From f5206ea8daa394ac566d1fdbfa8e068a98af032c Mon Sep 17 00:00:00 2001
From: Nuno Cruces <ncruces@users.noreply.github.com>
Date: Fri, 18 Apr 2025 03:01:09 +0100
Subject: [PATCH] Shellsort.

---
 sqlite3/libc/build.sh  |   4 +-
 sqlite3/libc/libc.wasm | Bin 2199 -> 2767 bytes
 sqlite3/libc/libc.wat  | 458 ++++++++++++++++++++++++++++++++++++++++-
 sqlite3/libc/stdlib.h  |  50 +++++
 4 files changed, 509 insertions(+), 3 deletions(-)
 create mode 100644 sqlite3/libc/stdlib.h

diff --git a/sqlite3/libc/build.sh b/sqlite3/libc/build.sh
index 855a393..cff8387 100755
--- a/sqlite3/libc/build.sh
+++ b/sqlite3/libc/build.sh
@@ -11,6 +11,7 @@ SRCS="${1:-libc.c}"
 
 trap 'rm -f libc.c libc.tmp' EXIT
 echo '#include <string.h>' > libc.c
+echo '#include <stdlib.h>' >> libc.c
 
 "$WASI_SDK/clang" --target=wasm32-wasi -std=c23 -g0 -O2 \
 	-o libc.wasm -I. "$SRCS" \
@@ -32,7 +33,8 @@ echo '#include <string.h>' > libc.c
 	-Wl,--export=strcspn \
 	-Wl,--export=strlen \
 	-Wl,--export=strncmp \
-	-Wl,--export=strspn
+	-Wl,--export=strspn \
+	-Wl,--export=qsort
 
 "$BINARYEN/wasm-ctor-eval" -g -c _initialize libc.wasm -o libc.tmp
 "$BINARYEN/wasm-opt" -g --strip --strip-producers -c -O3 \
diff --git a/sqlite3/libc/libc.wasm b/sqlite3/libc/libc.wasm
index 995c28efdf41a68a282b7cfd0a1b088f7c183730..4358642cd04b0811aab12e95c397af6fdd6fee9a 100755
GIT binary patch
delta 687
zcmY*X%Wl*#6uq{ulSv#mnQ7U;=GuE$v0_6!RV$64;tP;kP#0vT&`=~$c{oe9RTuCF
zY+wOD!dLJamAKA~K$C^fJ<hqu$B!@5-?#SR<l;;KfVhpvGDZ=PRUAR28o7?3Y$f=U
zkc6O6i*u$L1q?F3v{(20TK5;PUcUju!g82ZKfnLU;#ZVS6z^qA^1uR#AQuTpuG|jE
zQ1`B8IPYCSK(ccUvu5F@s!iHMmk|U(;>TQnnv0KfGko)cvOC@e(XmNKaO7qr$*z3q
zjB0ryL*>$IGLI~%*s`tmRp*hA%8uvN6d=H3D<q3qKIJ^K1JYDc(=B>tnu^Y8)-&W(
zxQxT#4FI@3o*Zo^iQX6hn$X^e7@E!@WuaO6%z~ksv&%@up_P>|glr{UW>_JXJgD9y
zwp@JZ#|84c?7vGnX)YeHv8LS-c%vL9PmmLGvq!c!LGeLyu(1!f`Y`pS29GhhPB^n3
z>BbsqvX)4e{8wyP`t?ttwiFw-rNER+2L{c%!S(J9pJuv+hg2P<w0uLR^M-afv>_Yh
tlQe5gAs`F`8X@2vzC1ZQg|b+^c=r6>+w>$)tLtC+)7-2d)^$Dk{1<;CWDfuU

delta 114
zcmX>vI$h9&A+b1@k%57MQJ6V_xxT)hu|9zbOfmwQ%skvcDF!A+21X!YWnpA!U}7)g
zndn@|&cn#a^^BK$<Kd@lirk!x42}&AEZmHY423LtiMgo^T%0+XNy&QUAk*WxCqL$l
Jo$Sul0RTP%8IS+~

diff --git a/sqlite3/libc/libc.wat b/sqlite3/libc/libc.wat
index 738ab4c..9373667 100644
--- a/sqlite3/libc/libc.wat
+++ b/sqlite3/libc/libc.wat
@@ -1,9 +1,11 @@
 (module $libc.wasm
- (type $0 (func (param i32 i32 i32) (result i32)))
- (type $1 (func (param i32 i32) (result i32)))
+ (type $0 (func (param i32 i32) (result i32)))
+ (type $1 (func (param i32 i32 i32) (result i32)))
  (type $2 (func (param i32) (result i32)))
+ (type $3 (func (param i32 i32 i32 i32)))
  (memory $0 256)
  (data $0 (i32.const 65536) "\01")
+ (table $0 1 1 funcref)
  (export "memory" (memory $0))
  (export "memset" (func $memset))
  (export "memcpy" (func $memcpy))
@@ -16,6 +18,7 @@
  (export "strchr" (func $strchr))
  (export "strspn" (func $strspn))
  (export "strcspn" (func $strcspn))
+ (export "qsort" (func $qsort))
  (func $memset (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
   (memory.fill
    (local.get $0)
@@ -1315,6 +1318,457 @@
    (local.get $0)
   )
  )
+ (func $qsort (param $0 i32) (param $1 i32) (param $2 i32) (param $3 i32)
+  (local $4 i32)
+  (local $5 i32)
+  (local $6 i32)
+  (local $7 i32)
+  (local $8 i32)
+  (local $9 i32)
+  (local $10 i32)
+  (local $11 i32)
+  (local $12 i32)
+  (local $13 i32)
+  (local $14 i32)
+  (local $15 i32)
+  (local $16 i32)
+  (local $17 i32)
+  (local $18 i32)
+  (local $19 i32)
+  (local $20 v128)
+  (local $scratch i32)
+  (block $block
+   (br_if $block
+    (i32.eqz
+     (local.get $2)
+    )
+   )
+   (br_if $block
+    (i32.lt_u
+     (local.get $1)
+     (i32.const 2)
+    )
+   )
+   (local.set $14
+    (i32.mul
+     (local.get $1)
+     (local.get $2)
+    )
+   )
+   (local.set $15
+    (i32.and
+     (local.get $2)
+     (i32.const 15)
+    )
+   )
+   (local.set $9
+    (i32.and
+     (local.get $2)
+     (i32.const -16)
+    )
+   )
+   (local.set $16
+    (i32.add
+     (local.get $0)
+     (local.get $2)
+    )
+   )
+   (local.set $17
+    (i32.lt_u
+     (local.get $2)
+     (i32.const 16)
+    )
+   )
+   (loop $label5
+    (local.set $6
+     (i32.eq
+      (local.get $1)
+      (i32.const 2)
+     )
+    )
+    (local.set $18
+     (i32.le_u
+      (i32.add
+       (local.get $0)
+       (i32.mul
+        (i32.add
+         (local.tee $13
+          (select
+           (i32.const 1)
+           (local.tee $1
+            (i32.wrap_i64
+             (i64.div_u
+              (i64.sub
+               (i64.mul
+                (i64.extend_i32_u
+                 (local.get $1)
+                )
+                (i64.const 5)
+               )
+               (i64.const 1)
+              )
+              (i64.const 11)
+             )
+            )
+           )
+           (local.get $6)
+          )
+         )
+         (i32.const 1)
+        )
+        (local.get $2)
+       )
+      )
+      (local.get $0)
+     )
+    )
+    (local.set $11
+     (local.tee $10
+      (i32.mul
+       (local.get $2)
+       (local.get $13)
+      )
+     )
+    )
+    (loop $label4
+     (block $block1
+      (br_if $block1
+       (i32.gt_u
+        (local.tee $5
+         (i32.sub
+          (local.get $11)
+          (local.get $10)
+         )
+        )
+        (local.get $11)
+       )
+      )
+      (loop $label3
+       (br_if $block1
+        (i32.le_s
+         (call_indirect $0 (type $0)
+          (local.tee $4
+           (i32.add
+            (local.get $0)
+            (local.tee $12
+             (local.get $5)
+            )
+           )
+          )
+          (local.tee $5
+           (i32.add
+            (local.get $4)
+            (local.get $10)
+           )
+          )
+          (local.get $3)
+         )
+         (i32.const 0)
+        )
+       )
+       (block $block3
+        (block $block4
+         (block $block2
+          (br_if $block2
+           (local.get $17)
+          )
+          (br_if $block2
+           (i32.and
+            (i32.eqz
+             (local.get $18)
+            )
+            (i32.lt_u
+             (local.get $5)
+             (i32.add
+              (local.get $12)
+              (local.get $16)
+             )
+            )
+           )
+          )
+          (local.set $5
+           (i32.add
+            (local.get $5)
+            (local.get $9)
+           )
+          )
+          (local.set $7
+           (i32.add
+            (local.get $4)
+            (local.get $9)
+           )
+          )
+          (local.set $6
+           (local.get $9)
+          )
+          (loop $label
+           (local.set $20
+            (v128.load align=1
+             (local.get $4)
+            )
+           )
+           (v128.store align=1
+            (local.get $4)
+            (v128.load align=1
+             (local.tee $8
+              (i32.add
+               (local.get $4)
+               (local.get $10)
+              )
+             )
+            )
+           )
+           (v128.store align=1
+            (local.get $8)
+            (local.get $20)
+           )
+           (local.set $4
+            (i32.add
+             (local.get $4)
+             (i32.const 16)
+            )
+           )
+           (br_if $label
+            (local.tee $6
+             (i32.sub
+              (local.get $6)
+              (i32.const 16)
+             )
+            )
+           )
+          )
+          (local.set $6
+           (local.get $15)
+          )
+          (br_if $block3
+           (i32.eq
+            (local.get $2)
+            (local.get $9)
+           )
+          )
+          (br $block4)
+         )
+         (local.set $7
+          (local.get $4)
+         )
+         (local.set $6
+          (local.get $2)
+         )
+        )
+        (br_if $block3
+         (i32.lt_u
+          (block (result i32)
+           (local.set $scratch
+            (i32.sub
+             (local.get $6)
+             (i32.const 1)
+            )
+           )
+           (if
+            (local.tee $4
+             (i32.and
+              (local.get $6)
+              (i32.const 3)
+             )
+            )
+            (then
+             (local.set $6
+              (i32.and
+               (local.get $6)
+               (i32.const -4)
+              )
+             )
+             (loop $label1
+              (local.set $19
+               (i32.load8_u
+                (local.get $7)
+               )
+              )
+              (i32.store8
+               (local.get $7)
+               (i32.load8_u
+                (local.get $5)
+               )
+              )
+              (i32.store8
+               (local.get $5)
+               (local.get $19)
+              )
+              (local.set $5
+               (i32.add
+                (local.get $5)
+                (i32.const 1)
+               )
+              )
+              (local.set $7
+               (i32.add
+                (local.get $7)
+                (i32.const 1)
+               )
+              )
+              (br_if $label1
+               (local.tee $4
+                (i32.sub
+                 (local.get $4)
+                 (i32.const 1)
+                )
+               )
+              )
+             )
+            )
+           )
+           (local.get $scratch)
+          )
+          (i32.const 3)
+         )
+        )
+        (loop $label2
+         (local.set $4
+          (i32.load8_u
+           (local.get $7)
+          )
+         )
+         (i32.store8
+          (local.get $7)
+          (i32.load8_u
+           (local.get $5)
+          )
+         )
+         (i32.store8
+          (local.get $5)
+          (local.get $4)
+         )
+         (local.set $8
+          (i32.load8_u
+           (local.tee $4
+            (i32.add
+             (local.get $7)
+             (i32.const 1)
+            )
+           )
+          )
+         )
+         (i32.store8
+          (local.get $4)
+          (i32.load8_u
+           (local.tee $4
+            (i32.add
+             (local.get $5)
+             (i32.const 1)
+            )
+           )
+          )
+         )
+         (i32.store8
+          (local.get $4)
+          (local.get $8)
+         )
+         (local.set $8
+          (i32.load8_u
+           (local.tee $4
+            (i32.add
+             (local.get $7)
+             (i32.const 2)
+            )
+           )
+          )
+         )
+         (i32.store8
+          (local.get $4)
+          (i32.load8_u
+           (local.tee $4
+            (i32.add
+             (local.get $5)
+             (i32.const 2)
+            )
+           )
+          )
+         )
+         (i32.store8
+          (local.get $4)
+          (local.get $8)
+         )
+         (local.set $8
+          (i32.load8_u
+           (local.tee $4
+            (i32.add
+             (local.get $7)
+             (i32.const 3)
+            )
+           )
+          )
+         )
+         (i32.store8
+          (local.get $4)
+          (i32.load8_u
+           (local.tee $4
+            (i32.add
+             (local.get $5)
+             (i32.const 3)
+            )
+           )
+          )
+         )
+         (i32.store8
+          (local.get $4)
+          (local.get $8)
+         )
+         (local.set $7
+          (i32.add
+           (local.get $7)
+           (i32.const 4)
+          )
+         )
+         (local.set $5
+          (i32.add
+           (local.get $5)
+           (i32.const 4)
+          )
+         )
+         (br_if $label2
+          (local.tee $6
+           (i32.sub
+            (local.get $6)
+            (i32.const 4)
+           )
+          )
+         )
+        )
+       )
+       (br_if $label3
+        (i32.le_u
+         (local.tee $5
+          (i32.sub
+           (local.get $12)
+           (local.get $10)
+          )
+         )
+         (local.get $12)
+        )
+       )
+      )
+     )
+     (br_if $label4
+      (i32.lt_u
+       (local.tee $11
+        (i32.add
+         (local.get $2)
+         (local.get $11)
+        )
+       )
+       (local.get $14)
+      )
+     )
+    )
+    (br_if $label5
+     (i32.ge_u
+      (local.get $13)
+      (i32.const 2)
+     )
+    )
+   )
+  )
+ )
  ;; features section: mutable-globals, nontrapping-float-to-int, simd, bulk-memory, sign-ext, reference-types, multivalue, bulk-memory-opt
 )
 
diff --git a/sqlite3/libc/stdlib.h b/sqlite3/libc/stdlib.h
new file mode 100644
index 0000000..ba91da6
--- /dev/null
+++ b/sqlite3/libc/stdlib.h
@@ -0,0 +1,50 @@
+#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
+extern "C" {
+#endif
+
+// Shellsort with Gonnet & Baeza-Yates gap sequence.
+// Simple, no recursion, doesn't use the C stack.
+// Clang auto-vectorizes the inner loop. 
+
+void qsort(void *base, size_t nel, size_t width,
+           int (*comp)(const void *, const void *)) {
+  if (width == 0) return;
+
+  size_t wnel = width * nel;
+  size_t gap = nel;
+  while (gap > 1) {
+    gap = (5ull * gap - 1) / 11;
+    if (gap == 0) gap = 1;
+
+    size_t wgap = width * gap;
+    __builtin_assume(wgap < wnel);
+    for (size_t i = wgap; i < wnel; i += width) {
+      for (size_t j = i; !__builtin_sub_overflow(j, wgap, &j);) {
+        char *a = j + (char *)base;
+        char *b = a + wgap;
+        if (comp(a, b) <= 0) break;
+
+        size_t s = width;
+        do {
+          char tmp = *a;
+          *a++ = *b;
+          *b++ = tmp;
+        } while (--s);
+      }
+    }
+  }
+}
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
+
+#endif  // _WASM_SIMD128_STDLIB_H
\ No newline at end of file