Optimize small structure copying on x86_64

lib/bcheck.c

@@ -292,6 +292,7 @@ DLL_EXPORT int __bound_strncmp(const char *s1, const char *s2, size_t n);
 DLL_EXPORT char *__bound_strcat(char *dest, const char *src);
 DLL_EXPORT char *__bound_strchr(const char *string, int ch);
 DLL_EXPORT char *__bound_strdup(const char *s);
+DLL_EXPORT void __bound_struct_copy(void *dst,void *src,size_t size);
 
 #if defined(__arm__) && defined(__ARM_EABI__)
 DLL_EXPORT void *__bound___aeabi_memcpy(void *dst, const void *src, size_t size);
@@ -424,6 +425,7 @@ static unsigned long long bound_strncmp_count;
 static unsigned long long bound_strcat_count;
 static unsigned long long bound_strchr_count;
 static unsigned long long bound_strdup_count;
+static unsigned long long bound_struct_copy_count;
 static unsigned long long bound_not_found;
 #define INCR_COUNT(x)          ++x
 #else
@@ -1765,6 +1767,16 @@ void *__bound_memset(void *s, int c, size_t n)
     return memset(s, c, n);
 }
 
+void __bound_struct_copy(void *dst,void *src,size_t size)
+{
+    dprintf(stderr, "Copy struct from %p to %p,size %lx\n",
+            src,dst,size);
+    INCR_COUNT(bound_struct_copy_count);
+    __bound_check(dst,size,"struct copy destination");
+    __bound_check(src,size,"struct copy source");
+    return;
+}
+
 #if defined(__arm__) && defined(__ARM_EABI__)
 void *__bound___aeabi_memcpy(void *dest, const void *src, size_t n)
 {

tcc.h

@@ -1644,6 +1644,8 @@ ST_FUNC void gen_increment_tcov (SValue *sv);
 
 /* ------------ x86_64-gen.c ------------ */
 #ifdef TCC_TARGET_X86_64
+#define TCC_TARGET_NATIVE_STRUCT_COPY
+ST_FUNC void gen_struct_copy(int size);
 ST_FUNC void gen_addr64(int r, Sym *sym, int64_t c);
 ST_FUNC void gen_opl(int op);
 #ifdef TCC_TARGET_PE

tccgen.c

@@ -3474,9 +3474,38 @@ ST_FUNC void vstore(void)
     if (sbt == VT_STRUCT) {
         /* if structure, only generate pointer */
         /* structure assignment : generate memcpy */
-        /* XXX: optimize if small size */
             size = type_size(&vtop->type, &align);
 
+#ifdef TCC_TARGET_NATIVE_STRUCT_COPY
+            if (size <= (PTR_SIZE << 4)) {
+		    vswap();
+#ifdef CONFIG_TCC_BCHECK
+                    if (vtop->r & VT_MUSTBOUND)
+                        gbound();
+#endif
+                    vtop->type.t = VT_PTR;
+                    gaddrof();
+
+                    vpushv(vtop - 1);
+#ifdef CONFIG_TCC_BCHECK
+                    if (vtop->r & VT_MUSTBOUND)
+                        gbound();
+#endif
+                    vtop->type.t = VT_PTR;
+                    gaddrof();                  /* src dest src */
+#ifdef CONFIG_TCC_BCHECK
+                    if (tcc_state->do_bounds_check) {
+                        vpush_helper_func(TOK___bound_struct_copy);
+                        vpushv(vtop - 2);
+                        vpushv(vtop - 2);
+                        vpushi(size);
+                        gfunc_call(3);
+                    }
+#endif
+
+                    gen_struct_copy(size);
+            } else {
+#endif
             /* destination */
             vswap();
 #ifdef CONFIG_TCC_BCHECK
@@ -3510,7 +3539,9 @@ ST_FUNC void vstore(void)
             vpushi(size);
             gfunc_call(3);
         /* leave source on stack */
-
+#ifdef TCC_TARGET_NATIVE_STRUCT_COPY
+            }
+#endif
     } else if (ft & VT_BITFIELD) {
         /* bitfield store handling */
 

tcctok.h

@@ -330,6 +330,7 @@
      DEF(TOK___bound_setjmp, "__bound_setjmp")
      DEF(TOK___bound_longjmp, "__bound_longjmp")
      DEF(TOK___bound_new_region, "__bound_new_region")
+     DEF(TOK___bound_struct_copy,"__bound_struct_copy")
 # ifdef TCC_TARGET_PE
 #  ifdef TCC_TARGET_X86_64
      DEF(TOK___bound_alloca_nr, "__bound_alloca_nr")

x86_64-gen.c

@@ -2282,6 +2282,56 @@ ST_FUNC void gen_vla_alloc(CType *type, int align) {
     }
 }
 
+/*
+ * Assmuing the top part of the stack looks like below,
+ *  src dest src
+ */
+void gen_struct_copy(int size)
+{
+    save_reg(TREG_RSI);
+    load(TREG_RSI,vtop);
+    vtop->r = TREG_RSI;
+    vswap();                    /* dest src src */
+    save_reg(TREG_RDI);
+    load(TREG_RDI,vtop);
+    vtop->r = TREG_RDI;
+    /*  Not aligned by 8bytes   */
+    if (size & 0x04) {
+        o(0xa5);
+    }
+    if (size & 0x02) {
+        o(0xa566);
+    }
+    if (size & 0x01) {
+        o(0xa4);
+    }
+
+    size >>= 3;
+    if (!size)
+        goto done;
+    /*  Although this function is only called when the struct is smaller    */
+    /*  than 32 bytes(4 * PTR_SIZE),a common implementation is included     */
+    if (size <= 4 && size) {
+        switch (size) {
+            case 4: o(0xa548);
+            case 3: o(0xa548);
+            case 2: o(0xa548);
+            case 1: o(0xa548);
+        }
+    } else {
+        save_reg(TREG_RCX);
+        vpushi(size);
+        load(TREG_RCX,vtop);
+        vtop->r = TREG_RCX;
+        o(0xa548f3);
+        vpop();
+    }
+done:
+    vpop();
+    vpop();
+    return;
+}
+
 
 /* end of x86-64 code generator */
 /*************************************************************/