build: full inline native C funcs (#619)

.gitmodules

@@ -1,4 +1,4 @@
-[submodule "tools/asm2asm"]
+[submodule "cloudwego"]
 	path = tools/asm2asm
 	url = https://github.com/cloudwego/asm2asm.git
 [submodule "tools/simde"]

fuzz/ast_fuzz_test.go

@@ -39,7 +39,7 @@ func fuzzASTGetFromObject(t *testing.T, data []byte, m map[string]interface{}) {
 		v, err := node.Interface()
 		require.NoErrorf(t, err, "error in node convert\n%s", msg)
 		require.Equalf(t, v, expv, "error in node equal\n%sGot:\n%s\nExp:\n%s\n", 
-			msg, spew.Sdump(v), spew.Sdump(expv))
+		msg, spew.Sdump(v), spew.Sdump(expv))
 	}
 }
 

internal/decoder/assembler_regabi_amd64.go

@@ -134,6 +134,7 @@ var (
     _R9 = jit.Reg("R9")
     _X0 = jit.Reg("X0")
     _X1 = jit.Reg("X1")
+    _X15 = jit.Reg("X15")
 )
 
 var (
@@ -421,6 +422,7 @@ func (self *_Assembler) call_go(fn obj.Addr) {
 func (self *_Assembler) callc(fn obj.Addr) {
     self.save(_IP)
     self.call(fn)
+    self.Emit("XORPS", _X15, _X15)
     self.load(_IP)
 }
 
@@ -605,7 +607,6 @@ func (self *_Assembler) skip_one() {
     self.Emit("TESTQ", _AX, _AX)                // TESTQ   AX, AX
     self.Sjmp("JS"   , _LB_parsing_error_v)     // JS      _parse_error_v
     self.Emit("MOVQ" , _VAR_pc, _R9)            // MOVQ    pc, R9
-    // self.Byte(0xcc)
     self.Rjmp("JMP"  , _R9)                     // JMP     (R9)
 }
 

internal/decoder/primitives.go

@@ -30,9 +30,7 @@ func decodeTypedPointer(s string, i int, vt *rt.GoType, vp unsafe.Pointer, sb *_
         return 0, err
     } else {
         rt.MoreStack(_FP_size + _VD_size + native.MaxFrameSize)
-        rt.StopProf()
         ret, err := fn(s, i, vp, sb, fv, "", nil)
-        rt.StartProf()
         return ret, err
     }
 }

internal/encoder/assembler_regabi_amd64.go

@@ -128,6 +128,7 @@ var (
 
 var (
     _X0 = jit.Reg("X0")
+    _X15 = jit.Reg("X15")
     _Y0 = jit.Reg("Y0")
 )
 
@@ -512,6 +513,7 @@ func (self *_Assembler) call_c(pc obj.Addr) {
     self.call(pc)               // CALL $pc
     self.xload(_REG_ffi...)     // LOAD $REG_ffi
     self.Emit("XCHGQ", _SP_p, _BX)
+    self.Emit("XORPS", _X15, _X15)
 }
 
 func (self *_Assembler) call_go(pc obj.Addr) {

internal/encoder/primitives.go

@@ -23,6 +23,7 @@ import (
 
     `github.com/bytedance/sonic/internal/jit`
     `github.com/bytedance/sonic/internal/native`
+    `github.com/bytedance/sonic/internal/native/types`
     `github.com/bytedance/sonic/internal/rt`
 )
 
@@ -73,15 +74,11 @@ func encodeTypedPointer(buf *[]byte, vt *rt.GoType, vp *unsafe.Pointer, sb *_Sta
         return err
     } else if vt.Indirect() {
         rt.MoreStack(_FP_size + native.MaxFrameSize)
-        rt.StopProf()
         err := fn(buf, *vp, sb, fv)
-        rt.StartProf()
         return err
     } else {
         rt.MoreStack(_FP_size + native.MaxFrameSize)
-        rt.StopProf()
         err := fn(buf, unsafe.Pointer(vp), sb, fv)
-        rt.StartProf()
         return err
     }
 }
@@ -123,8 +120,8 @@ func htmlEscape(dst []byte, src []byte) []byte {
     dbuf := (*rt.GoSlice)(unsafe.Pointer(&dst))
 
     /* grow dst if it is shorter */
-    if cap(dst) - len(dst) < len(src) + native.BufPaddingSize {
-        cap :=  len(src) * 3 / 2 + native.BufPaddingSize
+    if cap(dst) - len(dst) < len(src) + types.BufPaddingSize {
+        cap :=  len(src) * 3 / 2 + types.BufPaddingSize
         *dbuf = growslice(typeByte, *dbuf, cap)
     }