order based on order of insertion.
This should cause both our warnings about these and the modules
serialization to be deterministic as a consequence.
Found by inspection.
llvm-svn: 233343
DeclIDs so that in addition to be grouped by file, the order of these
groups is stable.
Found by inspection, no test case. Not sure this can be observed without
a randomized seed for the hash table, but we shouldn't be relying on the
hash table layout under any circumstances.
llvm-svn: 233339
rewritten decls for Objective-C modules.
Found by inspection and completely obvious, so no test case. Many of the
remaining determinism fixes won't have precise test cases at this point,
but these are the kinds of things we wouldn't ask for a specific test of
during code review but ask authors to fix. The functionality isn't
changing, and should (he he!) already be tested.
llvm-svn: 233333
traversing the identifier table.
No easy test case as this table is somewhere between hard and impossible
to observe as non-deterministically ordered. The table is a hash table
but we hash the string contents and never remove entries from the table
so the growth pattern, etc, is all completely fixed. However, relying on
the hash function being deterministic is specifically against the
long-term direction of LLVM's hashing datastructures, which are intended
to provide *no* ordering guarantees. As such, this defends against these
things by sorting the identifiers. Sorting identifiers right before we
emit them to a serialized form seems a low cost for predictability here.
llvm-svn: 233332
logic removed.
This logic was both inserting all builtins into the identifier table and
ensuring they would get serialized. The first happens unconditionally
now, and we always write out the entire identifier table. This code can
simply go away.
llvm-svn: 233331
constructors in the current lexical context even though name lookup
found them via some other context merged into the redecl chain.
This can only happen for implicit constructors which can only have the
name of the type of the current context, so we can fix this by simply
*always* merging those names first. This also has the advantage of
removing the walk of the current lexical context from the common case
when this is the only constructor name we need to deal with (implicit or
otherwise).
I've enhanced the tests to cover this case (and uncovered an unrelated
bug which I fixed in r233325).
llvm-svn: 233327
deserializing an inherited constructor.
This is the exact same logic we use when deserializing method overrides
for the same reason: the canonical decl may end up pinned to a different
decl when we are improting modules, we need to re-pin to the canonical
one during reading.
My test case for this will come in a subsequent commit. I was trying to
test a more tricky bug fix and the test case happened to tickle this bug
as well.
llvm-svn: 233325
Eric Christopher pointed out that we have a check for assembly code
generation in a clang test, which isn't cool. We already have Driver
and back-end CodeGen tests for the .abiversion handling, so this
testing is unnecessary anyway. Make it go away.
llvm-svn: 233314
Unlike most of the other platforms supported by Clang, CloudABI only
supports static linkage, for the reason that global filesystem access is
prohibited. Functions provided by dlfcn.h are not present. As we know
that applications will not try to do any symbol lookups at run-time, we
can garbage collect unused code quite aggressively. Because of this, it
makes sense to enable -ffunction-sections and -fdata-sections by
default.
Object files will be a bit larger than usual, but the resulting binary
will not be affected, as the sections are merged again. However, when
--gc-sections is used, the linker is able to remove unused code far more
more aggressively. It also has the advantage that transitive library
dependencies only need to be provided to the linker in case that
functionality is actually used.
Differential Revision: http://reviews.llvm.org/D8635
Reviewed by: echristo
llvm-svn: 233299
Now that CloudABI's target information and header search logic for Clang
has been submitted, the only thing that remains to be done is adding
support for CloudABI's linker.
CloudABI uses Binutils ld, although there is some work to use lld
instead. This means that this code is largely based on what we use on
FreeBSD. There are some exceptions, however:
- Only static linking is performed. CloudABI does not support any
dynamically linked executables.
- CloudABI uses compiler-rt, libc++ and libc++abi unconditionally. Link
in these libraries instead of using libgcc_s, libstdc++, etc.
- We must ensure that the .eh_frame_hdr is present to make C++
exceptions work properly.
Differential Revision: http://reviews.llvm.org/D8250
llvm-svn: 233269
Clang was inserting these into a dense map. While it never iterated the
dense map during normal compilation, it did when emitting a module. Fix
this by using a standard MapVector to preserve the order in which we
encounter the late parsed templates.
I suspect this still isn't ideal, as we don't seem to remove things from
this map even when we mark the templates as no longer late parsed. But
I don't know enough about this particular extension to craft a nice,
subtle test case covering this. I've managed to get the stress test to
at least do some late parsing and demonstrate the core problem here.
This patch fixes the test and provides deterministic behavior which is
a strict improvement over the prior state.
I've cleaned up some of the code here as well to be explicit about
inserting when that is what is actually going on.
llvm-svn: 233264
templates. Turns out all of this works correctly (so far). But it should
cover more code paths and will let me test some things that don't
actually work next.
llvm-svn: 233263
deterministically.
This fixes a latent issue where even Clang's Sema (and diagnostics) were
non-deterministic in the face of this pragma. The fix is super simple --
just use a MapVector so we track the order in which these are parsed (or
imported). Especially considering how rare they are, this seems like the
perfect tradeoff. I've also simplified the client code with judicious
use of auto and range based for loops.
I've added some pretty hilarious code to my stress test which now
survives the binary diff without issue.
llvm-svn: 233261
updated decl contexts get emitted.
Since this code was added, we have newer vastly simpler code for
handling this. The code I'm removing was very expensive and also
generated unstable order of declarations which made module outputs
non-deterministic.
All of the tests continue to pass for me and I'm able to check the
difference between the .pcm files after merging modules together.
llvm-svn: 233251
non-visible definition, skip the new definition and make the old one visible
instead of trying to parse it again and failing horribly. C++'s ODR allows
us to assume that the two definitions are identical.
llvm-svn: 233250
decl context lookup tables.
The first attepmt at this caused problems. We had significantly more
sources of non-determinism that I realized at first, and my change
essentially turned them from non-deterministic output into
use-after-free. Except that they weren't necessarily caught by tools
because the data wasn't really freed.
The new approach is much simpler. The first big simplification is to
inline the "visit" code and handle this directly. That works much
better, and I'll try to go and clean up the other caller of the visit
logic similarly.
The second key to the entire approach is that we need to *only* collect
names into a stable order at first. We then need to issue all of the
actual 'lookup()' calls in the stable order of the names so that we load
external results in a stable order. Once we have loaded all the results,
the table of results will stop being invalidated and we can walk all of
the names again and use the cheap 'noload_lookup()' method to quickly
get the results and serialize them.
To handle constructors and conversion functions (whose names can't be
stably ordered) in this approach, what we do is record only the visible
constructor and conversion function names at first. Then, if we have
any, we walk the decls of the class and add those names in the order
they occur in the AST. The rest falls out naturally.
This actually ends up simpler than the previous approach and seems much
more robust.
It uncovered a latent issue where we were building on-disk hash tables
for lookup results when the context was a linkage spec! This happened to
dodge all of the assert by some miracle. Instead, add a proper predicate
to the DeclContext class and use that which tests both for function
contexts and linkage specs.
It also uncovered PR23030 where we are forming somewhat bizarre negative
lookup results. I've just worked around this with a FIXME in place
because fixing this particular Clang bug seems quite hard.
I've flipped the first part of the test case I added for stability back
on in this commit. I'm taking it gradually to try and make sure the
build bots are happy this time.
llvm-svn: 233249
More than 2x speedup on modules builds with large redecl chains.
Roughly 15-20% speedup on non-modules builds for very large TUs.
Between 2-3% cost in memory on large TUs.
llvm-svn: 233228
they enable/disable.
This fixes two things:
a) sse4 isn't actually a target feature, don't treat it as one.
b) we weren't correctly disabling sse4.1 when we'd pass -mno-sse4
after enabling it, thus passing preprocessor directives and
(soon) passing the function attribute as well when we shouldn't.
llvm-svn: 233223
Summary:
On my system, clang tries to invoke /path/to/arm-linux-androideabi-ld
as the linker for Android, and the regex inside the test file considers
this as unacceptable.
Reviewers: samsonov
Subscribers: tberghammer, aemerson, cfe-commits
Differential Revision: http://reviews.llvm.org/D8598
llvm-svn: 233211
Summary:
Due to CUDA's implicit address space casting, the type of a static local
variable may be more specific (i.e. with address space qualifiers) than
the type expected by the constructor. Emit an addrspacecast in that
case.
Test Plan: Clang used to crash on the added test.
Reviewers: nlewycky, pcc, eliben, rsmith
Reviewed By: eliben, rsmith
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D8575
llvm-svn: 233208
This patch adds Hardware Transaction Memory (HTM) support supported by ISA 2.07
(POWER8). The intrinsic support is based on GCC one [1], with both 'PowerPC HTM
Low Level Built-in Functions' and 'PowerPC HTM High Level Inline Functions'
implemented.
Along with builtins a new driver switch is added to enable/disable HTM
instruction support (-mhtm) and a header with common definitions (mostly to
parse the TFHAR register value). The HTM switch also sets a preprocessor builtin
HTM.
The HTM usage requires a recently newer kernel with PPC HTM enabled. Tested on
powerpc64 and powerpc64le.
This is send along a llvm patch to enabled the builtins and option switch.
[1]
https://gcc.gnu.org/onlinedocs/gcc/PowerPC-Hardware-Transactional-Memory-Built-in-Functions.html
Phabricator Review: http://reviews.llvm.org/D8248
llvm-svn: 233205
On android x86_32 the long double is only 64 bits (compared to 80 bits
on linux x86_32) and on android x86_64 the long double is IEEEquad
(compared to x87DoubleExtended on linux x86_64). This CL creates new
TargetInfo classes for this targets to represent these differences.
Differential revision: http://reviews.llvm.org/D8357
llvm-svn: 233177
When we try to find the end loc for a token, we have to re-lex the
token. This was running into a problem when we'd store the end loc of
a macro's coverage region, since we wouldn't actually be at the
beginning of a token when we tried to re-lex it, leading us to do
silly things (and eventually assert) when whitespace or comments
followed.
This pushes our use of getPreciseTokenLocEnd earlier, so that we won't
call it when it doesn't make sense to. It also removes an unnecessary
adjustment by 1 that was working around this problem in some cases.
llvm-svn: 233169
This fixes my stress tests non-determinism so far. However, I've not
started playing with templates, friends, or terrible macros. I've found
at least two more seeming instabilities and am just waiting for a test
case to actually trigger them.
llvm-svn: 233162
lookup tables, we need to establish a stable ordering for constructing
the hash table. This is trickier than it might seem.
Most of these cases are easily handled by sorting the lookup results
associated with a specific name that has an identifier. However for
constructors and conversion functions, the story is more complicated.
Here we need to merge all of the constructors or conversion functions
together and this merge needs to be stable. We don't have any stable
ordering for either constructors or conversion functions as both would
require a stable ordering across types.
Instead, when we have constructors or conversion functions in the
results, we reconstruct a stable order by walking the decl context in
lexical order and merging them in the order their particular declaration
names are encountered. This doesn't generalize as there might be found
declaration names which don't actually occur within the lexical context,
but for constructors and conversion functions it is safe. It does
require loading the entire decl context if necessary to establish the
ordering but there doesn't seem to be a meaningful way around that.
Many thanks to Richard for talking through all of the design choices
here. While I wrote the code, he guided all the actual decisions about
how to establish the order of things.
No test case yet because the test case I have doesn't pass yet -- there
are still more sources of non-determinism. However, this is complex
enough that I wanted it to go into its own commit in case it causes some
unforseen issue or needs to be reverted.
llvm-svn: 233156
We previously defaulted to long double, but it's also possible to have
a half inc/dec amount, when LangOpts NativeHalfType is set.
Currently, that's only true for OpenCL.
llvm-svn: 233135
prune it when we have disabled implicit module generation and thus are
not using any cached modules.
Also update a test of explicitly generated modules to pass this CC1 flag
correctly.
This fixes an issue where Clang was dropping files into the source tree
while running its tests.
llvm-svn: 233117
There are two aspects of non-determinism fixed here, which was the
minimum required to cause at least an empty module to be deterministic.
First, the random number signature is only inserted into the module when
we are building modules implicitly. The use case for these random
signatures is to work around the very fact that modules are not
deterministic in their output when working with the implicitly built and
populated module cache. Eventually this should go away entirely when
we're confident that Clang is producing deterministic output.
Second, the on-disk hash table is populated based on the order of
iteration over a DenseMap. Instead, use a MapVector so that we can walk
it in insertion order.
I've added a test that an empty module, when built twice, produces the
same binary PCM file.
llvm-svn: 233115
I'm about to commit a patch for:
http://reviews.llvm.org/D8567
That patch will break this one existing test case in Clang.
I'm not sure if this file is intending to create a Clang
dependency on the LLVM IR optimizer, but that's the
consequence of specifying -O3 on this test file.
My hope is to avoid buildbot rage by removing this check,
committing the LLVM patch, and then fixing this check.
I don't know how to make a simultaneous commit to Clang
and LLVM.
I will commit the correct CHECK line fix for this test
shortly.
llvm-svn: 233109
Richard Smith