daily update on 20250208 by nitload youmu. lexical analyzer complete but little problem

内核源代码/StgSharp.C/StgSharpC.c

@@ -1,7 +1,6 @@
 #include "StgSharpC.h"
 #include "gl.h"
 #include "glfw3.h"
-#include "ssc_framework.h"
 #include <stdio.h>
 
 GladGLContext *currentContext;

内核源代码/StgSharp.C/StgSharpC.h

@@ -5,7 +5,6 @@
 #ifndef SSC
 #define SSC
 
-#include "ssc_framework.h"
 #include "ssc_internal.h"
 
 #define STBIDEF SSCAPI

内核源代码/StgSharp.C/StgSharpC.vcxproj

@@ -181,7 +181,6 @@
     <ClInclude Include="glfw3native.h" />
     <ClInclude Include="ssc_intrinsic.h" />
     <ClInclude Include="khrplatform.h" />
-    <ClInclude Include="ssc_framework.h" />
     <ClInclude Include="ssc_internal.h" />
     <ClInclude Include="stb_image.h" />
     <ClInclude Include="StgSharpC.h" />

内核源代码/StgSharp.C/dllmain.c

@@ -1,8 +1,4 @@
 
-#ifdef _WIN32
-#include "ssc_framework.h"
-#endif // __
-
 #include "StgSharpC.h"
 
 #ifdef _WIN32

内核源代码/StgSharp.C/hash.c

@@ -1,8 +1,6 @@
 #include "ssc_intrinsic.h"
 #include "StgSharpC.h"
 
-static inline uint32_t fmix32(uint32_t hash);
-
 INTERNAL int SSCDECL city_hash_simplify_sse(byte const *str, int const length)
 {
         uint32_t hash = length, size = length << 1;
@@ -42,5 +40,5 @@ INTERNAL int SSCDECL city_hash_simplify_sse(byte const *str, int const length)
         hash ^= hash >> 13;
         hash *= 0xc2b2ae35;
         hash ^= hash >> 16;
-        return fmix32(hash);
+        return hash;
 }

内核源代码/StgSharp.C/ssc_intrinsic.h

@@ -106,7 +106,7 @@ INTERNAL void SSCDECL dot_43_512(__4_columnset *transpose, __m128 *vector, __m12
 #pragma region string
 
 INTERNAL int SSCDECL city_hash_simplify_sse(byte const *str, int length);
-INTERNAL int SSCDECL index_pair_sse(byte const *str, uint32_t target, int length);
+INTERNAL int SSCDECL index_pair_sse(short const *str, uint32_t target, int length);
 
 #pragma endregion
 
@@ -122,7 +122,7 @@ typedef struct ssc_intrinsic {
         DOTPROC dot_41;
         DOTPROC dot_42;
         DOTPROC dot_43;
-        int (*index_pair)(byte *str, byte *target, int length, int start);
+        int (*index_pair)(short const *str, uint32_t target, int length);
         VECTORNORMALIZEPROC normalize_3;
         MATRIXTRANSPOSEPROC transpose_23;
         MATRIXTRANSPOSEPROC transpose_24;

内核源代码/StgSharp.C/string.c

@@ -1,27 +1,50 @@
 #include "StgSharpC.h"
 #include "ssc_intrinsic.h"
 
-int index_pair_sse(byte const *str, uint32_t target, int length)
+int index_pair_sse(short const *str, uint32_t target, int length)
 {
-        int index = 0, result = 0, size = length << 1, end = size - 14;
-        __m128i source = { .m128i_i32 = {
+        int index = 0, result = 0, size = length, end = size - 8;
+        __m128i s_even = { .m128i_u32 = {
                                    target,
                                    target,
                                    target,
                                    target,
                            } };
-        for (; (index < end) && (!result); index += 14) { // move 7*2 bytes instead of 8*2
-                __m128i vec = _mm_loadu_epi32(str + index);
-                __m128i com = _mm_cmpeq_epi32(vec, source);
+        __m128i vec, com;
+
+        while (index < end) {
+                vec = _mm_loadu_epi32(str + index);
+                com = _mm_cmpeq_epi16(vec, s_even);
                 result = _mm_movemask_epi8(com);
-        }
-        if (result) {
-                for (; result; result >> 2) {
-                        index++;
+                if (result) {
+                        for (size_t i = 0; i < 12; i++) {
+                                result >>= 2;
+                                if (((result & 0xF) == 0xF)) {
+                                        return index + i;
+                                }
+                                if (result == 0) {
+                                        break;
+                                }
+                        }
                 }
-                return index;
+                index += 1;
+                vec = _mm_loadu_epi32(str + index);
+                com = _mm_cmpeq_epi16(vec, s_even);
+                result = _mm_movemask_epi8(com);
+                if (result) {
+                        for (size_t i = 0; i < 12; i++) {
+                                result >>= 2;
+                                if (((result & 0xF) == 0xF)) {
+                                        return index + i;
+                                }
+                                if (result == 0) {
+                                        break;
+                                }
+                        }
+                }
+                index += 7;
         }
-        for (; (index < size - 2) && (!result); index += 2) {
+        for (; (index < size - 1) && (!result); index += 2) {
                 result = *(int *)(str + index) == target;
         }
         if (result) {