From 79bf171210cd7e1622caf1d3136ebec148984b1a Mon Sep 17 00:00:00 2001
From: Nuno Cruces <ncruces@users.noreply.github.com>
Date: Wed, 16 Apr 2025 16:39:36 +0100
Subject: [PATCH] Fix #263.

---
 sqlite3/libc/libc.wasm | Bin 2081 -> 2075 bytes
 sqlite3/libc/libc.wat  | 451 +++++++++++++++++++++--------------------
 sqlite3/sqlite_opt.h   |   4 +-
 sqlite3/strings.c      |  61 +++---
 sqlite3/vfs.c          |   5 +-
 5 files changed, 270 insertions(+), 251 deletions(-)

diff --git a/sqlite3/libc/libc.wasm b/sqlite3/libc/libc.wasm
index 68687cdf41200f41814c03e2e3bf4d548b2a34e2..f3761564dbea085bf113264ce25596df95eab1da 100755
GIT binary patch
delta 373
zcmY+9ze)o^5XR@vCKl_FwX&UEYg5~~TPOqx67mRI$X%fa=~7(b1E^f5m8CX{c`NZ1
z7M#5(Xfrb~eBbx`p52qXf5Fd(*T?P-i91>gq)F=yMT&afw1G1i5X_UXlMli;FHMwj
zzD%8%RU5EED<q~)t_dh~R-u5HCb`-(MUJx7|7i8g!%I+bkepXwcgf10dV2L@!K>Pj
ztc>~%-ga+D)Zv=qF(kJt3KE!A3RJQ{{V@Z8K$*ehMP6!#$rmRkCYacgQhU@Yt`4x}
zQ|Vk&t-zsbNU$tE)n1UW_Uc_0syFptG@S}J_WhUR4+#$>1wpv<rsQT7_D3xHdNy92
M|Aq5>bIl*-8~rXtk^lez

delta 382
zcmYk2J4ysW5QeMz5w*iKENCXxW2~v^bTFG3i0l!}++bY|HyDrLvWaM9sF-;WGf!YB
z;sqN0YX%=QbV2jg|LD7TAFm(ybNzVRFPK-R)gvayW7i6oI$7Ha6(SN__R6d>LOGr#
zQfA(mTpv{wf|#!2Ae_H=nOIl$vBA>;=MiV%%4SLwgu)dY(~p+WJm0+WW&fdXW@UQ8
z_C>CjNTE{62xDxiM<GXz*HD<QoRk?LSGE=M7=RGKmDA3(nQB;pjL8j~F?pm!0siUX
zfkLB0G&%ey6Ou(o^bxD~RFmYP3hQ98$-rkoHm~Mj^2g4@&Dd@IvNLJC%Ht#aZ=&`4
J?(@Pvxo@_uP=o*g

diff --git a/sqlite3/libc/libc.wat b/sqlite3/libc/libc.wat
index 4354504..65e52cc 100644
--- a/sqlite3/libc/libc.wat
+++ b/sqlite3/libc/libc.wat
@@ -136,94 +136,128 @@
   (i32.const 0)
  )
  (func $memchr (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
-  (local $3 v128)
-  (block $block2
-   (block $block1
-    (block $block
-     (if
-      (i32.ge_u
-       (local.get $2)
-       (i32.const 16)
-      )
-      (then
-       (local.set $3
-        (i8x16.splat
-         (local.get $1)
-        )
-       )
-       (loop $label
-        (br_if $block
-         (v128.any_true
-          (i8x16.eq
-           (v128.load align=1
-            (local.get $0)
-           )
-           (local.get $3)
+  (local $3 i32)
+  (local $4 i32)
+  (local $5 i32)
+  (local $6 v128)
+  (local $7 v128)
+  (local.set $4
+   (i32.and
+    (local.get $0)
+    (i32.const 15)
+   )
+  )
+  (block $block1
+   (block $block
+    (if
+     (v128.any_true
+      (local.tee $6
+       (i8x16.eq
+        (v128.load
+         (local.tee $3
+          (i32.and
+           (local.get $0)
+           (i32.const -16)
           )
          )
         )
-        (local.set $0
-         (i32.add
-          (local.get $0)
+        (local.tee $7
+         (i8x16.splat
+          (local.get $1)
+         )
+        )
+       )
+      )
+     )
+     (then
+      (br_if $block
+       (local.tee $1
+        (i32.and
+         (i8x16.bitmask
+          (local.get $6)
+         )
+         (i32.shl
+          (i32.const -1)
+          (local.get $4)
+         )
+        )
+       )
+      )
+     )
+    )
+    (br_if $block1
+     (i32.gt_u
+      (local.tee $1
+       (i32.sub
+        (i32.add
+         (local.get $2)
+         (local.get $4)
+        )
+        (i32.const 16)
+       )
+      )
+      (local.get $2)
+     )
+    )
+    (local.set $3
+     (i32.add
+      (i32.sub
+       (local.get $0)
+       (local.get $4)
+      )
+      (i32.const 16)
+     )
+    )
+    (block $block2
+     (loop $label
+      (br_if $block2
+       (v128.any_true
+        (local.tee $6
+         (i8x16.eq
+          (v128.load
+           (local.get $3)
+          )
+          (local.get $7)
+         )
+        )
+       )
+      )
+      (local.set $3
+       (i32.add
+        (local.get $3)
+        (i32.const 16)
+       )
+      )
+      (br_if $label
+       (i32.ge_u
+        (local.get $1)
+        (local.tee $1
+         (i32.sub
+          (local.get $1)
           (i32.const 16)
          )
         )
-        (br_if $label
-         (i32.gt_u
-          (local.tee $2
-           (i32.sub
-            (local.get $2)
-            (i32.const 16)
-           )
-          )
-          (i32.const 15)
-         )
-        )
        )
       )
      )
-     (br_if $block1
-      (i32.eqz
-       (local.get $2)
-      )
-     )
+     (br $block1)
     )
     (local.set $1
-     (i32.and
-      (local.get $1)
-      (i32.const 255)
-     )
-    )
-    (loop $label1
-     (br_if $block2
-      (i32.eq
-       (i32.load8_u
-        (local.get $0)
-       )
-       (local.get $1)
-      )
-     )
-     (local.set $0
-      (i32.add
-       (local.get $0)
-       (i32.const 1)
-      )
-     )
-     (br_if $label1
-      (local.tee $2
-       (i32.sub
-        (local.get $2)
-        (i32.const 1)
-       )
-      )
+     (i8x16.bitmask
+      (local.get $6)
      )
     )
    )
-   (local.set $0
-    (i32.const 0)
+   (local.set $5
+    (i32.add
+     (local.get $3)
+     (i32.ctz
+      (local.get $1)
+     )
+    )
    )
   )
-  (local.get $0)
+  (local.get $5)
  )
  (func $strlen (param $0 i32) (result i32)
   (local $1 i32)
@@ -309,123 +343,90 @@
   (local $2 i32)
   (local $3 i32)
   (local $4 v128)
-  (local $5 v128)
-  (local.set $3
-   (block $block (result i32)
-    (if
-     (i32.and
-      (i32.or
-       (local.get $0)
-       (local.get $1)
-      )
-      (i32.const 15)
-     )
-     (then
-      (local.set $2
-       (i32.load8_u
-        (local.get $0)
-       )
-      )
-      (br $block
-       (i32.load8_u
-        (local.get $1)
+  (block $block
+   (br_if $block
+    (i32.lt_u
+     (local.tee $2
+      (i32.sub
+       (i32.shl
+        (memory.size)
+        (i32.const 16)
        )
+       (i32.const 16)
       )
      )
+     (local.get $0)
     )
-    (if
+   )
+   (br_if $block
+    (i32.gt_u
+     (local.get $1)
+     (local.get $2)
+    )
+   )
+   (loop $label
+    (br_if $block
      (v128.any_true
       (v128.xor
-       (local.tee $5
-        (v128.load
-         (local.get $1)
-        )
+       (v128.load align=1
+        (local.get $1)
        )
        (local.tee $4
-        (v128.load
+        (v128.load align=1
          (local.get $0)
         )
        )
       )
      )
+    )
+    (if
+     (i32.eqz
+      (i8x16.all_true
+       (local.get $4)
+      )
+     )
      (then
-      (local.set $2
-       (i8x16.extract_lane_u 0
-        (local.get $4)
-       )
-      )
-      (br $block
-       (i8x16.extract_lane_u 0
-        (local.get $5)
-       )
+      (return
+       (i32.const 0)
       )
      )
     )
-    (loop $label
-     (if
-      (i32.eqz
-       (i8x16.all_true
-        (local.get $4)
-       )
-      )
-      (then
-       (return
-        (i32.const 0)
-       )
-      )
-     )
-     (local.set $4
-      (v128.load offset=16
-       (local.get $0)
-      )
-     )
-     (local.set $5
-      (v128.load offset=16
-       (local.get $1)
-      )
-     )
-     (local.set $1
-      (i32.add
-       (local.get $1)
-       (i32.const 16)
-      )
-     )
-     (local.set $0
-      (i32.add
-       (local.get $0)
-       (i32.const 16)
-      )
-     )
-     (br_if $label
-      (i32.eqz
-       (v128.any_true
-        (v128.xor
-         (local.get $5)
-         (local.get $4)
-        )
-       )
-      )
+    (local.set $1
+     (i32.add
+      (local.get $1)
+      (i32.const 16)
      )
     )
-    (local.set $2
-     (i8x16.extract_lane_u 0
-      (local.get $4)
+    (br_if $block
+     (i32.gt_u
+      (local.tee $0
+       (i32.add
+        (local.get $0)
+        (i32.const 16)
+       )
+      )
+      (local.get $2)
      )
     )
-    (i8x16.extract_lane_u 0
-     (local.get $5)
+    (br_if $label
+     (i32.le_u
+      (local.get $1)
+      (local.get $2)
+     )
     )
    )
   )
   (if
    (i32.eq
-    (i32.and
-     (local.get $2)
-     (i32.const 255)
+    (local.tee $2
+     (i32.load8_u
+      (local.get $0)
+     )
     )
-    (i32.and
-     (local.get $3)
-     (i32.const 255)
+    (local.tee $3
+     (i32.load8_u
+      (local.get $1)
+     )
     )
    )
    (then
@@ -441,16 +442,10 @@
       (i32.const 1)
      )
     )
-    (local.set $2
-     (local.get $3)
-    )
     (loop $label1
      (if
       (i32.eqz
-       (i32.and
-        (local.get $2)
-        (i32.const 255)
-       )
+       (local.get $2)
       )
       (then
        (return
@@ -490,81 +485,95 @@
    )
   )
   (i32.sub
-   (i32.and
-    (local.get $2)
-    (i32.const 255)
-   )
-   (i32.and
-    (local.get $3)
-    (i32.const 255)
-   )
+   (local.get $2)
+   (local.get $3)
   )
  )
  (func $strncmp (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
   (local $3 i32)
   (local $4 i32)
   (local $5 v128)
-  (block $block
-   (if
-    (i32.ge_u
-     (local.get $2)
-     (i32.const 16)
+  (block $block1
+   (block $block
+    (br_if $block
+     (i32.lt_u
+      (local.tee $3
+       (i32.sub
+        (i32.shl
+         (memory.size)
+         (i32.const 16)
+        )
+        (i32.const 16)
+       )
+      )
+      (local.get $0)
+     )
     )
-    (then
-     (loop $label
-      (br_if $block
-       (v128.any_true
-        (v128.xor
+    (loop $label
+     (br_if $block
+      (i32.gt_u
+       (local.get $1)
+       (local.get $3)
+      )
+     )
+     (br_if $block
+      (i32.lt_u
+       (local.get $2)
+       (i32.const 16)
+      )
+     )
+     (br_if $block1
+      (v128.any_true
+       (v128.xor
+        (v128.load align=1
+         (local.get $1)
+        )
+        (local.tee $5
          (v128.load align=1
-          (local.get $1)
-         )
-         (local.tee $5
-          (v128.load align=1
-           (local.get $0)
-          )
+          (local.get $0)
          )
         )
        )
       )
-      (if
-       (i32.eqz
-        (i8x16.all_true
-         (local.get $5)
-        )
-       )
-       (then
-        (return
-         (i32.const 0)
-        )
+     )
+     (if
+      (i32.eqz
+       (i8x16.all_true
+        (local.get $5)
        )
       )
-      (local.set $1
-       (i32.add
-        (local.get $1)
-        (i32.const 16)
+      (then
+       (return
+        (i32.const 0)
        )
       )
-      (local.set $0
-       (i32.add
-        (local.get $0)
-        (i32.const 16)
-       )
+     )
+     (local.set $2
+      (i32.sub
+       (local.get $2)
+       (i32.const 16)
       )
-      (br_if $label
-       (i32.gt_u
-        (local.tee $2
-         (i32.sub
-          (local.get $2)
-          (i32.const 16)
-         )
+     )
+     (local.set $1
+      (i32.add
+       (local.get $1)
+       (i32.const 16)
+      )
+     )
+     (br_if $label
+      (i32.le_u
+       (local.tee $0
+        (i32.add
+         (local.get $0)
+         (i32.const 16)
         )
-        (i32.const 15)
        )
+       (local.get $3)
       )
      )
     )
    )
-   (br_if $block
+   (br_if $block1
     (local.get $2)
    )
    (return
diff --git a/sqlite3/sqlite_opt.h b/sqlite3/sqlite_opt.h
index 0db37c3..73b1480 100644
--- a/sqlite3/sqlite_opt.h
+++ b/sqlite3/sqlite_opt.h
@@ -4,15 +4,15 @@
 #define SQLITE_THREADSAFE 0
 #define SQLITE_DEFAULT_WAL_SYNCHRONOUS 1
 #define SQLITE_LIKE_DOESNT_MATCH_BLOBS
-#define SQLITE_MAX_EXPR_DEPTH 0
 #define SQLITE_STRICT_SUBTYPE 1
-#define SQLITE_USE_ALLOCA
 #define SQLITE_OMIT_DEPRECATED
 #define SQLITE_OMIT_SHARED_CACHE
 #define SQLITE_OMIT_AUTOINIT
 
 // We need these:
 // #define SQLITE_DEFAULT_MEMSTATUS 0
+// #define SQLITE_MAX_EXPR_DEPTH 0
+// #define SQLITE_USE_ALLOCA
 // #define SQLITE_OMIT_DECLTYPE
 // #define SQLITE_OMIT_PROGRESS_CALLBACK
 
diff --git a/sqlite3/strings.c b/sqlite3/strings.c
index 24cb9d7..cdf5389 100644
--- a/sqlite3/strings.c
+++ b/sqlite3/strings.c
@@ -1,6 +1,7 @@
 #include <stddef.h>
 #include <stdint.h>
 #include <wasm_simd128.h>
+#include <__macro_PAGESIZE.h>
 
 #ifdef __wasm_bulk_memory__
 
@@ -42,23 +43,25 @@ int memcmp(const void *v1, const void *v2, size_t n) {
 }
 
 void *memchr(const void *v, int c, size_t n) {
-  c = (uint8_t)c;
-
+  uintptr_t align = (uintptr_t)v % sizeof(v128_t);
+  const v128_t *w = (void *)(v - align);
   const v128_t wc = wasm_i8x16_splat(c);
-  const v128_t *w = (void *)v;
-  for (; n >= sizeof(v128_t); n -= sizeof(v128_t)) {
-    if (wasm_v128_any_true(wasm_i8x16_eq(wasm_v128_load(w), wc))) {
-      break;
+
+  while (true) {
+    const v128_t cmp = 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 (void *)w + __builtin_ctz(mask);
+      }
     }
+    if (__builtin_sub_overflow(n, sizeof(v128_t) - align, &n)) {
+      return NULL;
+    }
+    align = 0;
     w++;
   }
-
-  const uint8_t *u = (void *)w;
-  while (n--) {
-    if (*u == c) return (void *)u;
-    u++;
-  }
-  return 0;
 }
 
 size_t strlen(const char *s) {
@@ -80,32 +83,40 @@ size_t strlen(const char *s) {
 }
 
 int strcmp(const char *s1, const char *s2) {
+  const v128_t *const limit =
+      (v128_t *)(__builtin_wasm_memory_size(0) * PAGESIZE) - 1;
+
   const v128_t *w1 = (void *)s1;
   const v128_t *w2 = (void *)s2;
-  if (((uintptr_t)s1 | (uintptr_t)s2) % sizeof(v128_t) == 0) {
-    while (!wasm_v128_any_true(*w1 ^ *w2)) {
-      if (!wasm_i8x16_all_true(*w1)) {
-        return 0;
-      }
-      w1++;
-      w2++;
+  while (w1 <= limit && w2 <= limit) {
+    if (wasm_v128_any_true(wasm_v128_load(w1) ^ wasm_v128_load(w2))) {
+      break;
     }
+    if (!wasm_i8x16_all_true(wasm_v128_load(w1))) {
+      return 0;
+    }
+    w1++;
+    w2++;
   }
 
   const uint8_t *u1 = (void *)w1;
   const uint8_t *u2 = (void *)w2;
   while (true) {
     if (*u1 != *u2) return *u1 - *u2;
-    if (*u1 == 0) return 0;
+    if (*u1 == 0) break;
     u1++;
     u2++;
   }
+  return 0;
 }
 
 int strncmp(const char *s1, const char *s2, size_t n) {
+  const v128_t *const limit =
+      (v128_t *)(__builtin_wasm_memory_size(0) * PAGESIZE) - 1;
+
   const v128_t *w1 = (void *)s1;
   const v128_t *w2 = (void *)s2;
-  for (; n >= sizeof(v128_t); n -= sizeof(v128_t)) {
+  for (; w1 <= limit && w2 <= limit && n >= sizeof(v128_t); n -= sizeof(v128_t)) {
     if (wasm_v128_any_true(wasm_v128_load(w1) ^ wasm_v128_load(w2))) {
       break;
     }
@@ -128,9 +139,7 @@ int strncmp(const char *s1, const char *s2, size_t n) {
 }
 
 char *strchrnul(const char *s, int c) {
-  c = (char)c;
-
-  if (__builtin_constant_p(c) && c == 0) {
+  if (__builtin_constant_p(c) && (char)c == 0) {
     return (char *)s + strlen(s);
   }
 
@@ -154,7 +163,7 @@ char *strchrnul(const char *s, int c) {
 
 char *strchr(const char *s, int c) {
   char *r = strchrnul(s, c);
-  return *(char *)r == (char)c ? r : 0;
+  return *(char *)r == (char)c ? r : NULL;
 }
 
 #endif
diff --git a/sqlite3/vfs.c b/sqlite3/vfs.c
index e69396e..e659f3b 100644
--- a/sqlite3/vfs.c
+++ b/sqlite3/vfs.c
@@ -137,9 +137,10 @@ sqlite3_vfs *sqlite3_vfs_find(const char *zVfsName) {
 
   // Create a new C wrapper.
   sqlite3_vfs *head = go_vfs_list;
-  go_vfs_list = malloc(sizeof(sqlite3_vfs) + strlen(zVfsName) + 1);
+  size_t vfsNameLen = strlen(zVfsName);
+  go_vfs_list = malloc(sizeof(sqlite3_vfs) + vfsNameLen + 1);
   char *name = (char *)(go_vfs_list + 1);
-  strcpy(name, zVfsName);
+  memcpy(name, zVfsName, vfsNameLen + 1);
   *go_vfs_list = (sqlite3_vfs){
       .iVersion = 2,
       .szOsFile = sizeof(struct go_file),