)
)
)
(block $block6
(block $block5
(block $block1
(br_if $block1
(i32.lt_u
@@ -2239,7 +2239,7 @@
)
(local.set $19
(i8x16.splat
(local.get $5)
(local.get $10)
)
)
(local.set $15
@@ -2248,12 +2248,18 @@
(i32.const 14)
)
)
(local.set $16
(local.set $10
(i32.add
(local.get $2)
(i32.const 1)
)
)
(local.set $16
(i32.lt_u
(local.get $3)
(i32.const 17)
)
)
(loop $label4
(block $block2
(br_if $block2
@@ -2310,134 +2316,118 @@
)
)
(loop $label3
(br_if $block6
(i32.eqz
(block $block5 (result i32)
(local.set $8
(i32.add
(local.tee $7
(i32.add
(local.get $0)
(i32.ctz
(local.get $11)
)
)
)
(i32.const 1)
(local.set $9
(i32.add
(local.tee $12
(i32.add
(local.get $0)
(i32.ctz
(local.get $11)
)
)
)
(i32.const 1)
)
)
(block $block4
(if
(i32.eqz
(local.get $16)
)
(then
(local.set $7
(local.get $10)
)
(local.set $5
(local.get $16)
(local.get $8)
)
(local.set $12
(i32.const 0)
)
(block $block4
(if
(i32.ge_u
(local.tee $9
(local.get $10)
)
(i32.const 16)
)
(then
(local.set $12
(i32.const 1)
)
(loop $label1
(br_if $block4
(v128.any_true
(v128.xor
(v128.load align=1
(local.get $5)
)
(v128.load align=1
(local.get $8)
)
)
)
)
(local.set $5
(i32.add
(local.get $5)
(local.tee $13
(i32.add
(i32.and
(i32.sub
(local.get $9)
(i32.const 1)
)
(i32.const 15)
)
(i32.const 1)
)
)
)
)
(local.set $8
(i32.add
(local.get $8)
(local.get $13)
)
)
(br_if $label1
(local.tee $9
(i32.sub
(local.get $9)
(local.get $13)
)
)
)
)
(br $block5
(i32.const 0)
)
)
)
(loop $label1
(br_if $block4
(i32.eqz
(local.get $9)
)
)
(loop $label2
(drop
(br_if $block5
(i32.const 1)
(i32.ne
(i32.load8_u
(local.get $8)
)
(i32.load8_u
(local.get $5)
)
(v128.any_true
(v128.xor
(v128.load align=1
(local.get $7)
)
(v128.load align=1
(local.get $9)
)
)
)
(local.set $5
(i32.add
(local.get $5)
(i32.const 1)
)
)
(local.set $8
(i32.add
(local.get $8)
(i32.const 1)
)
)
(br_if $label2
(local.tee $9
(i32.sub
(local.get $9)
)
(local.set $7
(i32.add
(local.get $7)
(local.tee $13
(i32.add
(i32.and
(i32.sub
(local.get $5)
(i32.const 1)
)
(i32.const 15)
)
(i32.const 1)
)
)
)
)
(local.set $9
(i32.add
(local.get $9)
(local.get $13)
)
)
(br_if $label1
(local.tee $5
(i32.sub
(local.get $5)
(local.get $13)
)
)
)
)
(local.get $12)
(br $block5)
)
)
(local.set $5
(local.get $8)
)
(local.set $7
(local.get $10)
)
(loop $label2
(br_if $block4
(i32.ne
(i32.load8_u
(local.get $9)
)
(i32.load8_u
(local.get $7)
)
)
)
(local.set $7
(i32.add
(local.get $7)
(i32.const 1)
)
)
(local.set $9
(i32.add
(local.get $9)
(i32.const 1)
)
)
(br_if $label2
(local.tee $5
(i32.sub
(local.get $5)
(i32.const 1)
)
)
)
)
(br $block5)
)
(br_if $label3
(local.tee $11
@@ -2476,17 +2466,17 @@
)
)
)
(block $block7
(block $block6
(if
(i32.ne
(local.get $1)
(i32.const -1)
)
(then
(local.set $7
(local.set $12
(i32.const 0)
)
(br_if $block6
(br_if $block5
(i32.lt_u
(local.get $1)
(local.tee $1
@@ -2497,19 +2487,19 @@
)
)
)
(br_if $block6
(br_if $block5
(i32.lt_u
(local.get $1)
(local.get $3)
)
)
(br $block7)
(br $block6)
)
)
(local.set $1
(i32.const -1)
)
(br_if $block7
(br_if $block6
(i8x16.all_true
(local.get $17)
)
@@ -2531,16 +2521,16 @@
)
)
)
(local.set $10
(local.set $8
(i32.sub
(local.get $1)
(local.get $3)
)
)
(local.set $4
(local.set $5
(i32.const 0)
)
(local.set $1
(local.set $4
(i32.ne
(local.get $1)
(i32.const -1)
@@ -2551,14 +2541,14 @@
(i32.const 0)
)
(loop $label6
(local.set $7
(local.set $12
(i32.const 0)
)
(br_if $block6
(br_if $block5
(i32.eqz
(i32.or
(local.get $1)
(local.tee $5
(local.get $4)
(local.tee $1
(i32.load8_u
(i32.add
(local.get $0)
@@ -2571,7 +2561,7 @@
)
(if
(i32.ne
(local.get $5)
(local.get $1)
(i32.load8_u
(i32.add
(local.get $2)
@@ -2588,16 +2578,16 @@
)
(br_if $label5
(i32.le_u
(local.tee $4
(local.tee $5
(i32.add
(local.get $4)
(local.get $5)
(i32.const 1)
)
)
(local.get $10)
(local.get $8)
)
)
(br $block6)
(br $block5)
)
)
(br_if $label6
@@ -2613,11 +2603,11 @@
)
)
)
(local.set $7
(local.set $12
(local.get $0)
)
)
(local.get $7)
(local.get $12)
)
(func $strstr (param $0 i32) (param $1 i32) (result i32)
(local $2 i32)

3
sqlite3/libc/stdlib.h
View File

@@ -1,9 +1,6 @@
#ifndef _WASM_SIMD128_STDLIB_H
#define _WASM_SIMD128_STDLIB_H
#include <stddef.h>
#include <stdint.h>
#include_next <stdlib.h> // the system stdlib.h
#ifdef __cplusplus

47
sqlite3/libc/string.h
View File

@@ -1,9 +1,7 @@
#ifndef _WASM_SIMD128_STRING_H
#define _WASM_SIMD128_STRING_H
#include <stddef.h>
#include <stdint.h>
#include <strings.h>
#include <wasm_simd128.h>
#include <__macro_PAGESIZE.h>
@@ -82,6 +80,49 @@ int memcmp(const void *v1, const void *v2, size_t n) {
return 0;
}
#ifdef __OPTIMIZE_SIZE__
// __memcmpeq is the same as memcmp but only compares for equality.
#define __memcmpeq(v1, v2, n) memcmp(v1, v2, n)
#else // __OPTIMIZE_SIZE__
static int __memcmpeq(const void *v1, const void *v2, size_t n) {
// Baseline algorithm.
if (n < sizeof(v128_t)) {
const unsigned char *u1 = (unsigned char *)v1;
const unsigned char *u2 = (unsigned char *)v2;
while (n--) {
if (*u1 != *u2) return 1;
u1++;
u2++;
}
return 0;
}
// memcmpeq is allowed to read up to n bytes from each object.
// Unaligned loads handle the case where the objects
// have mismatching alignments.
const v128_t *w1 = (v128_t *)v1;
const v128_t *w2 = (v128_t *)v2;
while (n) {
// Find any single bit difference.
if (wasm_v128_any_true(wasm_v128_load(w1) ^ wasm_v128_load(w2))) {
return 1;
}
// This makes n a multiple of sizeof(v128_t)
// for every iteration except the first.
size_t align = (n - 1) % sizeof(v128_t) + 1;
w1 = (v128_t *)((char *)w1 + align);
w2 = (v128_t *)((char *)w2 + align);
n -= align;
}
return 0;
}
#endif // __OPTIMIZE_SIZE__
__attribute__((weak))
void *memchr(const void *v, int c, size_t n) {
// When n is zero, a function that locates a character finds no occurrence.
@@ -510,7 +551,7 @@ static const char *__memmem_raita(const char *haystk, size_t sh,
// Each iteration clears that bit, tries again.
for (uint32_t mask = wasm_i8x16_bitmask(cmp); mask; mask &= mask - 1) {
size_t ctz = __builtin_ctz(mask);
if (!bcmp(haystk + ctz + 1, needle + 1, sn - 1)) {
if (!__memcmpeq(haystk + ctz + 1, needle + 1, sn - 1)) {
return haystk + ctz;
}
}

38
sqlite3/libc/strings.h
View File

@@ -1,8 +1,7 @@
#ifndef _WASM_SIMD128_STRINGS_H
#define _WASM_SIMD128_STRINGS_H
#include <stddef.h>
#include <wasm_simd128.h>
#include <string.h>
#include_next <strings.h> // the system strings.h
@@ -11,45 +10,12 @@ extern "C" {
#endif
#ifdef __wasm_simd128__
#ifndef __OPTIMIZE_SIZE__
__attribute__((weak))
int bcmp(const void *v1, const void *v2, size_t n) {
// bcmp is the same as memcmp but only compares for equality.
// Baseline algorithm.
if (n < sizeof(v128_t)) {
const unsigned char *u1 = (unsigned char *)v1;
const unsigned char *u2 = (unsigned char *)v2;
while (n--) {
if (*u1 != *u2) return 1;
u1++;
u2++;
}
return 0;
}
// bcmp is allowed to read up to n bytes from each object.
// Unaligned loads handle the case where the objects
// have mismatching alignments.
const v128_t *w1 = (v128_t *)v1;
const v128_t *w2 = (v128_t *)v2;
while (n) {
// Find any single bit difference.
if (wasm_v128_any_true(wasm_v128_load(w1) ^ wasm_v128_load(w2))) {
return 1;
}
// This makes n a multiple of sizeof(v128_t)
// for every iteration except the first.
size_t align = (n - 1) % sizeof(v128_t) + 1;
w1 = (v128_t *)((char *)w1 + align);
w2 = (v128_t *)((char *)w2 + align);
n -= align;
}
return 0;
return __memcmpeq(v1, v2, n);
}
#endif // __OPTIMIZE_SIZE__
#endif // __wasm_simd128__
#ifdef __cplusplus