Mehdi Amini [Mon, 7 Nov 2016 20:00:47 +0000 (20:00 +0000)]
Add some facilities to work with a git monorepo (experimental setup)
Add a new script in llvm/utils/git-svn/. When present in the $PATH,
it enables a `git llvm` command. It is providing at this
point only the ability to push from the git monorepo: `git llvm push`.
It is intended to evolves with more features, for instance I plan on
features like `git llvm show r284955` to help working with sequential
revision numbers.
The push feature is taken from Justin Lebar's script available here:
https://github.com/jlebar/llvm-repo-tools/
Kuba Brecka [Mon, 7 Nov 2016 19:09:56 +0000 (19:09 +0000)]
[tsan] Cast floating-point types correctly when instrumenting atomic accesses, LLVM part
Although rare, atomic accesses to floating-point types seem to be valid, i.e. `%a = load atomic float ...`. The TSan instrumentation pass however tries to emit inttoptr, which is incorrect, we should use a bitcast here. Anyway, IRBuilder already has a convenient helper function for this.
Matt Arsenault [Mon, 7 Nov 2016 19:09:27 +0000 (19:09 +0000)]
AMDGPU: Preserve vcc undef flags when inverting branch
If the branch was on a read-undef of vcc, passes that used
analyzeBranch to invert the branch condition wouldn't preserve
the undef flag resulting in a verifier error.
Fixes verifier failures in a future commit.
Also fix verifier error when inserting copy for vccz
corruption bug.
Benjamin Kramer [Mon, 7 Nov 2016 17:47:28 +0000 (17:47 +0000)]
[MemCpyOpt] Don't emit IR in an unspecified order
Argument evaluation order is one of the edge cases where Clang differs
from GCC, yielding different IR depending on which compiler LLVM was
built with. Make the order deterministic and tune the test to actually
verify the order instead of trying to hide it.
Mehdi Amini [Mon, 7 Nov 2016 17:40:28 +0000 (17:40 +0000)]
Add some facilities to work with a git monorepo (experimental setup)
Summary:
Some changes are made to cmake, especially the addition of a new
LLVM_ENABLE_PROJECTS option that makes the build system aware of
the monorepo directory structure.
Also a new script is added in llvm/utils/git-svn/. When present in
the $PATH, it enables a `git llvm` command. It is providing at this
point only the ability to push from the git monorepo: `git llvm push`.
It is intended to evolves with more features, for instance I plan on
features like `git llvm show r284955` to help working with sequential
revision numbers.
The push feature is taken from Justin Lebar's script available here:
https://github.com/jlebar/llvm-repo-tools/
Sanjay Patel [Mon, 7 Nov 2016 15:52:45 +0000 (15:52 +0000)]
[InstCombine] allow splat vector folds in adjustMinMax() (retry r285732)
This was reverted at r285866 because there was a crash handling a scalar
select of vectors. I added a check for that pattern and a test case based
on the example provided in the post-commit thread for r285732.
Jonas Paulsson [Mon, 7 Nov 2016 15:45:06 +0000 (15:45 +0000)]
[SystemZ] Correct the SchedModel regarding vector unit / instructions.
* Use a generic vector unit to model the issue unit more accurately.
* Update some vector instructions that actually use the vector unit for more
than one cycle.
James Molloy [Mon, 7 Nov 2016 13:38:21 +0000 (13:38 +0000)]
[Thumb1] Move padding earlier when synthesizing TBBs off of the PC
When the base register (register pointing to the jump table) is the PC, we expect the jump table to directly follow the jump sequence with no intervening padding.
If there is intervening padding, the calculated offsets will not be correct. One solution would be to account for any padding in the emitted LDRB instruction, but at the moment we don't support emitting MCExprs for the load offset.
In the meantime, it's correct and only a slight amount worse to just move the padding up, from just before the jump table to just before the jump instruction sequence. We can do that by emitting code alignment before the jump sequence, as we know the number of instructions in the sequence is always 4.
Craig Topper [Mon, 7 Nov 2016 00:13:46 +0000 (00:13 +0000)]
[X86] Remove GCCBuiltins from cvtsi2ss/cvtsi2sd/cvtss2sd intrinsics as they aren't used by clang. Add TODOs to remove these and some other unused intrinsics.
This handles the last case of the builtin function calls that we would
generate code which differed from Microsoft's ABI. Rather than
generating a call to `__pow{d,s}i2` we now promote the parameter to a
float or double and invoke `powf` or `pow` instead.
The clr/set/toggle-bit instructions (with the bit index given as an
immediate operand) had both, custom selection code that generated them,
and selection patterns at the same time. The selection patterns were
not used, because the custom selection code was executed first.
This patch removes the custom code in favor of the selection patterns.
The custom code handled 64-bit registers as well with an immediate bit
index, and so new patterns were added to implement that.
It was also the same case for the instruction "Rd += asr(Rs, Rt)",
except that the custom code did not offer any additional functionality,
and was simply removed.
Amaury Sechet [Sun, 6 Nov 2016 07:48:46 +0000 (07:48 +0000)]
Kill deprecated attribute API
Summary:
This kill various depreacated API related to attribute :
- The deprecated C API attribute based on LLVMAttribute enum.
- The Raw attribute set format (planned to be removed in 4.0).
Tim Shen [Sun, 6 Nov 2016 07:38:37 +0000 (07:38 +0000)]
[APFloat] Make functions that produce APFloaat objects use correct semantics.
Summary:
Fixes PR30869.
In D25977 I meant to change all functions that care about lifetime. I
changed constructors, factory functions, but I missed member/free
functions that return new instances. This patch changes them.
Craig Topper [Sun, 6 Nov 2016 02:03:58 +0000 (02:03 +0000)]
[AVX-512] Add -show-mc-encoding to legacy vector intrinsic tests so we can see when VEX or EVEX encoded instructions are being emitted. Make sure the tests all have an avx2 command line and an skx command line.
Justin Lebar [Sat, 5 Nov 2016 16:47:25 +0000 (16:47 +0000)]
[LoopStrengthReduce] Don't use a DenseSet<int64_t> when we might add any valid int64_t to the set.
Summary:
SmallSetVector uses DenseSet, but that means we need to reserve some
values for the empty and tombstone keys.
It seems to me we should have a general way to let us store full-range
ints inside of DenseSets, and furthermore that we probably shouldn't
silently let you add ints into DenseSets without explicitly promising
that they're in range. But that's a battle for another day; for now,
just fix this code, since we currently do something Very Bad when
compiling ffmpeg.
Craig Topper [Sat, 5 Nov 2016 05:35:23 +0000 (05:35 +0000)]
[AVX-512] Use an equality compare instead of StringRef::startswith in a few places in auto upgrade that were looking for the complete intrinsic name anyway.
Craig Topper [Sat, 5 Nov 2016 04:00:31 +0000 (04:00 +0000)]
[X86] Remove broken support for autoupgrading llvm.x86.fma4.* intrinsics to llvm.x86.fma.*.
It currently fires an assert if you even try. Looking back, I don't think it ever worked because it only changed the name of the function object, but not the intrinsic ID stored in it. Given that, I think it can be removed since no one has noticed or complained in the past 4 years.
Add a missing return to the move assignment operator for
SequenceNumberManager.
Sadly, we don't have any unittests for this class because it is
a private class. Since it seems to have a nice isolated and testable
interface, it'd be great to extract it to a detail namespace and write
unit tests for it as then we could catch issues. I'll probably pester
Lang about that or some alternative refactoring.
Remove dead code trying to handle when the amount of data read is
insufficient to populate the expected struct. Prior to this we already
bailed out of the routine when this situation comes up, so none of this
code had any effect.
If someone wants to bring it back to handle these cases, fixing the
earlier conditions and adding the necessary test cases that actually
exercises it, they can always revert this and go from there.
Both of these were noticed by PVS-Studio due to the identical (dead)
condition.
Only log the visit of a return instruction if we in fact found a return
instruction.
This avoids dereferencing null in the debug logging if the instruction
was not in fact a return instruction. This potential bug was found by
PVS-Studio.
This actually fixes the last of the "dereferenced a pointer before
checking it for null" reports in the recent PVS-Studio run. However,
there are quite a few reports of this nature that I did not do anything
to fix because they are pretty glaring false positives. They usually
took the form of quite clear correlated checks or a check made in
a separate function. I've even added asserts anywhere this correlation
wasn't pretty obvious and fundamental to the code.
Hoist check for TLI above all of the attempts to use it (including one
of which that is hidden inside a separate function call) and helpfully
before building expensive transaction infrastructure. This will avoid
crashing when running CGP in a generic mode if we ever managed to hit
this case.
Note that I spent some time looking at alternatives. CGP is actually
used without a TM or TLI in order to do some target-independent testing.
Further, all of the neighboring optimization techniques actually have
some paths that are effective even in the absence of TLI so this seemed
the correct scope at which to check and bypass logic. It still isn't
clear that long-term support for missing TM/TLI is the right
cost/benefit tradeoff for CGP -- we seem to get relatively little for it
and the code is just littered with checks (and assumptions which
I suspect are still missing some checks).
This at least fixes the potential bug in this code spotted by
PVS-Studio, so we've got that going for us. ;]
Brian Gesiak [Thu, 3 Nov 2016 23:41:49 +0000 (23:41 +0000)]
[lit] Remove TODO
Summary:
Instead of keeping track of TODOs for lit in a file checked into source
control, use LLVM's bug tracker. The TODOs have been migrated to the
following bugs:
Sink all of the code relying on the MachO MachineModuleInfo to live
behind the test that the MachineModuleInfo analysis was
actually available and can be used.
While the MachO bits may well be reasonable to assume in the darwin
assembly printer, the analysis isn't constructively guaranteed anywhere
I could find so it seems safest to avoid crashing here.
This issue was found with PVS-Studio. Pretty sure the Clang Static
Anaylzer flags similar issues but we've probably never pointed it at
this code effectively.
Weiming Zhao [Thu, 3 Nov 2016 21:49:08 +0000 (21:49 +0000)]
[Cortex-M0] Atomic lowering
Summary: ARMv6m supports dmb etc fench instructions but not ldrex/strex etc. So for some atomic load/store, LLVM should inline instructions instead of lowering to __sync_ calls.
projects / llvm / log