Author: Richard Plangger <planri...@gmail.com>
Branch: vmprof-native
Changeset: r90020:dffb13845738
Date: 2017-02-09 12:22 +0100
http://bitbucket.org/pypy/pypy/changeset/dffb13845738/
Log: apply changes made in pypy source repo, and never ported back to
vmprof source
diff --git a/pypy/module/_vmprof/test/test__vmprof.py
b/pypy/module/_vmprof/test/test__vmprof.py
--- a/pypy/module/_vmprof/test/test__vmprof.py
+++ b/pypy/module/_vmprof/test/test__vmprof.py
@@ -24,10 +24,11 @@
i += 5 * WORD # header
assert s[i ] == '\x05' # MARKER_HEADER
assert s[i + 1] == '\x00' # 0
- assert s[i + 2] == '\x02' # VERSION_THREAD_ID
- assert s[i + 3] == chr(4) # len('pypy')
- assert s[i + 4: i + 8] == 'pypy'
- i += 8
+ assert s[i + 2] == '\x06' # VERSION_TIMESTAMP
+ assert s[i + 3] == '\x08' # PROFILE_RPYTHON
+ assert s[i + 4] == chr(4) # len('pypy')
+ assert s[i + 5: i + 9] == 'pypy'
+ i += 9
while i < len(s):
if s[i] == '\x03':
break
@@ -41,6 +42,8 @@
_, size = struct.unpack("ll", s[i:i + 2 * WORD])
count += 1
i += 2 * WORD + size
+ elif s[i] == '\x06':
+ i += 8+8+8
else:
raise AssertionError(ord(s[i]))
return count
diff --git a/pypy/module/_vmprof/test/test_direct.py
b/pypy/module/_vmprof/test/test_direct.py
--- a/pypy/module/_vmprof/test/test_direct.py
+++ b/pypy/module/_vmprof/test/test_direct.py
@@ -43,7 +43,7 @@
}
-""" + open(str(srcdir.join("vmprof_get_custom_offset.h"))).read(),
include_dirs=[str(srcdir)])
+""" + open(str(srcdir.join("shared/vmprof_get_custom_offset.h"))).read(),
include_dirs=[str(srcdir)])
class TestDirect(object):
def test_infrastructure(self):
diff --git a/rpython/rlib/rvmprof/rvmprof.py b/rpython/rlib/rvmprof/rvmprof.py
--- a/rpython/rlib/rvmprof/rvmprof.py
+++ b/rpython/rlib/rvmprof/rvmprof.py
@@ -185,15 +185,16 @@
This function can only be called once during translation.
It generates a C function called __vmprof_eval_vmprof which is used by
- the C source code as an extern function. This is necessary while walking
- the native stack. If you see __vmprof_eval_vmprof defined twice during
+ the vmprof C source code and is bound as an extern function.
+ This is necessary while walking the native stack.
+ If you see __vmprof_eval_vmprof defined twice during
translation, read on:
To remove this restriction do the following:
*) Extend the macro IS_VMPROF_EVAL in the vmprof source repo to check
several
sybmols.
- *) Give each function provided to this decorator a unique name in C
+ *) Give each function provided to this decorator a unique symbol name in C
"""
if _hack_update_stack_untranslated:
from rpython.rtyper.annlowlevel import llhelper
diff --git a/rpython/rlib/rvmprof/src/shared/vmprof_common.h
b/rpython/rlib/rvmprof/src/shared/vmprof_common.h
--- a/rpython/rlib/rvmprof/src/shared/vmprof_common.h
+++ b/rpython/rlib/rvmprof/src/shared/vmprof_common.h
@@ -60,8 +60,9 @@
char *vmprof_init(int fd, double interval, int memory,
int lines, const char *interp_name, int native)
{
- if (interval < 1e-6 || interval >= 1.0)
+ if (!(interval >= 1e-6 && interval < 1.0)) { /* also if it is NaN */
return "bad value for 'interval'";
+ }
prepare_interval_usec = (int)(interval * 1000000.0);
if (prepare_concurrent_bufs() < 0)
diff --git a/rpython/rlib/rvmprof/src/shared/vmprof_get_custom_offset.h
b/rpython/rlib/rvmprof/src/shared/vmprof_get_custom_offset.h
--- a/rpython/rlib/rvmprof/src/shared/vmprof_get_custom_offset.h
+++ b/rpython/rlib/rvmprof/src/shared/vmprof_get_custom_offset.h
@@ -1,120 +1,53 @@
#pragma once
-#ifdef PYPY_JIT_CODEMAP
void *pypy_find_codemap_at_addr(long addr, long *start_addr);
long pypy_yield_codemap_at_addr(void *codemap_raw, long addr,
long *current_pos_addr);
-long pypy_jit_stack_depth_at_loc(long loc);
-#endif
+#define MAX_INLINE_DEPTH 384
-#ifdef CPYTHON_GET_CUSTOM_OFFSET
-static void *tramp_start, *tramp_end;
-#endif
-
-
-static ptrdiff_t vmprof_unw_get_custom_offset(void* ip, void *cp) {
-
-#if defined(PYPY_JIT_CODEMAP)
-
- intptr_t ip_l = (intptr_t)ip;
- return pypy_jit_stack_depth_at_loc(ip_l);
-
-#elif defined(CPYTHON_GET_CUSTOM_OFFSET)
-
- if (ip >= tramp_start && ip <= tramp_end) {
- // XXX the return value is wrong for all the places before push and
- // after pop, fix
- void *bp;
- void *sp;
-
- /* This is a stage2 trampoline created by hotpatch:
-
- push %rbx
- push %rbp
- mov %rsp,%rbp
- and $0xfffffffffffffff0,%rsp // make sure the stack is
aligned
- movabs $0x7ffff687bb10,%rbx
- callq *%rbx
- leaveq
- pop %rbx
- retq
-
- the stack layout is like this:
-
- +-----------+ high addresses
- | ret addr |
- +-----------+
- | saved rbx | start of the function frame
- +-----------+
- | saved rbp |
- +-----------+
- | ........ | <-- rbp
- +-----------+ low addresses
-
- So, the trampoline frame starts at rbp+16, and the return address,
- is at rbp+24. The vmprof API requires us to return the offset of
- the frame relative to sp, hence we have this weird computation.
-
- XXX (antocuni): I think we could change the API to return directly
- the frame address instead of the offset; however, this require a
- change in the PyPy code too
- */
-
- unw_get_reg (cp, UNW_REG_SP, (unw_word_t*)&sp);
- unw_get_reg (cp, UNW_X86_64_RBP, (unw_word_t*)&bp);
- return bp+16+8-sp;
- }
- return -1;
-
-#else
-
- return -1;
-
-#endif
-}
-
-static long vmprof_write_header_for_jit_addr(void **result, long n,
- void *ip, int max_depth)
+static long vmprof_write_header_for_jit_addr(intptr_t *result, long n,
+ intptr_t addr, int max_depth)
{
#ifdef PYPY_JIT_CODEMAP
void *codemap;
long current_pos = 0;
- intptr_t id;
+ intptr_t ident, local_stack[MAX_INLINE_DEPTH];
+ long m;
long start_addr = 0;
- intptr_t addr = (intptr_t)ip;
- int start, k;
- void *tmp;
codemap = pypy_find_codemap_at_addr(addr, &start_addr);
- if (codemap == NULL)
- // not a jit code at all
+ if (codemap == NULL || n >= max_depth - 2)
+ // not a jit code at all or almost max depth
return n;
// modify the last entry to point to start address and not the random one
// in the middle
- result[n - 1] = (void*)start_addr;
- result[n] = (void*)2;
- n++;
- start = n;
- while (n < max_depth) {
- id = pypy_yield_codemap_at_addr(codemap, addr, ¤t_pos);
- if (id == -1)
+ result[n++] = VMPROF_ASSEMBLER_TAG;
+ result[n++] = start_addr;
+
+ // build the list of code idents corresponding to the current
+ // position inside this particular piece of assembler. If (very
+ // unlikely) we get more than MAX_INLINE_DEPTH recursion levels
+ // all inlined inside this single piece of assembler, then stop:
+ // there will be some missing frames then. Otherwise, we need to
+ // first collect 'local_stack' and then write it to 'result' in the
+ // opposite order, stopping at 'max_depth'. Previous versions of
+ // the code would write the oldest calls and then stop---whereas
+ // what we really need it to write the newest calls and then stop.
+ m = 0;
+ while (m < MAX_INLINE_DEPTH) {
+ ident = pypy_yield_codemap_at_addr(codemap, addr, ¤t_pos);
+ if (ident == -1)
// finish
break;
- if (id == 0)
+ if (ident == 0)
continue; // not main codemap
- result[n++] = (void *)id;
+ local_stack[m++] = ident;
}
- k = 0;
- while (k < (n - start) / 2) {
- tmp = result[start + k];
- result[start + k] = result[n - k - 1];
- result[n - k - 1] = tmp;
- k++;
- }
- if (n < max_depth) {
- result[n++] = (void*)3;
+ while (m > 0 && n < max_depth) {
+ result[n++] = VMPROF_JITTED_TAG;
+ result[n++] = local_stack[--m];
}
#endif
return n;
_______________________________________________
pypy-commit mailing list
pypy-commit@python.org
https://mail.python.org/mailman/listinfo/pypy-commit
