From 60ab485b2968c6499f496352fc4d28e80ab3d130 Mon Sep 17 00:00:00 2001
From: Nuno Cruces <ncruces@users.noreply.github.com>
Date: Fri, 9 May 2025 11:17:22 +0100
Subject: [PATCH] Comments.

---
 sqlite3/libc/libc.wasm |  Bin 5536 -> 5636 bytes
 sqlite3/libc/libc.wat  | 1378 +++++++++++++++++++++-------------------
 sqlite3/libc/string.h  |   69 +-
 3 files changed, 790 insertions(+), 657 deletions(-)

diff --git a/sqlite3/libc/libc.wasm b/sqlite3/libc/libc.wasm
index 9d2dab0530d96f35f707239367b44cd71cf91cc4..f2f37223e1f63b817318cc255307e723ce644b3f 100755
GIT binary patch
delta 1317
zcmZ`(!HN_y5KYoa_w;n9yJyA2xR>e(UQ`5+A~;!Dmvx!lgBMSNAfnwD2fcY{1rhWl
z%nyh+FCy$Wc=Uff81*GR?##~4HWZyoRlQVI>gDs*ch|m8xcG79c!&Q=yN{Vjy1{GW
z6z$82Wo0HNUTMw1mIoBs>S(qrC&cbBTfM@*EnAaLu_IvQt%)0#Fji+I!e5iHys49D
z0-TH(MGd?hW6WJ9O{`zJLsPg1yP7?w!aXq(s|aV1eAXC+cuFX)*sfM25lwPtq-I%$
z#JH_mqZX-BMCzc=npDjRDK>^tI)yWQqqZJo^9$9@ZZOp%ml^zl*BH4M&`_-XM7HY-
zNhq;@_<f_c?HlnsW5_?ks<n*Nwuvb@(u`u;y+lc7h`xj##7!+Jb|aIyJZFq)NC7D!
z?i;TQf}tBgOCdu@wy;#*BB3Zq3_%-1$jBS8W*j0Q5yWF{QA|VoVrJ5&Mm}y-Yf5W{
z&Pm@^Q@Rh_edQ*1+YE&@==YYJBC>P&UpTCaJA`!<tecGPJ%_zI*A8QGt`WiRDH`$+
zXA+X(H$paHgZmA!)P>ihr5;g9*?gwQcTcJaaWVic{nQ#xyoQqmoS%B!oyHlg;iPLg
z!!?}D<2aq`!r$o0x9D{FztbVdoS^e%j@s+DSBfqC#!9gz?kBGRx+?umON??Y=WmNk
z^Do7?D>nNTqyiuOyr(q0r#W~?VJ0JrTPxarJg&2TAk;0~h^GLC7CZx!Y`5UT27V_e
z)b3Wf9iQbk*nFDZRbV#;XcM_HyPn+gG`TIS6@lOwtc#^@qE8xpmaXB>_qBi}@;?-C
whUAFl5niuaiV5f-;DYGjivONQ=nC@*#K&8KF0lSk|0Zz4Bf+xcN7at{2Th0wN&o-=

delta 1165
zcmbVMJ&P1U5bdw|?%tUV7IZJEZZT33%rtT5;9TT5FEBCFKty&E56lea0y!`f<_Bmb
zn22z~<^B`@g@KK}nmcdT^-LTnx~seT^{cLW^ZNSxYoBLQeY<+REx#Athf<|o{)!~&
z`)1aPx>7SWBi+_}W=7(!Xq%U$mu<5Q=`L2JH&#`oNH1EsjF}L=BsGi{r_|DCq*}S~
zxn`V_#95>eB3>a&A`Y|oAh`H&r*nc_d>YJ-6B_rCdA<k<&QnfVBX(Ru2KSIT2*!!3
zB1w|Dm@~4Z$$%E~CgUZg1j36T_7QAl)xM?nCsG4MNGZj9f2%!IKPmxkx^3niQcC04
z+h{F#IoJqR!@P?&$5_BhSjB5~v<71WOJPV~_$KmEAz=!pVSbFuUL1tXC#1YVW`@%a
zqxeR~k1|woOMu&gr0t)$sX38k`$IP^PGmg7oa+9ayOi=WIi7aIMtl~^_y|A*0672*
zzGPpme8s+A`I<&lMrV?g37>zPWCkp=ft?K4-3lz-0MAar^P=PzI{z*N&o{t}0eppB
zZeUlZ*w*!5Mw7>*->)_hwFH0d5ghw#k1%3??Gb7kN85nky2VctwKC`#Z~C^Lb|~=r
zKF>L(Mwec=;V9vLOwey=zU-cpERJk}!fVQwEC?=tazhLS=WaB(j~nq0>d38*Q10<y
z+HMX#R8ZhQgg0RShwvb^Kc!bgdQe{dm-6ho^!l64<hF~GnhV4UKv@NNlH7HAP#y+W
oouC4mqjZniLx}ic__UweTbHq<tgbj1$id7xf`m_wzc$;+FMZA2%K!iX

diff --git a/sqlite3/libc/libc.wat b/sqlite3/libc/libc.wat
index 1dbd28a..460ac39 100644
--- a/sqlite3/libc/libc.wat
+++ b/sqlite3/libc/libc.wat
@@ -2010,7 +2010,7 @@
   (local $7 i32)
   (local $8 i32)
   (local $9 i32)
-  (local $10 v128)
+  (local $10 i32)
   (local $11 v128)
   (local $12 v128)
   (local $13 v128)
@@ -2018,6 +2018,7 @@
   (local $15 v128)
   (local $16 v128)
   (local $17 v128)
+  (local $18 v128)
   (if
    (i32.lt_u
     (local.get $1)
@@ -2029,562 +2030,30 @@
     )
    )
   )
-  (block $block2
-   (block $block
-    (if
-     (i32.le_u
-      (local.get $3)
-      (i32.const 15)
-     )
-     (then
-      (if
-       (i32.gt_u
-        (local.get $0)
-        (local.tee $6
-         (i32.sub
-          (i32.sub
-           (i32.shl
-            (memory.size)
-            (i32.const 16)
-           )
-           (local.get $3)
-          )
-          (i32.const 16)
-         )
-        )
-       )
-       (then
-        (local.set $4
-         (local.get $1)
-        )
-        (br $block)
-       )
-      )
-      (local.set $10
-       (v128.load8_splat
+  (block $block1
+   (if
+    (i32.eqz
+     (i32.and
+      (i32.gt_u
+       (local.tee $8
         (i32.sub
-         (i32.add
-          (local.get $2)
-          (local.get $3)
-         )
-         (i32.const 1)
-        )
-       )
-      )
-      (local.set $12
-       (v128.load8_splat
-        (local.get $2)
-       )
-      )
-      (local.set $7
-       (i32.sub
-        (local.get $3)
-        (i32.const 2)
-       )
-      )
-      (local.set $8
-       (i32.add
-        (local.get $2)
-        (i32.const 1)
-       )
-      )
-      (loop $label1
-       (block $block1
-        (br_if $block1
-         (i32.eqz
-          (v128.any_true
-           (local.tee $13
-            (v128.and
-             (i8x16.eq
-              (local.get $10)
-              (v128.load align=1
-               (i32.sub
-                (i32.add
-                 (local.get $0)
-                 (local.get $3)
-                )
-                (i32.const 1)
-               )
-              )
-             )
-             (i8x16.eq
-              (local.get $12)
-              (v128.load align=1
-               (local.get $0)
-              )
-             )
-            )
-           )
-          )
-         )
-        )
-        (br_if $block1
-         (i32.eqz
-          (local.tee $4
-           (i8x16.bitmask
-            (local.get $13)
-           )
-          )
-         )
-        )
-        (loop $label
-         (br_if $block2
-          (i32.eqz
-           (call $bcmp
-            (i32.add
-             (local.tee $5
-              (i32.add
-               (local.get $0)
-               (i32.ctz
-                (local.get $4)
-               )
-              )
-             )
-             (i32.const 1)
-            )
-            (local.get $8)
-            (local.get $7)
-           )
-          )
-         )
-         (br_if $label
-          (local.tee $4
-           (i32.and
-            (i32.sub
-             (local.get $4)
-             (i32.const 1)
-            )
-            (local.get $4)
-           )
-          )
-         )
-        )
-       )
-       (if
-        (i32.lt_u
          (local.get $1)
-         (local.tee $4
-          (i32.sub
-           (local.get $1)
-           (i32.const 16)
-          )
-         )
-        )
-        (then
-         (return
-          (i32.const 0)
-         )
-        )
-       )
-       (if
-        (i32.gt_u
          (local.get $3)
-         (local.get $4)
-        )
-        (then
-         (return
-          (i32.const 0)
-         )
         )
        )
-       (local.set $1
-        (local.get $4)
-       )
-       (br_if $label1
-        (i32.ge_u
-         (local.get $6)
-         (local.tee $0
-          (i32.add
-           (local.get $0)
-           (i32.const 16)
-          )
-         )
-        )
-       )
-      )
-      (br $block)
-     )
-    )
-    (memory.fill
-     (i32.const 4112)
-     (select
-      (i32.const 255)
-      (local.tee $6
-       (i32.sub
-        (local.get $3)
-        (i32.const 1)
-       )
+       (i32.const 15)
       )
       (i32.ge_u
-       (local.get $6)
-       (i32.const 255)
-      )
-     )
-     (i32.const 256)
-    )
-    (block $block3
-     (if
-      (i32.ne
        (local.get $3)
        (i32.const 16)
       )
-      (then
-       (local.set $12
-        (i32x4.splat
-         (local.get $6)
-        )
-       )
-       (local.set $13
-        (v128.const i32x4 0x0000000c 0x0000000d 0x0000000e 0x0000000f)
-       )
-       (local.set $15
-        (v128.const i32x4 0x00000008 0x00000009 0x0000000a 0x0000000b)
-       )
-       (local.set $16
-        (v128.const i32x4 0x00000004 0x00000005 0x00000006 0x00000007)
-       )
-       (local.set $17
-        (v128.const i32x4 0x00000000 0x00000001 0x00000002 0x00000003)
-       )
-       (local.set $5
-        (local.tee $7
-         (i32.and
-          (local.get $6)
-          (i32.const -16)
-         )
-        )
-       )
-       (local.set $4
-        (local.get $2)
-       )
-       (loop $label2
-        (v128.store8_lane 0
-         (i32.add
-          (i32x4.extract_lane 0
-           (local.tee $11
-            (i32x4.extend_low_i16x8_u
-             (i16x8.extend_low_i8x16_u
-              (local.tee $14
-               (v128.load align=1
-                (local.get $4)
-               )
-              )
-             )
-            )
-           )
-          )
-          (i32.const 4112)
-         )
-         (local.tee $10
-          (i8x16.narrow_i16x8_u
-           (i16x8.narrow_i32x4_u
-            (i32x4.min_u
-             (i32x4.add
-              (local.get $12)
-              (v128.not
-               (local.get $17)
-              )
-             )
-             (v128.const i32x4 0x000000ff 0x000000ff 0x000000ff 0x000000ff)
-            )
-            (i32x4.min_u
-             (i32x4.add
-              (local.get $12)
-              (v128.not
-               (local.get $16)
-              )
-             )
-             (v128.const i32x4 0x000000ff 0x000000ff 0x000000ff 0x000000ff)
-            )
-           )
-           (i16x8.narrow_i32x4_u
-            (i32x4.min_u
-             (i32x4.add
-              (local.get $12)
-              (v128.not
-               (local.get $15)
-              )
-             )
-             (v128.const i32x4 0x000000ff 0x000000ff 0x000000ff 0x000000ff)
-            )
-            (i32x4.min_u
-             (i32x4.add
-              (local.get $12)
-              (v128.not
-               (local.get $13)
-              )
-             )
-             (v128.const i32x4 0x000000ff 0x000000ff 0x000000ff 0x000000ff)
-            )
-           )
-          )
-         )
-        )
-        (v128.store8_lane 1
-         (i32.add
-          (i32x4.extract_lane 1
-           (local.get $11)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 2
-         (i32.add
-          (i32x4.extract_lane 2
-           (local.get $11)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 3
-         (i32.add
-          (i32x4.extract_lane 3
-           (local.get $11)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 4
-         (i32.add
-          (i32x4.extract_lane 0
-           (local.tee $11
-            (i32x4.extend_low_i16x8_u
-             (i16x8.extend_low_i8x16_u
-              (i8x16.shuffle 4 5 6 7 0 0 0 0 0 0 0 0 0 0 0 0
-               (local.get $14)
-               (local.get $10)
-              )
-             )
-            )
-           )
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 5
-         (i32.add
-          (i32x4.extract_lane 1
-           (local.get $11)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 6
-         (i32.add
-          (i32x4.extract_lane 2
-           (local.get $11)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 7
-         (i32.add
-          (i32x4.extract_lane 3
-           (local.get $11)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 8
-         (i32.add
-          (i32x4.extract_lane 0
-           (local.tee $11
-            (i32x4.extend_low_i16x8_u
-             (i16x8.extend_low_i8x16_u
-              (i8x16.shuffle 8 9 10 11 0 0 0 0 0 0 0 0 0 0 0 0
-               (local.get $14)
-               (local.get $10)
-              )
-             )
-            )
-           )
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 9
-         (i32.add
-          (i32x4.extract_lane 1
-           (local.get $11)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 10
-         (i32.add
-          (i32x4.extract_lane 2
-           (local.get $11)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 11
-         (i32.add
-          (i32x4.extract_lane 3
-           (local.get $11)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 12
-         (i32.add
-          (i32x4.extract_lane 0
-           (local.tee $14
-            (i32x4.extend_low_i16x8_u
-             (i16x8.extend_low_i8x16_u
-              (i8x16.shuffle 12 13 14 15 0 0 0 0 0 0 0 0 0 0 0 0
-               (local.get $14)
-               (local.get $10)
-              )
-             )
-            )
-           )
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 13
-         (i32.add
-          (i32x4.extract_lane 1
-           (local.get $14)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 14
-         (i32.add
-          (i32x4.extract_lane 2
-           (local.get $14)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (v128.store8_lane 15
-         (i32.add
-          (i32x4.extract_lane 3
-           (local.get $14)
-          )
-          (i32.const 4112)
-         )
-         (local.get $10)
-        )
-        (local.set $17
-         (i32x4.add
-          (local.get $17)
-          (v128.const i32x4 0x00000010 0x00000010 0x00000010 0x00000010)
-         )
-        )
-        (local.set $4
-         (i32.add
-          (local.get $4)
-          (i32.const 16)
-         )
-        )
-        (local.set $16
-         (i32x4.add
-          (local.get $16)
-          (v128.const i32x4 0x00000010 0x00000010 0x00000010 0x00000010)
-         )
-        )
-        (local.set $15
-         (i32x4.add
-          (local.get $15)
-          (v128.const i32x4 0x00000010 0x00000010 0x00000010 0x00000010)
-         )
-        )
-        (local.set $13
-         (i32x4.add
-          (local.get $13)
-          (v128.const i32x4 0x00000010 0x00000010 0x00000010 0x00000010)
-         )
-        )
-        (br_if $label2
-         (local.tee $5
-          (i32.sub
-           (local.get $5)
-           (i32.const 16)
-          )
-         )
-        )
-       )
-       (br_if $block3
-        (i32.eq
-         (local.get $6)
-         (local.get $7)
-        )
-       )
-      )
-     )
-     (local.set $5
-      (i32.add
-       (local.get $2)
-       (local.get $7)
-      )
-     )
-     (local.set $4
-      (i32.sub
-       (i32.sub
-        (local.get $3)
-        (local.get $7)
-       )
-       (i32.const 2)
-      )
-     )
-     (loop $label3
-      (i32.store8
-       (i32.add
-        (i32.load8_u
-         (local.get $5)
-        )
-        (i32.const 4112)
-       )
-       (select
-        (i32.const 255)
-        (local.get $4)
-        (i32.ge_u
-         (local.get $4)
-         (i32.const 255)
-        )
-       )
-      )
-      (local.set $5
-       (i32.add
-        (local.get $5)
-        (i32.const 1)
-       )
-      )
-      (br_if $label3
-       (i32.ne
-        (local.tee $4
-         (i32.sub
-          (local.get $4)
-          (i32.const 1)
-         )
-        )
-        (i32.const -1)
-       )
-      )
      )
     )
-    (block $block4
+    (then
      (if
-      (i32.gt_u
+      (i32.le_u
        (local.get $0)
-       (local.tee $6
+       (local.tee $9
         (i32.sub
          (i32.sub
           (i32.shl
@@ -2593,55 +2062,686 @@
           )
           (local.get $3)
          )
-         (i32.const 16)
+         (i32.const 15)
         )
        )
       )
       (then
-       (local.set $4
-        (local.get $1)
+       (local.set $11
+        (v128.load8_splat
+         (i32.add
+          (local.get $2)
+          (local.tee $6
+           (i32.sub
+            (local.get $3)
+            (i32.const 1)
+           )
+          )
+         )
+        )
        )
-       (br $block4)
-      )
-     )
-     (local.set $10
-      (v128.load8_splat
-       (i32.sub
+       (local.set $13
+        (v128.load8_splat
+         (local.get $2)
+        )
+       )
+       (local.set $7
         (i32.add
          (local.get $2)
+         (i32.const 1)
+        )
+       )
+       (loop $label1
+        (block $block
+         (br_if $block
+          (i32.eqz
+           (v128.any_true
+            (local.tee $14
+             (v128.and
+              (i8x16.eq
+               (local.get $11)
+               (v128.load align=1
+                (i32.sub
+                 (i32.add
+                  (local.get $0)
+                  (local.get $3)
+                 )
+                 (i32.const 1)
+                )
+               )
+              )
+              (i8x16.eq
+               (local.get $13)
+               (v128.load align=1
+                (local.get $0)
+               )
+              )
+             )
+            )
+           )
+          )
+         )
+         (br_if $block
+          (i32.eqz
+           (local.tee $4
+            (i8x16.bitmask
+             (local.get $14)
+            )
+           )
+          )
+         )
+         (loop $label
+          (br_if $block1
+           (i32.eqz
+            (call $bcmp
+             (i32.add
+              (local.tee $5
+               (i32.add
+                (local.get $0)
+                (i32.ctz
+                 (local.get $4)
+                )
+               )
+              )
+              (i32.const 1)
+             )
+             (local.get $7)
+             (local.get $6)
+            )
+           )
+          )
+          (br_if $label
+           (local.tee $4
+            (i32.and
+             (i32.sub
+              (local.get $4)
+              (i32.const 1)
+             )
+             (local.get $4)
+            )
+           )
+          )
+         )
+        )
+        (if
+         (i32.lt_u
+          (local.get $1)
+          (local.tee $5
+           (i32.sub
+            (local.get $1)
+            (i32.const 16)
+           )
+          )
+         )
+         (then
+          (return
+           (i32.const 0)
+          )
+         )
+        )
+        (if
+         (i32.gt_u
+          (local.get $3)
+          (local.get $5)
+         )
+         (then
+          (return
+           (i32.const 0)
+          )
+         )
+        )
+        (local.set $1
+         (local.get $5)
+        )
+        (br_if $label1
+         (i32.le_u
+          (local.tee $0
+           (i32.add
+            (local.get $0)
+            (i32.const 16)
+           )
+          )
+          (local.get $9)
+         )
+        )
+       )
+       (local.set $8
+        (i32.sub
+         (local.get $1)
          (local.get $3)
         )
-        (i32.const 1)
        )
       )
      )
-     (local.set $12
+     (local.set $1
+      (i32.const 0)
+     )
+     (loop $label2
+      (local.set $4
+       (i32.const 0)
+      )
+      (loop $label3
+       (if
+        (i32.ne
+         (i32.load8_u
+          (i32.add
+           (local.get $2)
+           (local.get $4)
+          )
+         )
+         (i32.load8_u
+          (i32.add
+           (local.get $0)
+           (local.get $4)
+          )
+         )
+        )
+        (then
+         (local.set $0
+          (i32.add
+           (local.get $0)
+           (i32.const 1)
+          )
+         )
+         (local.set $5
+          (i32.const 0)
+         )
+         (br_if $label2
+          (i32.le_u
+           (local.tee $1
+            (i32.add
+             (local.get $1)
+             (i32.const 1)
+            )
+           )
+           (local.get $8)
+          )
+         )
+         (br $block1)
+        )
+       )
+       (br_if $label3
+        (i32.ne
+         (local.get $3)
+         (local.tee $4
+          (i32.add
+           (local.get $4)
+           (i32.const 1)
+          )
+         )
+        )
+       )
+      )
+     )
+     (return
+      (local.get $0)
+     )
+    )
+   )
+   (memory.fill
+    (i32.const 4112)
+    (select
+     (i32.const -1)
+     (local.tee $7
+      (i32.sub
+       (local.get $3)
+       (i32.const 1)
+      )
+     )
+     (local.tee $5
+      (i32.gt_s
+       (local.get $7)
+       (i32.const 254)
+      )
+     )
+    )
+    (i32.const 256)
+   )
+   (block $block2
+    (br_if $block2
+     (i32.ge_u
+      (local.tee $6
+       (select
+        (i32.sub
+         (local.get $3)
+         (i32.const 256)
+        )
+        (i32.const 0)
+        (local.get $5)
+       )
+      )
+      (local.get $7)
+     )
+    )
+    (if
+     (i32.ge_u
+      (local.tee $10
+       (i32.add
+        (i32.xor
+         (local.get $6)
+         (i32.const -1)
+        )
+        (local.get $3)
+       )
+      )
+      (i32.const 16)
+     )
+     (then
+      (local.set $4
+       (i32.add
+        (local.get $2)
+        (local.get $6)
+       )
+      )
+      (local.set $14
+       (i32x4.add
+        (local.tee $11
+         (i32x4.splat
+          (local.get $6)
+         )
+        )
+        (v128.const i32x4 0x0000000c 0x0000000d 0x0000000e 0x0000000f)
+       )
+      )
+      (local.set $16
+       (i32x4.add
+        (local.get $11)
+        (v128.const i32x4 0x00000008 0x00000009 0x0000000a 0x0000000b)
+       )
+      )
+      (local.set $17
+       (i32x4.add
+        (local.get $11)
+        (v128.const i32x4 0x00000004 0x00000005 0x00000006 0x00000007)
+       )
+      )
+      (local.set $18
+       (i32x4.add
+        (local.get $11)
+        (v128.const i32x4 0x00000000 0x00000001 0x00000002 0x00000003)
+       )
+      )
+      (local.set $6
+       (i32.add
+        (local.get $6)
+        (local.tee $9
+         (i32.and
+          (local.get $10)
+          (i32.const -16)
+         )
+        )
+       )
+      )
+      (local.set $13
+       (i32x4.splat
+        (local.get $7)
+       )
+      )
+      (local.set $5
+       (local.get $9)
+      )
+      (loop $label4
+       (v128.store8_lane 0
+        (i32.add
+         (i32x4.extract_lane 0
+          (local.tee $12
+           (i32x4.extend_low_i16x8_u
+            (i16x8.extend_low_i8x16_u
+             (local.tee $15
+              (v128.load align=1
+               (local.get $4)
+              )
+             )
+            )
+           )
+          )
+         )
+         (i32.const 4112)
+        )
+        (local.tee $11
+         (i8x16.narrow_i16x8_u
+          (i16x8.narrow_i32x4_u
+           (v128.and
+            (i32x4.add
+             (local.get $13)
+             (v128.not
+              (local.get $18)
+             )
+            )
+            (v128.const i32x4 0x000000ff 0x000000ff 0x000000ff 0x000000ff)
+           )
+           (v128.and
+            (i32x4.add
+             (local.get $13)
+             (v128.not
+              (local.get $17)
+             )
+            )
+            (v128.const i32x4 0x000000ff 0x000000ff 0x000000ff 0x000000ff)
+           )
+          )
+          (i16x8.narrow_i32x4_u
+           (v128.and
+            (i32x4.add
+             (local.get $13)
+             (v128.not
+              (local.get $16)
+             )
+            )
+            (v128.const i32x4 0x000000ff 0x000000ff 0x000000ff 0x000000ff)
+           )
+           (v128.and
+            (i32x4.add
+             (local.get $13)
+             (v128.not
+              (local.get $14)
+             )
+            )
+            (v128.const i32x4 0x000000ff 0x000000ff 0x000000ff 0x000000ff)
+           )
+          )
+         )
+        )
+       )
+       (v128.store8_lane 1
+        (i32.add
+         (i32x4.extract_lane 1
+          (local.get $12)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 2
+        (i32.add
+         (i32x4.extract_lane 2
+          (local.get $12)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 3
+        (i32.add
+         (i32x4.extract_lane 3
+          (local.get $12)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 4
+        (i32.add
+         (i32x4.extract_lane 0
+          (local.tee $12
+           (i32x4.extend_low_i16x8_u
+            (i16x8.extend_low_i8x16_u
+             (i8x16.shuffle 4 5 6 7 0 0 0 0 0 0 0 0 0 0 0 0
+              (local.get $15)
+              (local.get $11)
+             )
+            )
+           )
+          )
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 5
+        (i32.add
+         (i32x4.extract_lane 1
+          (local.get $12)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 6
+        (i32.add
+         (i32x4.extract_lane 2
+          (local.get $12)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 7
+        (i32.add
+         (i32x4.extract_lane 3
+          (local.get $12)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 8
+        (i32.add
+         (i32x4.extract_lane 0
+          (local.tee $12
+           (i32x4.extend_low_i16x8_u
+            (i16x8.extend_low_i8x16_u
+             (i8x16.shuffle 8 9 10 11 0 0 0 0 0 0 0 0 0 0 0 0
+              (local.get $15)
+              (local.get $11)
+             )
+            )
+           )
+          )
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 9
+        (i32.add
+         (i32x4.extract_lane 1
+          (local.get $12)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 10
+        (i32.add
+         (i32x4.extract_lane 2
+          (local.get $12)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 11
+        (i32.add
+         (i32x4.extract_lane 3
+          (local.get $12)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 12
+        (i32.add
+         (i32x4.extract_lane 0
+          (local.tee $15
+           (i32x4.extend_low_i16x8_u
+            (i16x8.extend_low_i8x16_u
+             (i8x16.shuffle 12 13 14 15 0 0 0 0 0 0 0 0 0 0 0 0
+              (local.get $15)
+              (local.get $11)
+             )
+            )
+           )
+          )
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 13
+        (i32.add
+         (i32x4.extract_lane 1
+          (local.get $15)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 14
+        (i32.add
+         (i32x4.extract_lane 2
+          (local.get $15)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (v128.store8_lane 15
+        (i32.add
+         (i32x4.extract_lane 3
+          (local.get $15)
+         )
+         (i32.const 4112)
+        )
+        (local.get $11)
+       )
+       (local.set $18
+        (i32x4.add
+         (local.get $18)
+         (v128.const i32x4 0x00000010 0x00000010 0x00000010 0x00000010)
+        )
+       )
+       (local.set $4
+        (i32.add
+         (local.get $4)
+         (i32.const 16)
+        )
+       )
+       (local.set $17
+        (i32x4.add
+         (local.get $17)
+         (v128.const i32x4 0x00000010 0x00000010 0x00000010 0x00000010)
+        )
+       )
+       (local.set $16
+        (i32x4.add
+         (local.get $16)
+         (v128.const i32x4 0x00000010 0x00000010 0x00000010 0x00000010)
+        )
+       )
+       (local.set $14
+        (i32x4.add
+         (local.get $14)
+         (v128.const i32x4 0x00000010 0x00000010 0x00000010 0x00000010)
+        )
+       )
+       (br_if $label4
+        (local.tee $5
+         (i32.sub
+          (local.get $5)
+          (i32.const 16)
+         )
+        )
+       )
+      )
+      (br_if $block2
+       (i32.eq
+        (local.get $9)
+        (local.get $10)
+       )
+      )
+     )
+    )
+    (local.set $4
+     (i32.add
+      (local.get $2)
+      (local.get $6)
+     )
+    )
+    (local.set $5
+     (i32.sub
+      (i32.sub
+       (local.get $3)
+       (local.get $6)
+      )
+      (i32.const 2)
+     )
+    )
+    (loop $label5
+     (i32.store8
+      (i32.add
+       (i32.load8_u
+        (local.get $4)
+       )
+       (i32.const 4112)
+      )
+      (local.get $5)
+     )
+     (local.set $4
+      (i32.add
+       (local.get $4)
+       (i32.const 1)
+      )
+     )
+     (br_if $label5
+      (i32.ne
+       (local.tee $5
+        (i32.sub
+         (local.get $5)
+         (i32.const 1)
+        )
+       )
+       (i32.const -1)
+      )
+     )
+    )
+   )
+   (if
+    (i32.le_u
+     (local.get $0)
+     (local.tee $9
+      (i32.sub
+       (i32.sub
+        (i32.shl
+         (memory.size)
+         (i32.const 16)
+        )
+        (local.get $3)
+       )
+       (i32.const 15)
+      )
+     )
+    )
+    (then
+     (local.set $11
+      (v128.load8_splat
+       (i32.add
+        (local.get $2)
+        (local.get $7)
+       )
+      )
+     )
+     (local.set $13
       (v128.load8_splat
        (local.get $2)
       )
      )
-     (local.set $7
-      (i32.sub
-       (local.get $3)
-       (i32.const 2)
-      )
-     )
-     (local.set $8
+     (local.set $6
       (i32.add
        (local.get $2)
        (i32.const 1)
       )
      )
-     (loop $label5
-      (block $block5
-       (br_if $block5
+     (loop $label7
+      (block $block3
+       (br_if $block3
         (i32.eqz
          (v128.any_true
-          (local.tee $15
+          (local.tee $16
            (v128.and
             (i8x16.eq
-             (local.get $10)
-             (local.tee $13
+             (local.get $11)
+             (local.tee $14
               (v128.load align=1
                (i32.sub
                 (i32.add
@@ -2654,7 +2754,7 @@
              )
             )
             (i8x16.eq
-             (local.get $12)
+             (local.get $13)
              (v128.load align=1
               (local.get $0)
              )
@@ -2664,17 +2764,17 @@
          )
         )
        )
-       (br_if $block5
+       (br_if $block3
         (i32.eqz
          (local.tee $4
           (i8x16.bitmask
-           (local.get $15)
+           (local.get $16)
           )
          )
         )
        )
-       (loop $label4
-        (br_if $block2
+       (loop $label6
+        (br_if $block1
          (i32.eqz
           (call $bcmp
            (i32.add
@@ -2688,12 +2788,12 @@
             )
             (i32.const 1)
            )
-           (local.get $8)
+           (local.get $6)
            (local.get $7)
           )
          )
         )
-        (br_if $label4
+        (br_if $label6
          (local.tee $4
           (i32.and
            (i32.sub
@@ -2709,17 +2809,18 @@
       (local.set $5
        (i32.const 0)
       )
-      (br_if $block2
-       (i32.gt_u
-        (local.tee $4
+      (br_if $block1
+       (i32.lt_u
+        (local.get $1)
+        (local.tee $1
          (i32.sub
           (local.get $1)
-          (local.tee $9
+          (local.tee $4
            (i32.add
             (i32.load8_u
              (i32.add
               (i8x16.extract_lane_s 15
-               (local.get $13)
+               (local.get $14)
               )
               (i32.const 4112)
              )
@@ -2729,100 +2830,95 @@
           )
          )
         )
+       )
+      )
+      (br_if $block1
+       (i32.lt_u
         (local.get $1)
-       )
-      )
-      (br_if $block2
-       (i32.gt_u
         (local.get $3)
-        (local.get $4)
        )
       )
-      (local.set $1
-       (local.get $4)
-      )
-      (br_if $label5
+      (br_if $label7
        (i32.le_u
         (local.tee $0
          (i32.add
           (local.get $0)
-          (local.get $9)
+          (local.get $4)
          )
         )
-        (local.get $6)
+        (local.get $9)
        )
       )
      )
-    )
-   )
-   (local.set $6
-    (i32.sub
-     (local.get $4)
-     (local.get $3)
+     (local.set $8
+      (i32.sub
+       (local.get $1)
+       (local.get $3)
+      )
+     )
     )
    )
    (local.set $1
     (i32.const 0)
    )
-   (loop $label6
+   (loop $label9
     (local.set $4
      (i32.const 0)
     )
-    (loop $label7
-     (if
-      (i32.ne
-       (i32.load8_u
-        (i32.add
-         (local.get $2)
-         (local.get $4)
+    (block $block4
+     (loop $label8
+      (br_if $block4
+       (i32.ne
+        (i32.load8_u
+         (i32.add
+          (local.get $2)
+          (local.get $4)
+         )
         )
-       )
-       (i32.load8_u
-        (i32.add
-         (local.get $0)
-         (local.get $4)
+        (i32.load8_u
+         (i32.add
+          (local.get $0)
+          (local.get $4)
+         )
         )
        )
       )
-      (then
-       (local.set $0
-        (i32.add
-         (local.get $0)
-         (i32.const 1)
-        )
-       )
-       (local.set $5
-        (i32.const 0)
-       )
-       (br_if $label6
-        (i32.le_u
-         (local.tee $1
-          (i32.add
-           (local.get $1)
-           (i32.const 1)
-          )
+      (br_if $label8
+       (i32.ne
+        (local.get $3)
+        (local.tee $4
+         (i32.add
+          (local.get $4)
+          (i32.const 1)
          )
-         (local.get $6)
         )
        )
-       (br $block2)
       )
      )
-     (br_if $label7
-      (i32.ne
-       (local.get $3)
-       (local.tee $4
-        (i32.add
-         (local.get $4)
-         (i32.const 1)
-        )
-       )
-      )
+     (return
+      (local.get $0)
+     )
+    )
+    (local.set $0
+     (i32.add
+      (local.get $0)
+      (i32.const 1)
+     )
+    )
+    (local.set $5
+     (i32.const 0)
+    )
+    (br_if $label9
+     (i32.le_u
+      (local.tee $1
+       (i32.add
+        (local.get $1)
+        (i32.const 1)
+       )
+      )
+      (local.get $8)
      )
     )
-   )
-   (local.set $5
-    (local.get $0)
    )
   )
   (local.get $5)
diff --git a/sqlite3/libc/string.h b/sqlite3/libc/string.h
index 72a864e..c05543a 100644
--- a/sqlite3/libc/string.h
+++ b/sqlite3/libc/string.h
@@ -93,7 +93,7 @@ void *memchr(const void *v, int c, size_t n) {
 
   // memchr must behave as if it reads characters sequentially
   // and stops as soon as a match is found.
-  // Aligning ensures loads beyond the first match don't fail.
+  // Aligning ensures loads beyond the first match are safe.
   uintptr_t align = (uintptr_t)v % sizeof(v128_t);
   const v128_t *w = (v128_t *)((char *)v - align);
   const v128_t wc = wasm_i8x16_splat(c);
@@ -111,7 +111,8 @@ void *memchr(const void *v, int c, size_t n) {
       // If the mask is zero because of alignment,
       // it's as if we didn't find anything.
       if (mask) {
-        // We found a match, unless it is beyond the end of the object.
+        // Find the offset of the first one bit (little-endian).
+        // That's a match, unless it is beyond the end of the object.
         // Recall that we decremented n, so less-than-or-equal-to is correct.
         size_t ctz = __builtin_ctz(mask);
         return ctz <= n + align ? (char *)w + ctz : NULL;
@@ -136,6 +137,7 @@ void *memrchr(const void *v, int c, size_t n) {
     const v128_t cmp = wasm_i8x16_eq(wasm_v128_load(--w), wc);
     // Bitmask is slow on AArch64, any_true is much faster.
     if (wasm_v128_any_true(cmp)) {
+      // Find the offset of the last one bit (little-endian).
       size_t clz = __builtin_clz(wasm_i8x16_bitmask(cmp)) - 15;
       return (char *)(w + 1) - clz;
     }
@@ -152,7 +154,7 @@ void *memrchr(const void *v, int c, size_t n) {
 __attribute__((weak))
 size_t strlen(const char *s) {
   // strlen must stop as soon as it finds the terminator.
-  // Aligning ensures loads beyond the terminator don't fail.
+  // Aligning ensures loads beyond the terminator are safe.
   uintptr_t align = (uintptr_t)s % sizeof(v128_t);
   const v128_t *w = (v128_t *)(s - align);
 
@@ -167,6 +169,7 @@ size_t strlen(const char *s) {
       // Knowing this helps the compiler.
       __builtin_assume(mask || align);
       if (mask) {
+        // Find the offset of the first one bit (little-endian).
         return (char *)w - s + __builtin_ctz(mask);
       }
     }
@@ -191,7 +194,7 @@ static int __strcmp(const char *s1, const char *s2) {
       // if the terminator is found before that difference.
       break;
     }
-    // All characters are equal.
+    // We know all characters are equal.
     // If any is a terminator the strings are equal.
     if (!wasm_i8x16_all_true(wasm_v128_load(w1))) {
       return 0;
@@ -251,7 +254,7 @@ int strncmp(const char *s1, const char *s2, size_t n) {
       // if the terminator is found before that difference.
       break;
     }
-    // All characters are equal.
+    // We know all characters are equal.
     // If any is a terminator the strings are equal.
     if (!wasm_i8x16_all_true(wasm_v128_load(w1))) {
       return 0;
@@ -274,7 +277,7 @@ int strncmp(const char *s1, const char *s2, size_t n) {
 
 static char *__strchrnul(const char *s, int c) {
   // strchrnul must stop as soon as a match is found.
-  // Aligning ensures loads beyond the first match don't fail.
+  // Aligning ensures loads beyond the first match are safe.
   uintptr_t align = (uintptr_t)s % sizeof(v128_t);
   const v128_t *w = (v128_t *)(s - align);
   const v128_t wc = wasm_i8x16_splat(c);
@@ -290,6 +293,7 @@ static char *__strchrnul(const char *s, int c) {
       // Knowing this helps the compiler.
       __builtin_assume(mask || align);
       if (mask) {
+        // Find the offset of the first one bit (little-endian).
         return (char *)w + __builtin_ctz(mask);
       }
     }
@@ -388,6 +392,7 @@ size_t strspn(const char *s, const char *c) {
       const v128_t cmp = wasm_i8x16_eq(wasm_v128_load(w), wc);
       // Bitmask is slow on AArch64, all_true is much faster.
       if (!wasm_i8x16_all_true(cmp)) {
+        // Find the offset of the first zero bit (little-endian).
         size_t ctz = __builtin_ctz(~wasm_i8x16_bitmask(cmp));
         return (char *)w + ctz - s;
       }
@@ -410,6 +415,7 @@ size_t strspn(const char *s, const char *c) {
     const v128_t cmp = _WASM_SIMD128_CHKBITS(bitmap, wasm_v128_load(w));
     // Bitmask is slow on AArch64, all_true is much faster.
     if (!wasm_i8x16_all_true(cmp)) {
+      // Find the offset of the first zero bit (little-endian).
       size_t ctz = __builtin_ctz(~wasm_i8x16_bitmask(cmp));
       return (char *)w + ctz - s;
     }
@@ -442,6 +448,7 @@ size_t strcspn(const char *s, const char *c) {
     const v128_t cmp = _WASM_SIMD128_CHKBITS(bitmap, wasm_v128_load(w));
     // Bitmask is slow on AArch64, any_true is much faster.
     if (wasm_v128_any_true(cmp)) {
+      // Find the offset of the first one bit (little-endian).
       size_t ctz = __builtin_ctz(wasm_i8x16_bitmask(cmp));
       return (char *)w + ctz - s;
     }
@@ -462,14 +469,18 @@ static const char *__memmem_rabin(const char *haystk, size_t sh,
                                   const char *needle, size_t sn,
                                   uint8_t bmbc[256]) {
   // http://0x80.pl/notesen/2016-11-28-simd-strfind.html
+
+  // We've handled empty and single character needles.
+  // The needle is no longer than haystack.
   __builtin_assume(2 <= sn && sn <= sh);
 
   const v128_t fst = wasm_i8x16_splat(needle[0]);
   const v128_t lst = wasm_i8x16_splat(needle[sn - 1]);
-  const char *N =
-      (char *)(__builtin_wasm_memory_size(0) * PAGESIZE - sn - sizeof(v128_t));
+  // The last haystk offset for which loading blk_lst is safe.
+  const char *H = (char *)(__builtin_wasm_memory_size(0) * PAGESIZE - sn + 1 -
+                           sizeof(v128_t));
 
-  while (haystk <= N) {
+  while (haystk <= H) {
     const v128_t blk_fst = wasm_v128_load((v128_t *)(haystk));
     const v128_t blk_lst = wasm_v128_load((v128_t *)(haystk + sn - 1));
     const v128_t eq_fst = wasm_i8x16_eq(fst, blk_fst);
@@ -477,17 +488,26 @@ static const char *__memmem_rabin(const char *haystk, size_t sh,
 
     const v128_t cmp = eq_fst & eq_lst;
     if (wasm_v128_any_true(cmp)) {
+      // Find the offset of the first one bit (little-endian).
+      // 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 - 2)) {
+        // This could compare one less byte.
+        // Since bcmp compares left-to-right (this could change)
+        // the last byte only matters if we just found the needle.
+        // Otherwise this may help bcmp vectorize.
+        if (!bcmp(haystk + ctz + 1, needle + 1, sn - 1)) {
           return haystk + ctz;
         }
       }
     }
 
     size_t skip = sizeof(v128_t);
+    // Apply the bad-character rule to the last byte of the haystack.
     if (bmbc) skip += bmbc[wasm_i8x16_extract_lane(blk_lst, 15)];
+    // Have we reached the end of the haystack?
     if (__builtin_sub_overflow(sh, skip, &sh)) return NULL;
+    // Is the needle longer than the haystack?
     if (sn > sh) return NULL;
     haystk += skip;
   }
@@ -506,16 +526,31 @@ static const char *__memmem_rabin(const char *haystk, size_t sh,
 static const char *__memmem_raita(const char *haystk, size_t sh,
                                   const char *needle, size_t sn) {
   // https://www-igm.univ-mlv.fr/~lecroq/string/node22.html
+
+  // We've handled empty and single character needles.
+  // The needle is no longer than haystack.
   __builtin_assume(2 <= sn && sn <= sh);
 
+  // Compute Boyer-Moore's bad-character shift function.
+  // Only the last 255 characters of the needle matter for shifts up to 255,
+  // which is good enough for most needles.
+  size_t n = sn - 1;
+  size_t i = 0;
+  int c = n;
+  if (c >= 255) {
+    c = 255;
+    i = n - 255;
+  }
+
 #ifndef _REENTRANT
   static
 #endif
   uint8_t bmbc[256];
-  memset(bmbc, sn - 1 < 255 ? sn - 1 : 255, sizeof(bmbc));
-  for (size_t i = 0; i < sn - 1; i++) {
-    size_t t = sn - 1 - i - 1;
-    if (t > 255) t = 255;
+  memset(bmbc, c, sizeof(bmbc));
+  for (; i < n; i++) {
+    // One less than the usual offset.
+    // Added back later (as vector).
+    size_t t = n - i - 1;
     bmbc[(unsigned char)needle[i]] = t;
   }
 
@@ -527,8 +562,10 @@ static const char *__memmem(const char *haystk, size_t sh,  //
   // Return when needle is longer than haystack.
   if (sn > sh) return NULL;
 
-  return sn < sizeof(v128_t) ? __memmem_rabin(haystk, sh, needle, sn, NULL)
-                             : __memmem_raita(haystk, sh, needle, sn);
+  // Decide if Boyer-Moore's bad-character rule will be useful.
+  return sn < sizeof(v128_t) || sh - sn < sizeof(v128_t)
+             ? __memmem_rabin(haystk, sh, needle, sn, NULL)
+             : __memmem_raita(haystk, sh, needle, sn);
 }
 
 __attribute__((weak))