From 276d79b1fa139756477e4acdb6b5fac51f2acb90 Mon Sep 17 00:00:00 2001
From: Fabrice <fabrice@schaub-dev.xyz>
Date: Mon, 23 Mar 2026 08:12:21 +0100
Subject: adding sse2 path

---
 cheesemap.c | 46 ++++++++++++++++++++++++++++++++++++++++++++--
 cheesemap.h | 21 +++++++++------------
 makefile    |  5 +++++
 3 files changed, 58 insertions(+), 14 deletions(-)

diff --git a/cheesemap.c b/cheesemap.c
index ade0659..b0a7b62 100644
--- a/cheesemap.c
+++ b/cheesemap.c
@@ -6,6 +6,11 @@
 #include <stdlib.h>
 #include <string.h>
 
+#include CM_OPT_ASSERT_PATH
+#ifndef assert
+#error "assert is not defined"
+#endif
+
 #define CM_ATTR(...) __attribute__((__VA_ARGS__))
 
 CM_ATTR(hot) static inline uintptr_t cm_ctz(uintptr_t val) {
@@ -34,7 +39,7 @@ CM_ATTR(hot) static inline uintptr_t cm_bitmask_lowest_set_bit(bitmask_t mask) {
 #if CM_GROUP_SIZE == 8
   return cm_ctz(mask) / CHAR_BIT;
 #elif CM_GROUP_SIZE == 16
-  return cm_ctz(mask)
+  return cm_ctz(mask);
 #else
 #error "unknown group size"
 #endif
@@ -105,7 +110,33 @@ static inline group_t cm_group_load(const uint8_t* ctrl);
 static inline bitmask_t cm_group_match_empty_or_deleted(group_t group);
 static inline bitmask_t cm_group_match_full(group_t group);
 
+/* sse2 implementation */
+#ifdef CM_SSE2
+static inline group_t cm_group_load(const uint8_t* ctrl) {
+  assert(ctrl != NULL);
+  return _mm_loadu_si128((const group_t*)ctrl);
+}
+
+static inline bitmask_t cm_group_match_tag(group_t group, uint8_t tag) {
+  __m128i cmp = _mm_cmpeq_epi8(group, _mm_set1_epi8(tag));
+  return _mm_movemask_epi8(cmp);
+}
+
+static inline bitmask_t cm_group_match_empty_or_deleted(group_t group) {
+  return _mm_movemask_epi8(group);
+}
+
+static inline bitmask_t cm_group_match_empty(group_t group) {
+  return cm_group_match_tag(group, CM_CTRL_EMPTY);
+}
+
+static inline bitmask_t cm_group_match_full(group_t group) {
+  return ~cm_group_match_empty_or_deleted(group);
+}
+#endif
+
 /* scalar implementation */
+#ifndef CM_NO_FALLBACK
 static inline group_t cm_group_repeat(uint8_t v) {
   return (group_t)v * (((group_t)-1) / (uint8_t)~0);
 }
@@ -136,8 +167,19 @@ static inline bitmask_t cm_group_match_tag(group_t group, uint8_t tag) {
   return (cmp - cm_group_repeat(CM_CTRL_END)) & ~cmp &
          cm_group_repeat(CM_CTRL_DELETED);
 }
+#endif
+
+/* ctrl's n stuff */
+
+static inline uintptr_t cm_h1(cm_hash_t hash) {
+  return (uintptr_t)(hash >> CM_FP_SIZE);
+}
+
+static inline uint8_t cm_h2(cm_hash_t hash) {
+  uintptr_t top = hash >> (sizeof(cm_hash_t) * CHAR_BIT - CM_FP_SIZE);
+  return (uint8_t)(top & CM_H2_MASK);
+}
 
-/* static ctrl's */
 const uint8_t CM_CTRL_STATIC_EMPTY[CM_GROUP_SIZE] = {[0 ... CM_GROUP_SIZE - 1] =
                                                          CM_CTRL_EMPTY};
 
diff --git a/cheesemap.h b/cheesemap.h
index 7cdb2de..5d87a6c 100644
--- a/cheesemap.h
+++ b/cheesemap.h
@@ -13,14 +13,20 @@ extern "C" {
 #include <stdbool.h>
 #include <stdint.h>
 
-#include CM_OPT_ASSERT_PATH
-#ifndef assert
-#error "assert is not defined"
+#ifdef CM_SSE2
+#include <emmintrin.h>
+
+typedef __m128i group_t;
+typedef uint16_t bitmask_t;
+#define CM_GROUP_SIZE 16
+#define CM_NO_FALLBACK
 #endif
 
+#ifndef CM_NO_FALLBACK
 typedef uintptr_t group_t;
 typedef group_t bitmask_t;
 #define CM_GROUP_SIZE __SIZEOF_POINTER__
+#endif
 
 ////////////////////////////////
 // cheesemap callback functions
@@ -63,15 +69,6 @@ enum {
 
 };
 
-static inline uintptr_t cm_h1(cm_hash_t hash) {
-  return (uintptr_t)(hash >> CM_FP_SIZE);
-}
-
-static inline uint8_t cm_h2(cm_hash_t hash) {
-  uintptr_t top = hash >> (sizeof(cm_hash_t) * CHAR_BIT - CM_FP_SIZE);
-  return (uint8_t)(top & CM_H2_MASK);
-}
-
 extern const uint8_t CM_CTRL_STATIC_EMPTY[CM_GROUP_SIZE];
 
 struct cheesemap_raw {
diff --git a/makefile b/makefile
index dacbbe5..982698c 100644
--- a/makefile
+++ b/makefile
@@ -5,6 +5,7 @@ CM_OPT_ASSERT_PATH ?= <assert.h>
 CM_OPT_RELEASE ?= 1
 CM_OPT_ENABLE_UBSAN ?= 0
 CM_OPT_ENABLE_ASAN ?= 0
+CM_OPT_ENABLE_SSE2 ?= 0
 
 CC ?= gcc
 
@@ -39,6 +40,10 @@ ifeq ($(CM_OPT_ENABLE_ASAN),1)
 	CM_CC_FLAGS += -fsanitize=address
 endif
 
+ifeq ($(CM_OPT_ENABLE_SSE2),1)
+	CM_CC_FLAGS += -DCM_SSE2=1
+endif
+
 .PHONY: all
 all: $(CM_OBJECT) $(CM_DEMO)
 
-- 
cgit v1.2.3