Merge branch 'master' of github.com:vmprof/vmprof-python

Author: planrich

setup.py

@@ -25,6 +25,10 @@
         extra_compile_args = ['-Wno-unused']
         extra_compile_args += ['-DVMPROF_APPLE=1']
         extra_compile_args += ['-DVMPROF_UNIX=1']
+        # overwrite the optimization level, if it is not optimized enough,
+        # it might use the regiter rbx...
+        extra_compile_args += ['-g']
+        extra_compile_args += ['-O2']
     elif sys.platform.startswith('linux'):
         libraries = ['dl','unwind']
         extra_compile_args = ['-Wno-unused']
@@ -74,7 +78,7 @@
     name='vmprof',
     author='vmprof team',
     author_email='[email protected]',
-    version="0.4.0.dev9",
+    version="0.4.0.dev10",
     packages=find_packages(),
     description="Python's vmprof client",
     long_description='See https://vmprof.readthedocs.org/',

src/_vmprof.c

@@ -27,17 +27,29 @@ static destructor Original_code_dealloc = 0;
 PyObject* (*_default_eval_loop)(PyFrameObject *, int) = 0;
 
 #ifdef VMPROF_UNIX
-#ifdef __GNUC__
-__attribute__((optimize("O1")))
-#elif defined(__clang__)
+#ifdef __clang__
 __attribute__((disable_tail_calls))
+#elif defined(__GNUC__)
+__attribute__((optimize("O1")))
 #endif
 PyObject* vmprof_eval(PyFrameObject *f, int throwflag)
 {
+#ifdef X86_64
     register PyFrameObject * callee_saved asm("rbx");
+#elif defined(X86_32)
+    register PyFrameObject * callee_saved asm("edi");
+#else
+#    error "platform not supported"
+#endif
 
     asm volatile(
+#ifdef X86_64
         "movq %1, %0\t\n"
+#elif defined(X86_32)
+        "mov %1, %0\t\n"
+#else
+#    error "platform not supported"
+#endif
         : "=r" (callee_saved)
         : "r" (f) );
     return _default_eval_loop(f, throwflag);

src/_vmprof.h

@@ -49,6 +49,12 @@ PyObject* vmprof_eval(PyFrameObject *f, int throwflag);
 #define VMP_SUPPORTS_NATIVE_PROFILING
 #endif
 
+#ifdef __x86_64__
+#define X86_64
+#elif defined(__i386__)
+#define X86_32
+#endif
+
 #define MARKER_STACKTRACE '\x01'
 #define MARKER_VIRTUAL_IP '\x02'
 #define MARKER_TRAILER '\x03'

src/stack.c

@@ -15,6 +15,11 @@
 #ifdef VMP_SUPPORTS_NATIVE_PROFILING
 #define UNW_LOCAL_ONLY
 #include <libunwind.h>
+#  ifdef X86_64
+#    define REG_RBX UNW_X86_64_RBX
+#  elif defined(X86_32)
+#    define REG_RBX UNW_X86_EDI
+#  endif
 #endif
 
 #ifdef __APPLE__
@@ -161,12 +166,9 @@ int vmp_walk_and_record_stack(PY_STACK_FRAME_T *frame, void ** result,
         if ((void*)pip.start_ip == (void*)vmprof_eval) {
             // yes we found one stack entry of the python frames!
             unw_word_t rbx = 0;
-            if (unw_get_reg(&cursor, UNW_X86_64_RBX, &rbx) < 0) {
+            if (unw_get_reg(&cursor, REG_RBX, &rbx) < 0) {
                 break;
             }
-            if (top_most_frame == NULL) {
-                top_most_frame = (PyFrameObject*)rbx;
-            }
             if (rbx != (unw_word_t)top_most_frame) {
                 // uh we are screwed! the ip indicates we are have context
                 // to a PyEval_EvalFrameEx function, but when we tried to retrieve

src/trampoline.c

@@ -1,8 +1,9 @@
 #include "trampoline.h"
 
 #include "machine.h"
+#include "_vmprof.h"
 
-#define _GNU_SOURCE
+#define _GNU_SOURCE 1
 #include <string.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -59,17 +60,11 @@ static char * g_trampoline = NULL;
 // the machine code size copied over from the callee
 static int g_trampoline_length;
 
-void _jmp_to(char * a, uintptr_t addr, int call) {
-
-    // TODO 32-bit
-
+int _jmp_to(char * a, uintptr_t addr) {
+#ifdef X86_64
     // moveabsq <addr>, <reg>
     a[0] = 0x48; // REX.W
-    if (call) {
-        a[1] = 0xb8; // %rax
-    } else {
-        a[1] = 0xba; // %rdx
-    }
+    a[1] = 0xba; // %rdx
     a[2] = addr & 0xff;
     a[3] = (addr >> 8) & 0xff;
     a[4] = (addr >> 16) & 0xff;
@@ -79,21 +74,67 @@ void _jmp_to(char * a, uintptr_t addr, int call) {
     a[8] = (addr >> 48) & 0xff;
     a[9] = (addr >> 56) & 0xff;
 
-    if (call) {
-        a[10] = 0xff;
-        a[11] = 0xd0;
-    } else {
-        a[10] = 0xff;
-        a[11] = 0xe2;
+    // jmp %edx
+    a[10] = 0xff;
+    a[11] = 0xe2;
+    return 12;
+#elif defined(X86_32)
+    // mov <addr>, %edx
+    a[0] = 0xba;
+    a[1] = addr & 0xff;
+    a[2] = (addr >> 8) & 0xff;
+    a[3] = (addr >> 16) & 0xff;
+    a[4] = (addr >> 24) & 0xff;
+    // jmp %edx
+    a[5] = 0xff;
+    a[6] = 0xe2;
+    return 7;
+#endif
+    return 0;
+}
+
+#ifdef X86_32
+int patch_relative_call(void * base, char * rel_call, char *rel_call_end, int bytes_after) {
+    if (bytes_after != 0) {
+        return 0;
     }
+
+    char * r = rel_call+1;
+
+    int off = r[0] |
+              ((r[1] & 0xff) << 8) |
+              ((r[2] & 0xff) << 16) |
+              ((r[3] & 0xff) << 24);
+    // instruction pointer is just after the whole instruction
+    intptr_t addr = (intptr_t)base + 5 + off;
+
+    rel_call[0] = 0xb8;
+    rel_call[1] = addr & 0xff;
+    rel_call[2] = (addr >> 8) & 0xff;
+    rel_call[3] = (addr >> 16) & 0xff;
+    rel_call[4] = (addr >> 24) & 0xff;
+    // jmp %edx
+    rel_call[5] = 0xff;
+    rel_call[6] = 0xd0;
+
+    return 2;
 }
+#endif
 
 // a hilarious typo, tramp -> trump :)
 int _redirect_trampoline_and_back(char * eval, char * trump, char * vmprof_eval) {
 
     char * trump_first_byte = trump;
+#ifdef X86_64
     int needed_bytes = 12;
+#elif defined(X86_32)
+    int needed_bytes = 8;
+    int relative_call_at_pos = -1;
+#else
+#   error "platform not supported"
+#endif
     int bytes = 0;
+    int off = 0;
     char * ptr = eval;
 
     // 1) copy the instructions that should be redone in the trampoline
@@ -102,6 +143,12 @@ int _redirect_trampoline_and_back(char * eval, char * trump, char * vmprof_eval)
         if (res == 0) {
             return 1;
         }
+#ifdef X86_32
+        if (ptr[0] == '\xe8') {
+            // occur on 32bit linux
+            relative_call_at_pos = bytes;
+        }
+#endif
         bytes += res;
         ptr += res;
     }
@@ -110,40 +157,18 @@ int _redirect_trampoline_and_back(char * eval, char * trump, char * vmprof_eval)
     // 2) initiate the first few instructions of the eval loop
     {
         (void)memcpy(trump, eval, bytes);
-        _jmp_to(trump+bytes, (uintptr_t)eval+bytes, 0);
-        //char * wptr = trump;
-        //*wptr++ = 0x55;
-
-        //*wptr++ = 0x48;
-        //*wptr++ = 0x89;
-        //*wptr++ = 0xe5;
-
-        //*wptr++ = 0x53;
-        //*wptr++ = 0x53;
-
-        //*wptr++ = 0x48;
-        //*wptr++ = 0x89;
-        //*wptr++ = 0xfb;
-
-        //char * trampcall = wptr;
-        //wptr += 12;
-
-        //// pop 
-        //*wptr++ = 0x5b;
-        //*wptr++ = 0x5b;
-        //*wptr++ = 0x5d;
-        //*wptr++ = 0xc3;
-
-        //_jmp_to(trampcall, (uintptr_t)wptr, 1);
-
-        //(void)memcpy(wptr, eval, bytes);
-        //wptr += bytes;
-        //_jmp_to(wptr, (uintptr_t)eval+bytes, 0);
+#ifdef X86_32
+        if (relative_call_at_pos != -1) {
+            off = patch_relative_call(eval+relative_call_at_pos, trump+relative_call_at_pos,
+                                          trump+relative_call_at_pos+5, bytes-relative_call_at_pos-5);
+        }
+#endif
+        _jmp_to(trump+bytes+off, (uintptr_t)eval+bytes);
     }
 
     // 3) overwrite the first few bytes of callee to jump to tramp
     // callee must call back 
-    _jmp_to(eval, (uintptr_t)vmprof_eval, 0);
+    _jmp_to(eval, (uintptr_t)vmprof_eval);
 
     return 0;
 }

travis/script.sh

@@ -9,7 +9,7 @@ fi
 py.test vmprof/ -vrs
 py.test jitlog/ -vrs
 
-if [[ -n "TRAVIS_TAG" && "$BUILD_LINUX_WHEEL" == "1" ]]; then
+if [[ -n "$TRAVIS_TAG" && "$BUILD_LINUX_WHEEL" == "1" ]]; then
     docker pull quay.io/pypa/manylinux1_x86_64
     docker pull quay.io/pypa/manylinux1_i686
     docker run --rm -v `pwd`:/io:Z quay.io/pypa/manylinux1_x86_64 bash /io/travis/build-wheels.sh;

vmprof/test/test_c_symboltable.py

@@ -15,7 +15,7 @@ class TestSymbolTable(object):
     def setup_class(cls):
         stack_ffi = FFI()
         stack_ffi.cdef("""
-        void dump_all_known_symbols(int fd);
+        //void dump_all_known_symbols(int fd);
         int test_extract(char ** name, int * lineno, char ** src);
         """)
         with open("src/symboltable.c", "rb") as fd:

vmprof/test/test_run.py

@@ -367,6 +367,9 @@ def walk(parent):
             if parent is None or len(parent.children) == 0:
                 return False
 
+            if 'n:native_gzipgzipgzip:' in parent.name:
+                return True
+
             for child in parent.children.values():
                 if 'n:native_gzipgzipgzip:' in child.name:
                     p = float(child.count) / parent.count