function being redirected but also the soname of the object that it is
in. This makes it a bit easier to diagnose redirection problems.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@14051
* add a trace_DIE function
* use it to trace a bad DIE
and to trace all DIEs that are (maybe) read
(due to the "avoid read twice" optimisation, the tracing was not
so easy to read anymore => add an explicit trace_DIE call at the beginning
of read_DIE)
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@14050
by a DIE parser
Instead of pre-reading the DIE, first let the parser(s) possibly
parse the DIE. Read (to skip) the DIE data if no parser has parsed it.
OTherwise, just jump to the end of the DIE as established by the parser
that has read the DIE.
This slightly improves the reading of inlined info.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@14049
A recent gnatpro version is needed for this to work.
Thanks to these intercepts, some false positive errors are avoided,
and helgrind properly recuperates some internal memory associated
to the terminated task.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@14046
Wrong place for the assertion for the inlparser
+ move the "zero the parsers" out of the "if VG_(clo*)" conditions
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@14044
of DIEs when one or more parsers will read them also)
+ add the name of the parser in the barf output.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@14041
showing inlined function calls.
See 278972 valgrind stacktraces and suppression do not handle inlined function call debuginfo
Reading the inlined dwarf call info is activated using the new clo
--read-inline-info=yes
Default is currently no but an objective is to optimise the performance
and memory in order to possibly set it on by default.
(see below discussion about performances).
Basically, the patch provides the following pieces:
1. Implement a new dwarf3 reader that reads the inlined call info
2. Some performance improvements done for this new parser, and
on some common code between the new parser and the var info parser.
3. Use the parsed inlined info to produce stacktrace showing inlined calls
4. Use the parsed inlined info in the suppression matching and suppression generation
5. and of course, some reg tests
1. new dwarf3 reader:
---------------------
Two options were possible: add the reading of the inlined info
in the current var info dwarf reader, or add a 2nd reader.
The 2nd approach was preferred, for the following reasons:
The var info reader is slow, memory hungry and quite complex.
Having a separate parsing phase for the inlined information
is simpler/faster when just reading the inlined info.
Possibly, a single parser would be faster when using both
--read-var-info=yes and --read-inline-info=yes.
However, var-info being extremely memory/cpu hungry, it is unlikely
to be used often, and having a separate parsing for inlined info
does in any case make not much difference.
(--read-var-info=yes is also now less interesting thanks to commit
r13991, which provides a fast and low memory "reasonable" location
for an address).
The inlined info parser reads the dwarf info to make calls
to priv_storage.h ML_(addInlInfo).
2. performance optimisations
----------------------------
* the abbrev cache has been improved in revision r14035.
* The new parser skips the non interesting DIEs
(the var-info parser has no logic to skip uninteresting DIEs).
* Some other minor perf optimisation here and there.
In total now, on a big executable, 15 seconds CPU are needed to
create the inlined info (on my slow x86 pentium).
With regards to memory, the dinfo arena:
with inlined info: 172281856/121085952 max/curr mmap'd
without : 157892608/106721280 max/curr mmap'd,
So, basically, inlined information costs about 15Mb of memory for
my big executable (compared to first version of the patch, this is
already using less memory, thanks to the strpool deduppoolalloc.
The needed memory can probably be decreased somewhat more.
3. produce better stack traces
------------------------------
VG_(describe_IP) has a new argument InlIPCursor *iipc which allows
to describe inlined function calls by doing repetitive calls
to describe_IP. See pub_tool_debuginfo.h for a description.
4. suppression generation and matching
--------------------------------------
* suppression generation now also uses an InlIPCursor *iipc
to generate a line for each inlined fn call.
* suppression matching: to allow suppression matching to
match one IP to several function calls in a suppression entry,
the 'inputCompleter' object (that allows to lazily generate
function or object names for a stacktrace when matching
an error with a suppression) has been generalised a little bit
more to also lazily generate the input sequence.
VG_(generic_match) has been updated so as to be more generic
with respect to the input completer : when providing an
input completer, VG_(generic_match) does not need anymore
to produce/compute any input itself : this is all delegated
to the input completer.
5. various regtests
-------------------
to test stack traces with inlined calls, and suppressions
of (some of) these errors using inlined fn calls matching.
Work still to do:
-----------------
* improve parsing performance
* improve the memory overhead.
* handling the directory name for files of the inlined function calls is not yet done.
(probably implies to refactor some code)
* see if m_errormgr.c *offsets arrays cannot be managed via xarray
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@14036
For each DIE, the dwarf3 reader must know which data elements to read.
These elements are described by an abbreviation.
Re-reading these abbreviations for each DIE is costly as
the location of the needed abbreviation is found by scanning the full
abbv section, which is very costly.
(A small cache of 32 abbv offsets in the abbv section somewhat decreases
the cost, but reading the abbvs is still a hot spot, in particular for
big debug informations).
This patch:
* adds an hash table of parsed abbreviations
* all abbreviations for a CU are read in one single scan of the abbv
section, when the CU header is read
So, with the patch, the di image is not accessed anymore for reading the abbvs
after the CU header parsing.
On a big executable, --read-var-info=yes user cpu changes from
trunk: 320 seconds
to
abbv cache: 270 seconds
This further improves on a previous (not committed) abbv cache that
was just caching up to 513 entries in the abbv pos cache and populating
the cache with an initial scan. The user cpu for this version was 285 seconds.
NB: this is some work in anticipation of a following patch that
will add reading dwarf3 inlined information, with the hope to make
this reading fast enough to activate it by default.
Note: on the examples I looked at, all abbreviations were numbered starting
from 1, with no holes. If that would always be the case, then one could use
an xarray of parsed abbreviations rather than an hash table. However,
I found nothing in the dwarf standard that guarantees that abbreviations
are numbered from 1. So, the hash table.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@14035
It is possible that a debug info contains no string (and so strpool
is never allocated).
A protection to avoid accessing strpool was already necessary
in ML_(canonicaliseTables) :
if (di->strpool)
VG_(freezeDedupPA) (di->strpool);
So, if a similar debug info is released, we need the same protection
to avoid accessing a NULL strpool.
Detect by Julian on arm64, but not (at least easily) reproduced on amd64.
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@14033
include/pub_tool_deduppoolalloc.h
coregrind/pub_core_deduppoolalloc.h
coregrind/m_deduppoolalloc.c
and uses it (currently only) for the strings in m_debuginfo/storage.c
The idea is that such ddup pool allocator will also be used for other
highly duplicated information (e.g. the DiCFSI information), where
significant gains can also be achieved.
The dedup pool for strings also decreases significantly the memory
needed by the read inline information (patch still to be committed,
see bug 278972).
When testing with a big executable (tacot_process),
this reduces the size of the dinfo arena from
trunk: 158941184/109760512 max/curr mmap'd, 156775944/107882728 max/curr,
to
ddup: 157892608/106614784 max/curr mmap'd, 156362160/101414712 max/curr
(so 3Mb less mmap-ed once debug info is read, 1Mb less mmap-ed in peak,
6Mb less allocated once debug info is read).
This is all gained due to the string which changes from:
trunk: 17,434,704 in 266: di.storage.addStr.1
to
ddup: 10,966,608 in 750: di.storage.addStr.1
(6.5Mb less memory used by strings)
The gain in mmap-ed memory is smaller due to fragmentation.
Probably one could decrease the fragmentation by using bigger
size for the dedup pool, but then we would lose memory on the last
allocated pool (and for small libraries, we often do not use much
of a big pool block).
Solution might be to increase the pool size but have a "shrink_block"
operation. To be looked at in the future.
In terms of performance, startup of a big executable (on an old pentium)
is not influenced significantly (something like 0.1 seconds on 15 seconds
startup for a big executable, on a slow pentium).
The dedup pool uses a hash table. The hash function used currently
is the VG_(adler32) check sum. It is reported (and visible also here)
that this checksum is not a very good hash function (many collisions).
To have statistics about collisions, use --stats -v -v -v
As an example of the collisions, on the strings in debug info of memcheck tool on x86,
one obtain:
--4789-- dedupPA:di.storage.addStr.1 9983 allocs (8174 uniq) 11 pools (4820 bytes free in last pool)
--4789-- nr occurences of chains of len N, N-plicated keys, N-plicated elts
--4789-- N: 0 : nr chain 6975, nr keys 0, nr elts 0
--4789-- N: 1 : nr chain 3670, nr keys 6410, nr elts 8174
--4789-- N: 2 : nr chain 1070, nr keys 226, nr elts 0
--4789-- N: 3 : nr chain 304, nr keys 100, nr elts 0
--4789-- N: 4 : nr chain 104, nr keys 84, nr elts 0
--4789-- N: 5 : nr chain 72, nr keys 42, nr elts 0
--4789-- N: 6 : nr chain 44, nr keys 34, nr elts 0
--4789-- N: 7 : nr chain 18, nr keys 13, nr elts 0
--4789-- N: 8 : nr chain 17, nr keys 8, nr elts 0
--4789-- N: 9 : nr chain 4, nr keys 6, nr elts 0
--4789-- N:10 : nr chain 9, nr keys 4, nr elts 0
--4789-- N:11 : nr chain 1, nr keys 0, nr elts 0
--4789-- N:13 : nr chain 1, nr keys 1, nr elts 0
--4789-- total nr of unique chains: 12289, keys 6928, elts 8174
which shows that on 8174 different strings, we have only 6410 strings which have
a unique hash value. As other examples, N:13 line shows we have 13 strings
mapping to the same key. N:14 line shows we have 4 groups of 10 strings mapping to the
same key, etc.
So, adler32 is definitely a bad hash function.
Trials have been done with another hash function, giving a much lower
collision rate. So, a better (but still fast) hash function would probably
be beneficial. To be looked at ...
git-svn-id: svn://svn.valgrind.org/valgrind/trunk@14029
