From 3bd11a0a86349e58c5de2dfacf0cebaebc1d6328 Mon Sep 17 00:00:00 2001
From: Nuno Cruces <ncruces@users.noreply.github.com>
Date: Sat, 5 Apr 2025 11:15:16 +0100
Subject: [PATCH] More SIMD.

---
 sqlite3/libc/build.sh     |   1 +
 sqlite3/libc/libc.wasm    | Bin 858 -> 905 bytes
 sqlite3/libc/libc.wat     | 240 ++++++++++++++++++++++----------------
 sqlite3/libc/libc_test.go |  18 +++
 sqlite3/strings.c         |  58 +++++----
 5 files changed, 195 insertions(+), 122 deletions(-)

diff --git a/sqlite3/libc/build.sh b/sqlite3/libc/build.sh
index 3211d01..9dcd5be 100755
--- a/sqlite3/libc/build.sh
+++ b/sqlite3/libc/build.sh
@@ -20,6 +20,7 @@ trap 'rm -f libc.tmp' EXIT
 	-Wl,--export=memset \
 	-Wl,--export=memcpy \
 	-Wl,--export=memcmp \
+	-Wl,--export=strlen \
 	-Wl,--export=strcmp \
 	-Wl,--export=strncmp
 
diff --git a/sqlite3/libc/libc.wasm b/sqlite3/libc/libc.wasm
index ce4e30611e65b1cc26f4be82942621c92b7daade..1ed0412ff11723022580d149382a48375ab5b312 100755
GIT binary patch
delta 360
zcmYk2u}T9$6h+^AGrNn~EG|e&@ntG4jNlKrQ$!F+A%eDvOAupr5jS9AF}8MwZTthV
zva<3I`~?5OPgwBHBF5%%VeZ^>xnJR9v<H)_KmfoEHCE0!bYt{~yi`%>p9lht@IaA`
zmOI;9^JxhxtQNO<aRWLMQF;}05j!6`TwXrAGbBeDgTVu^>B6ShhK>_Z8Dk@fO+HMJ
z(WWzlOrr$in8CV=an|=HuwF3kf*h|=EDZ4*5;?r5Y?aDzl!%G{$MZROiqXwUb1EP$
zJe^rsk6UIR1t4hfpQOfL1Hl<;sYV^@RgJcR)S~(w2gIuW#+Tkumy==&owfPZMLxLm
UGwz4=D{XdVAN@N|P=C_X4<Z#s@&Et;

delta 313
zcmYk2El&eM5I|>UXYayYp(jcm$t(&;C=yJQu4p~s542nm5U!yu0fAU7j^+#=!QVln
z(RjoKcFusAx0$^8+Nbm**@xMxLIA)mnwk&-hAC+@iPbDvMtG$3Bv~Kr^~>r8OuD*V
z&h^C|#O#&r`tBusAyJ?Q3Ic*l-(B!X;1Ix(U_=LV!3G$1tw2j|PMfnRano1HMnuVI
zq8#(cLpN{D!f{(!I4kIjD~AFK5F6FlW4sBu9RF*RJ0s+Oa3MzHL^tuM6{tWlj^dnF
osXuX${ke6I-nNH4pH&x-ZCqcTmxFu#p95RJ(+<a@x&Ef-U-LIafB*mh

diff --git a/sqlite3/libc/libc.wat b/sqlite3/libc/libc.wat
index 42e54ad..bd223c9 100644
--- a/sqlite3/libc/libc.wat
+++ b/sqlite3/libc/libc.wat
@@ -1,12 +1,14 @@
 (module $libc.wasm
  (type $0 (func (param i32 i32 i32) (result i32)))
- (type $1 (func (param i32 i32) (result i32)))
+ (type $1 (func (param i32) (result i32)))
+ (type $2 (func (param i32 i32) (result i32)))
  (memory $0 256)
  (data $0 (i32.const 1024) "\01")
  (export "memory" (memory $0))
  (export "memset" (func $memset))
  (export "memcpy" (func $memcpy))
  (export "memcmp" (func $memcmp))
+ (export "strlen" (func $strlen))
  (export "strcmp" (func $strcmp))
  (export "strncmp" (func $strncmp))
  (func $memset (param $0 i32) (param $1 i32) (param $2 i32) (result i32)
@@ -128,127 +130,173 @@
   )
   (i32.const 0)
  )
- (func $strcmp (param $0 i32) (param $1 i32) (result i32)
+ (func $strlen (param $0 i32) (result i32)
+  (local $1 i32)
   (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)
+  (local $scratch i32)
+  (block $block
+   (br_if $block
+    (i32.gt_u
+     (local.tee $1
+      (local.get $0)
      )
-     (then
-      (local.set $2
-       (i32.load8_u
-        (local.get $0)
+     (local.tee $2
+      (i32.sub
+       (i32.shl
+        (memory.size)
+        (i32.const 16)
        )
+       (i32.const 16)
       )
-      (br $block
-       (i32.load8_u
+     )
+    )
+   )
+   (loop $label
+    (br_if $block
+     (i32.eqz
+      (i8x16.all_true
+       (v128.load align=1
         (local.get $1)
        )
       )
      )
     )
-    (if
-     (v128.any_true
-      (v128.xor
-       (local.tee $5
-        (v128.load
-         (local.get $1)
-        )
-       )
-       (local.tee $4
-        (v128.load
-         (local.get $0)
-        )
-       )
-      )
-     )
-     (then
-      (local.set $2
-       (i8x16.extract_lane_u 0
-        (local.get $4)
-       )
-      )
-      (br $block
-       (i8x16.extract_lane_u 0
-        (local.get $5)
+    (br_if $label
+     (i32.le_u
+      (local.tee $1
+       (i32.add
+        (local.get $1)
+        (i32.const 16)
        )
       )
+      (local.get $2)
      )
     )
-    (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.set $0
+   (i32.add
+    (i32.xor
+     (local.get $0)
+     (i32.const -1)
+    )
+    (local.get $1)
+   )
+  )
+  (loop $label1
+   (local.set $0
+    (i32.add
+     (local.get $0)
+     (i32.const 1)
+    )
+   )
+   (br_if $label1
+    (block (result i32)
+     (local.set $scratch
+      (i32.load8_u
        (local.get $1)
       )
      )
      (local.set $1
       (i32.add
        (local.get $1)
+       (i32.const 1)
+      )
+     )
+     (local.get $scratch)
+    )
+   )
+  )
+  (local.get $0)
+ )
+ (func $strcmp (param $0 i32) (param $1 i32) (result i32)
+  (local $2 i32)
+  (local $3 i32)
+  (local $4 v128)
+  (block $block
+   (br_if $block
+    (i32.lt_u
+     (local.tee $2
+      (i32.sub
+       (i32.shl
+        (memory.size)
+        (i32.const 16)
+       )
        (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.get $0)
+    )
+   )
+   (br_if $block
+    (i32.gt_u
+     (local.get $1)
+     (local.get $2)
+    )
+   )
+   (loop $label
+    (br_if $block
+     (v128.any_true
+      (v128.xor
+       (v128.load align=1
+        (local.get $1)
+       )
+       (local.tee $4
+        (v128.load align=1
+         (local.get $0)
         )
        )
       )
      )
     )
-    (local.set $2
-     (i8x16.extract_lane_u 0
-      (local.get $4)
+    (if
+     (i32.eqz
+      (i8x16.all_true
+       (local.get $4)
+      )
+     )
+     (then
+      (return
+       (i32.const 0)
+      )
      )
     )
-    (i8x16.extract_lane_u 0
-     (local.get $5)
+    (local.set $1
+     (i32.add
+      (local.get $1)
+      (i32.const 16)
+     )
+    )
+    (br_if $block
+     (i32.gt_u
+      (local.tee $0
+       (i32.add
+        (local.get $0)
+        (i32.const 16)
+       )
+      )
+      (local.get $2)
+     )
+    )
+    (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
@@ -264,16 +312,10 @@
       (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
@@ -313,14 +355,8 @@
    )
   )
   (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)
diff --git a/sqlite3/libc/libc_test.go b/sqlite3/libc/libc_test.go
index 7ee710e..701dd6b 100644
--- a/sqlite3/libc/libc_test.go
+++ b/sqlite3/libc/libc_test.go
@@ -25,6 +25,7 @@ var (
 	memset  api.Function
 	memcpy  api.Function
 	memcmp  api.Function
+	strlen  api.Function
 	strcmp  api.Function
 	strncmp api.Function
 	stack   [8]uint64
@@ -49,6 +50,7 @@ func TestMain(m *testing.M) {
 	memset = mod.ExportedFunction("memset")
 	memcpy = mod.ExportedFunction("memcpy")
 	memcmp = mod.ExportedFunction("memcmp")
+	strlen = mod.ExportedFunction("strlen")
 	strcmp = mod.ExportedFunction("strcmp")
 	strncmp = mod.ExportedFunction("strncmp")
 	memory, _ = mod.Memory().Read(0, mod.Memory().Size())
@@ -114,6 +116,22 @@ func Benchmark_memcmp(b *testing.B) {
 	}
 }
 
+func Benchmark_strlen(b *testing.B) {
+	clear(memory)
+	call(memset, ptr1, 5, size-1)
+
+	b.SetBytes(size)
+	b.ResetTimer()
+	for range b.N {
+		call(strlen, ptr1)
+	}
+	b.StopTimer()
+
+	if got := int32(call(strlen, ptr1)); got != size-1 {
+		b.Fatal(got)
+	}
+}
+
 func Benchmark_strcmp(b *testing.B) {
 	clear(memory)
 	call(memset, ptr1, 7, size-1)
diff --git a/sqlite3/strings.c b/sqlite3/strings.c
index 89364cf..74aca2b 100644
--- a/sqlite3/strings.c
+++ b/sqlite3/strings.c
@@ -1,3 +1,4 @@
+#include <__macro_PAGESIZE.h>
 #include <limits.h>
 #include <stddef.h>
 #include <stdint.h>
@@ -21,8 +22,6 @@ void *memmove(void *dest, const void *src, size_t n) {
 
 #ifdef __wasm_simd128__
 
-#define UNALIGNED(x) ((uintptr_t)x % sizeof(*x))
-
 int memcmp(const void *v1, const void *v2, size_t n) {
   const v128_t *w1 = v1;
   const v128_t *w2 = v2;
@@ -44,20 +43,41 @@ int memcmp(const void *v1, const void *v2, size_t n) {
   return 0;
 }
 
-int strcmp(const char *c1, const char *c2) {
-  const v128_t *w1 = (void *)c1;
-  const v128_t *w2 = (void *)c2;
-  if (!(UNALIGNED(w1) | UNALIGNED(w2))) {
-    while (true) {
-      if (wasm_v128_any_true(*w1 ^ *w2)) {
-        break;  // *w1 != *w2
-      }
-      if (!wasm_i8x16_all_true(*w1)) {
-        return 0;  // *w1 == *w2 and have a NUL
-      }
-      w1++;
-      w2++;
+size_t strlen(const char *s) {
+  const v128_t *const limit =
+      (v128_t *)(__builtin_wasm_memory_size(0) * PAGESIZE) - 1;
+
+  const v128_t *w = (void *)s;
+  while (w <= limit) {
+    if (!wasm_i8x16_all_true(wasm_v128_load(w))) {
+      break;  // *w has a NUL
     }
+    w++;
+  }
+
+  const char *ss = (void *)w;
+  while (true) {
+    if (*ss == 0) break;
+    ss++;
+  }
+  return ss - 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;
+  while (w1 <= limit && w2 <= limit) {
+    if (wasm_v128_any_true(wasm_v128_load(w1) ^ wasm_v128_load(w2))) {
+      break;  // *w1 != *w2
+    }
+    if (!wasm_i8x16_all_true(wasm_v128_load(w1))) {
+      return 0;  // *w1 == *w2 and have a NUL
+    }
+    w1++;
+    w2++;
   }
 
   const uint8_t *u1 = (void *)w1;
@@ -71,9 +91,9 @@ int strcmp(const char *c1, const char *c2) {
   return 0;
 }
 
-int strncmp(const char *c1, const char *c2, size_t n) {
-  const v128_t *w1 = (void *)c1;
-  const v128_t *w2 = (void *)c2;
+int strncmp(const char *s1, const char *s2, size_t n) {
+  const v128_t *w1 = (void *)s1;
+  const v128_t *w2 = (void *)s2;
   for (; n >= sizeof(v128_t); n -= sizeof(v128_t)) {
     if (wasm_v128_any_true(wasm_v128_load(w1) ^ wasm_v128_load(w2))) {
       break;  // *w1 != *w2
@@ -96,6 +116,4 @@ int strncmp(const char *c1, const char *c2, size_t n) {
   return 0;
 }
 
-#undef UNALIGNED
-
 #endif