fix: poor performance in atof32

bytedance · Sep 13, 2022 · be18179 · be18179
1 parent 734b06e
commit be18179
Show file tree

Hide file tree

Showing 17 changed files with 2,586 additions and 4,793 deletions.
diff --git a/encoder/assembler_amd64_go116.go b/encoder/assembler_amd64_go116.go
@@ -852,7 +852,8 @@ func (self *_Assembler) _asm_OP_f32(_ *_Instr) {
     self.save_c()                                       // SAVE     $C_regs
     self.rbuf_di()                                      // MOVQ     RP, DI
     self.Emit("MOVSS"   , jit.Ptr(_SP_p, 0), _X0)       // MOVSS    (SP.p), X0
-    self.call_c(_F_f32toa)                              // CALL_C   f64toa
+    self.Emit("CVTSS2SD", _X0, _X0)                     // CVTSS2SD X0, X0
+    self.call_c(_F_f32toa)                              // CALL_C   f32toa
     self.Emit("ADDQ"    , _AX, _RL)                     // ADDQ     AX, RL
 }
 

diff --git a/encoder/assembler_amd64_go117.go b/encoder/assembler_amd64_go117.go
@@ -865,7 +865,8 @@ func (self *_Assembler) _asm_OP_f32(_ *_Instr) {
     self.save_c()                                       // SAVE     $C_regs
     self.rbuf_di()                                      // MOVQ     RP, DI
     self.Emit("MOVSS"   , jit.Ptr(_SP_p, 0), _X0)       // MOVSS    (SP.p), X0
-    self.call_c(_F_f32toa)                              // CALL_C   f64toa
+    self.Emit("CVTSS2SD", _X0, _X0)                     // CVTSS2SD X0, X0
+    self.call_c(_F_f32toa)                              // CALL_C   f32toa
     self.Emit("ADDQ"    , _AX, _RL)                     // ADDQ     AX, RL
 }
 

diff --git a/internal/native/avx/fastfloat_amd64_test.go b/internal/native/avx/fastfloat_amd64_test.go
diff --git a/internal/native/avx/native_amd64.go b/internal/native/avx/native_amd64.go
diff --git a/internal/native/avx/native_amd64.s b/internal/native/avx/native_amd64.s
diff --git a/internal/native/avx/native_subr_amd64.go b/internal/native/avx/native_subr_amd64.go
diff --git a/internal/native/avx2/fastfloat_amd64_test.go b/internal/native/avx2/fastfloat_amd64_test.go
diff --git a/internal/native/avx2/native_amd64.go b/internal/native/avx2/native_amd64.go