nop instuction byte reversed, optimisation added

.github/workflows/ci.yml

@@ -27,7 +27,7 @@ jobs:
       run: python -m pip install -e .[dev]
 
     - name: Build ops obj files and runner
-      run: bash build.sh
+      run: bash src/copapy/obj/nativecompile.sh
 
     - name: Run tests with pytest
       run: pytest

.gitignore

@@ -13,9 +13,11 @@ __pycache__
 /bin/*
 token.txt
 /src/copapy/obj/*.o
+/src/copapy/obj/old/*.o
 runmem2
 src/*.so
 bin/*
 build/*
 /*.obj
 /src/*.pyd
+src/copapy/stencils.c

src/copapy/generate_stencils.py

@@ -9,7 +9,7 @@ def get_function_start() -> str:
     return """
     int function_start(){
         result_int(0);  // dummy call instruction before marker gets striped
-        asm volatile (".long 0x0F1F4400E2");
+        asm volatile (".long 0xE200441F0F");
         return 1;
     }
     """
@@ -19,7 +19,7 @@ def get_function_end() -> str:
     return """
     int function_end(){
         result_int(0);
-        asm volatile (".long 0x0F1F4400E1");
+        asm volatile (".long 0xE100441F0F");
         return 1;
     }
     """
@@ -28,9 +28,9 @@ def get_function_end() -> str:
 def get_op_code(op: str, type1: str, type2: str, type_out: str) -> str:
     return f"""
     void {op}_{type1}_{type2}({type1} arg1, {type2} arg2) {{
-        asm volatile (".long 0x0F1F4400E1");
+        asm volatile (".long 0xE100441F0F");
         result_{type_out}_{type2}(arg1 {op_signs[op]} arg2, arg2);
-        asm volatile (".long 0x0F1F4400E2");
+        asm volatile (".long 0xE200441F0F");
     }}
     """
 
@@ -38,9 +38,9 @@ def get_op_code(op: str, type1: str, type2: str, type_out: str) -> str:
 def get_op_code_float(op: str, type1: str, type2: str) -> str:
     return f"""
     void {op}_{type1}_{type2}({type1} arg1, {type2} arg2) {{
-        asm volatile (".long 0x0F1F4400E1");
+        asm volatile (".long 0xE100441F0F");
         result_float_{type2}((float)arg1 {op_signs[op]} arg2, arg2);
-        asm volatile (".long 0x0F1F4400E2");
+        asm volatile (".long 0xE200441F0F");
     }}
     """
 
@@ -48,9 +48,9 @@ def get_op_code_float(op: str, type1: str, type2: str) -> str:
 def get_op_code_int(op: str, type1: str, type2: str) -> str:
     return f"""
     void {op}_{type1}_{type2}({type1} arg1, {type2} arg2) {{
-        asm volatile (".long 0x0F1F4400E1");
+        asm volatile (".long 0xE100441F0F");
         result_int_{type2}((int)(arg1 {op_signs[op]} arg2), arg2);
-        asm volatile (".long 0x0F1F4400E2");
+        asm volatile (".long 0xE200441F0F");
     }}
     """
 
@@ -70,9 +70,9 @@ def get_result_stubs2(type1: str, type2: str) -> str:
 def get_read_reg0_code(type1: str, type2: str, type_out: str) -> str:
     return f"""
     void read_{type_out}_reg0_{type1}_{type2}({type1} arg1, {type2} arg2) {{
-        asm volatile (".long 0x0F1F4400E1");
+        asm volatile (".long 0xE100441F0F");
         result_{type_out}_{type2}(dummy_{type_out}, arg2);
-        asm volatile (".long 0x0F1F4400E2");
+        asm volatile (".long 0xE200441F0F");
     }}
     """
 
@@ -80,9 +80,9 @@ def get_read_reg0_code(type1: str, type2: str, type_out: str) -> str:
 def get_read_reg1_code(type1: str, type2: str, type_out: str) -> str:
     return f"""
     void read_{type_out}_reg1_{type1}_{type2}({type1} arg1, {type2} arg2) {{
-        asm volatile (".long 0x0F1F4400E1");
+        asm volatile (".long 0xE100441F0F");
         result_{type1}_{type_out}(arg1, dummy_{type_out});
-        asm volatile (".long 0x0F1F4400E2");
+        asm volatile (".long 0xE200441F0F");
     }}
     """
 
@@ -90,10 +90,10 @@ def get_read_reg1_code(type1: str, type2: str, type_out: str) -> str:
 def get_write_code(type1: str) -> str:
     return f"""
     void write_{type1}({type1} arg1) {{
-        asm volatile (".long 0x0F1F4400E1");
+        asm volatile (".long 0xE100441F0F");
         dummy_{type1} = arg1;
         result_{type1}(arg1);
-        asm volatile (".long 0x0F1F4400E2");
+        asm volatile (".long 0xE200441F0F");
     }}
     """
 

src/copapy/obj/crosscompile.sh

@@ -19,7 +19,7 @@ OPT=O3
 mkdir -p $DEST
 
 # Native x86_64
-gcc-12 -c $SRC -o $DEST/stencils_x86_64_$OPT.o
+gcc-12 -$OPT -c $SRC -o $DEST/stencils_x86_64_$OPT.o
 
 # ARM64
 aarch64-linux-gnu-gcc-12 -$OPT -c $SRC -o $DEST/stencils_aarch64_$OPT.o

src/copapy/obj/nativecompile.sh

@@ -0,0 +1,14 @@
+#!/bin/bash
+set -e
+set -v
+
+SRC=src/copapy/stencils.c
+DEST=src/copapy/obj
+
+python src/copapy/generate_stencils.py
+
+mkdir -p $DEST
+gcc-12 -c $SRC -O0 -o $DEST/stencils_x86_64_O0.o
+gcc-12 -c $SRC -O1 -o $DEST/stencils_x86_64_O1.o
+gcc-12 -c $SRC -O2 -o $DEST/stencils_x86_64_O2.o
+gcc-12 -c $SRC -O3 -o $DEST/stencils_x86_64_O3.o

src/copapy/stencil_db.py

@@ -5,8 +5,8 @@ from enum import Enum
 
 ByteOrder = Literal['little', 'big']
 
-START_MARKER = 0x0F1F4400E1  # Nop on x86-64 
+START_MARKER = 0xE100441F0F    # Nop on x86-64 
-END_MARKER = 0x0F1F4400E2    # Nop on x86-64 
+END_MARKER = 0xE200441F0F    # Nop on x86-64 
 MARKER_LENGTH = 5
 
  # on x86_64: call or jmp instruction when tail call optimized 

src/copapy/stencils.c

@@ -1,291 +0,0 @@
-
-    // Auto-generated stencils for copapy
-    // Do not edit manually
-
-    volatile int dummy_int = 1337;
-    volatile float dummy_float = 1337;
-    
-    void result_int(int arg1);
-    
-    void result_float(float arg1);
-    
-    void result_int_int(int arg1, int arg2);
-    
-    void result_int_float(int arg1, float arg2);
-    
-    void result_float_int(float arg1, int arg2);
-    
-    void result_float_float(float arg1, float arg2);
-    
-    void add_int_int(int arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_int(arg1 + arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void add_int_float(int arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1 + arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void add_float_int(float arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_int(arg1 + arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void add_float_float(float arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1 + arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void sub_int_int(int arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_int(arg1 - arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void sub_int_float(int arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1 - arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void sub_float_int(float arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_int(arg1 - arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void sub_float_float(float arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1 - arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void mul_int_int(int arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_int(arg1 * arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void mul_int_float(int arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1 * arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void mul_float_int(float arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_int(arg1 * arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void mul_float_float(float arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1 * arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void div_int_int(int arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_int((float)arg1 / arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void div_int_float(int arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float((float)arg1 / arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void div_float_int(float arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_int(arg1 / arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void div_float_float(float arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1 / arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void gt_int_int(int arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_int(arg1 > arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void gt_int_float(int arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1 > arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void gt_float_int(float arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_int(arg1 > arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void gt_float_float(float arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1 > arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void eq_int_int(int arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_int(arg1 == arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void eq_int_float(int arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1 == arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void eq_float_int(float arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_int(arg1 == arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void eq_float_float(float arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1 == arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void mod_int_int(int arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_int(arg1 % arg2, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_int_reg0_int_int(int arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_int(dummy_int, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_int_reg1_int_int(int arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_int(arg1, dummy_int);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_float_reg0_int_int(int arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_int(dummy_float, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_float_reg1_int_int(int arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_float(arg1, dummy_float);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_int_reg0_int_float(int arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_float(dummy_int, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_int_reg1_int_float(int arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_int(arg1, dummy_int);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_float_reg0_int_float(int arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(dummy_float, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_float_reg1_int_float(int arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_float(arg1, dummy_float);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_int_reg0_float_int(float arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_int(dummy_int, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_int_reg1_float_int(float arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_int(arg1, dummy_int);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_float_reg0_float_int(float arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_int(dummy_float, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_float_reg1_float_int(float arg1, int arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1, dummy_float);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_int_reg0_float_float(float arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_int_float(dummy_int, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_int_reg1_float_float(float arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_int(arg1, dummy_int);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_float_reg0_float_float(float arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(dummy_float, arg2);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void read_float_reg1_float_float(float arg1, float arg2) {
-        asm volatile (".long 0xF17ECAFE");
-        result_float_float(arg1, dummy_float);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void write_int(int arg1) {
-        asm volatile (".long 0xF17ECAFE");
-        dummy_int = arg1;
-        result_int(arg1);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    void write_float(float arg1) {
-        asm volatile (".long 0xF17ECAFE");
-        dummy_float = arg1;
-        result_float(arg1);
-        asm volatile (".long 0xF27ECAFE");
-    }
-    
-    int function_start(){
-        result_int(0);  // dummy call instruction before marker gets striped
-        asm volatile (".long 0xF27ECAFE");
-        return 1;
-    }
-    
-    int function_end(){
-        result_int(0);
-        asm volatile (".long 0xF17ECAFE");
-        return 1;
-    }
-