OSDN Git Service

android-x86/external-llvm.git
7 years agoAdd an index for Module Metadata record in the bitcode
Mehdi Amini [Wed, 28 Dec 2016 22:30:28 +0000 (22:30 +0000)]
Add an index for Module Metadata record in the bitcode

This index record the position for each metadata record in
the bitcode, so that the reader will be able to lazy-load
on demand each individual record.

We also make sure that every abbrev is emitted upfront so
that the block can be skipped while reading.

I don't plan to commit this before having the reader
counterpart, but I figured this can be reviewed mostly
independently.

Recommit r290684 (was reverted in r290686 because a test
was broken) after adding a threshold to avoid emitting
the index when unnecessary (little amount of metadata).
This optimization "hides" a limitation of the ability
to backpatch in the bitstream: we can only backpatch
safely when the position has been flushed. So if we emit
an index for one metadata, it is possible that (part of)
the offset placeholder hasn't been flushed and the backpatch
will fail.

Differential Revision: https://reviews.llvm.org/D28083

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290690 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "Add an index for Module Metadata record in the bitcode"
Saleem Abdulrasool [Wed, 28 Dec 2016 20:37:22 +0000 (20:37 +0000)]
Revert "Add an index for Module Metadata record in the bitcode"

This reverts commit a0ca6ae2d38339e4ede0dfa588086fc23d87e836.  Revert at
Mehdi's request as it is breaking bots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290686 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NewGVN] replace emplace_back with push_back
Piotr Padlewski [Wed, 28 Dec 2016 20:36:08 +0000 (20:36 +0000)]
[NewGVN] replace emplace_back with push_back

emplace_back is not faster if it is equivalent to push_back. In this cases emplaced value had the
same type that the one stored in container. It is ugly and it might be even slower (see
Scott Meyers presentation about emplacement).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290685 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd an index for Module Metadata record in the bitcode
Mehdi Amini [Wed, 28 Dec 2016 19:44:19 +0000 (19:44 +0000)]
Add an index for Module Metadata record in the bitcode

Summary:
This index record the position for each metadata record in
the bitcode, so that the reader will be able to lazy-load
on demand each individual record.

We also make sure that every abbrev is emitted upfront so
that the block can be skipped while reading.

I don't plan to commit this before having the reader
counterpart, but I figured this can be reviewed mostly
independently.

Reviewers: pcc, tejohnson

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D28083

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290684 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NewGVN] Simplyfy loop NFC
Piotr Padlewski [Wed, 28 Dec 2016 19:42:49 +0000 (19:42 +0000)]
[NewGVN] Simplyfy loop NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290683 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ThinLTO] Honor -O{0,1,2,4} passed through the libLTO interface for ThinLTO
Mehdi Amini [Wed, 28 Dec 2016 19:37:16 +0000 (19:37 +0000)]
[ThinLTO] Honor -O{0,1,2,4} passed through the libLTO interface for ThinLTO

This was hardcoded to be O3 till now, without any way to change it
without changing the code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290682 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NewGVN] replace typedefs with usings
Piotr Padlewski [Wed, 28 Dec 2016 19:29:26 +0000 (19:29 +0000)]
[NewGVN] replace typedefs with usings

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290680 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NewGVN] NFC fixes
Piotr Padlewski [Wed, 28 Dec 2016 19:17:17 +0000 (19:17 +0000)]
[NewGVN] NFC fixes

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290679 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[WinEH] Don't assume endFunction is called while in .text
Reid Kleckner [Wed, 28 Dec 2016 19:05:12 +0000 (19:05 +0000)]
[WinEH] Don't assume endFunction is called while in .text

Jump table emission can switch to .rdata before
WinException::endFunction gets called. Just remember the appropriate
text section we started in and reset back to it when we end the
function. We were already switching sections back from .xdata anyway.

Fixes the first problem in PR31488, so that now COFF switch tables can
live in .rdata if we want them to.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290678 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NewGVN] Global sweep replacing NULL with nullptr. NFCI.
Davide Italiano [Wed, 28 Dec 2016 14:00:11 +0000 (14:00 +0000)]
[NewGVN] Global sweep replacing NULL with nullptr. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290670 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NewGVN] Remove redundant code. NFCI.
Davide Italiano [Wed, 28 Dec 2016 13:54:16 +0000 (13:54 +0000)]
[NewGVN] Remove redundant code. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290669 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NewGVN] equals() for loads/stores is the same. Unify.
Davide Italiano [Wed, 28 Dec 2016 13:37:17 +0000 (13:37 +0000)]
[NewGVN] equals() for loads/stores is the same. Unify.

Differential Revision:  https://reviews.llvm.org/D28116

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290667 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Introduce a devirtualization iteration layer for the new PM.
Chandler Carruth [Wed, 28 Dec 2016 11:07:33 +0000 (11:07 +0000)]
[PM] Introduce a devirtualization iteration layer for the new PM.

This is an orthogonal and separated layer instead of being embedded
inside the pass manager. While it adds a small amount of complexity, it
is fairly minimal and the composability and control seems worth the
cost.

The logic for this ends up being nicely isolated and targeted. It should
be easy to experiment with different iteration strategies wrapped around
the CGSCC bottom-up walk using this kind of facility.

The mechanism used to track devirtualization is the simplest one I came
up with. I think it handles most of the cases the existing iteration
machinery handles, but I haven't done a *very* in depth analysis. It
does however match the basic intended semantics, and we can tweak or
tune its exact behavior incrementally as necessary. One thing that we
may want to revisit is freshly building the value handle set on each
iteration. While I don't think this will be a significant cost (it is
strictly fewer value handles but more churn of value handes than the old
call graph), it is conceivable that we'll want a somewhat more clever
tracking mechanism. My hope is to layer that on as a follow up patch
with data supporting any implementation complexity it adds.

This code also provides for a basic count heuristic: if the number of
indirect calls decreases and the number of direct calls increases for
a given function in the SCC, we assume devirtualization is responsible.
This matches the heuristics currently used in the legacy pass manager.

Differential Revision: https://reviews.llvm.org/D23114

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290665 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Teach the CGSCC's CG update utility to more carefully invalidate
Chandler Carruth [Wed, 28 Dec 2016 10:34:50 +0000 (10:34 +0000)]
[PM] Teach the CGSCC's CG update utility to more carefully invalidate
analyses when we're about to break apart an SCC.

We can't wait until after breaking apart the SCC to invalidate things:
1) Which SCC do we then invalidate? All of them?
2) Even if we invalidate all of them, a newly created SCC may not have
   a proxy that will convey the invalidation to functions!

Previously we only invalidated one of the SCCs and too late. This led to
stale analyses remaining in the cache. And because the caching strategy
actually works, they would get used and chaos would ensue.

Doing invalidation early is somewhat pessimizing though if we *know*
that the SCC structure won't change. So it turns out that the design to
make the mutation API force the caller to know the *kind* of mutation in
advance was indeed 100% correct and we didn't do enough of it. So this
change also splits two cases of switching a call edge to a ref edge into
two separate APIs so that callers can clearly test for this and take the
easy path without invalidating when appropriate. This is particularly
important in this case as we expect most inlines to be between functions
in separate SCCs and so the common case is that we don't have to so
aggressively invalidate analyses.

The LCG API change in turn needed some basic cleanups and better testing
in its unittest. No interesting functionality changed there other than
more coverage of the returned sequence of SCCs.

While this seems like an obvious improvement over the current state, I'd
like to revisit the core concept of invalidating within the CG-update
layer at all. I'm wondering if we would be better served forcing the
callers to handle the invalidation beforehand in the cases that they
can handle it. An interesting example is when we want to teach the
inliner to *update and preserve* analyses. But we can cross that bridge
when we get there.

With this patch, the new pass manager an build all of the LLVM test
suite at -O3 and everything passes. =D I haven't bootstrapped yet and
I'm sure there are still plenty of bugs, but this gives a nice baseline
so I'm going to increasingly focus on fleshing out the missing
functionality, especially the bits that are just turned off right now in
order to let us establish this baseline.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290664 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoThis is a large patch for X86 AVX-512 of an optimization for reducing code size by...
Gadi Haber [Wed, 28 Dec 2016 10:12:48 +0000 (10:12 +0000)]
This is a large patch for X86 AVX-512 of an optimization for reducing code size by encoding EVEX AVX-512 instructions using the shorter VEX encoding when possible.

There are cases of AVX-512 instructions that have two possible encodings. This is the case with instructions that use vector registers with low indexes of 0 - 15 and do not use the zmm registers or the mask k registers.
The EVEX encoding prefix requires 4 bytes whereas the VEX prefix can take only up to 3 bytes. Consequently, using the VEX encoding for these instructions results in a code size reduction of ~2 bytes even though it is compiled with the AVX-512 features enabled.

Reviewers: Craig Topper, Zvi Rackoover, Elena Demikhovsky
Differential Revision: https://reviews.llvm.org/D27901

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290663 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Teach the inliner's call graph update to handle inserting new edges
Chandler Carruth [Wed, 28 Dec 2016 03:13:12 +0000 (03:13 +0000)]
[PM] Teach the inliner's call graph update to handle inserting new edges
when they are call edges at the leaf but may (transitively) be reached
via ref edges.

It turns out there is a simple rule: insert everything as a ref edge
which is a safe conservative default. Then we let the existing update
logic handle promoting some of those to call edges.

Note that it would be fairly cheap to make these call edges right away
if that is desirable by testing whether there is some existing call path
from the source to the target. It just seemed like slightly more
complexity in this code path that isn't strictly necessary. If anyone
feels strongly about handling this differently I'm happy to change it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290649 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine] Remove a piece of a comment that said that InstCombiner contains pass...
Craig Topper [Wed, 28 Dec 2016 03:12:42 +0000 (03:12 +0000)]
[InstCombine] Remove a piece of a comment that said that InstCombiner contains pass infrastructure. That hasn't been true since r226618. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290648 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Actually commit the test update that was supposed to accompany
Chandler Carruth [Wed, 28 Dec 2016 02:31:24 +0000 (02:31 +0000)]
[PM] Actually commit the test update that was supposed to accompany
r290644. Sorry for this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290646 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LCG] Teach the ref edge removal to handle a ref edge that is trivial
Chandler Carruth [Wed, 28 Dec 2016 02:24:58 +0000 (02:24 +0000)]
[LCG] Teach the ref edge removal to handle a ref edge that is trivial
due to a call cycle.

This actually crashed the ref removal before.

I've added a unittest that covers this kind of interesting graph
structure and mutation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290645 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Disable the loop vectorizer from the new PM's pipeline as it
Chandler Carruth [Wed, 28 Dec 2016 02:24:55 +0000 (02:24 +0000)]
[PM] Disable the loop vectorizer from the new PM's pipeline as it
currenty relies on the old PM's dependency system forming LCSSA.

The new PM will require a different design for this, and for now this is
causing most of the issues I'm currently seeing in testing. I'd like to
get to a testable baseline and then work on re-enabling things one at
a time.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290644 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine] Canonicalize insert splat sequences into an insert + shuffle
Michael Kuperstein [Wed, 28 Dec 2016 00:18:08 +0000 (00:18 +0000)]
[InstCombine] Canonicalize insert splat sequences into an insert + shuffle

This adds a combine that canonicalizes a chain of inserts which broadcasts
a value into a single insert + a splat shufflevector.

This fixes PR31286.

Differential Revision: https://reviews.llvm.org/D27992

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290641 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[libFuzzer] add an experimental flag -experimental_len_control=1 that sets max_len...
Kostya Serebryany [Tue, 27 Dec 2016 23:24:55 +0000 (23:24 +0000)]
[libFuzzer] add an experimental flag -experimental_len_control=1 that sets max_len to 1M and tries to increases the actual max sizes of mutations very gradually (second attempt)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290637 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMark comparator call operator as const
Eric Fiselier [Tue, 27 Dec 2016 23:15:58 +0000 (23:15 +0000)]
Mark comparator call operator as const

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290636 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[libFuzzer] don't create large random mutations when given an empty seed
Kostya Serebryany [Tue, 27 Dec 2016 22:15:04 +0000 (22:15 +0000)]
[libFuzzer] don't create large random mutations when given an empty seed

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290634 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[sanitizer-coverage] sort the switch cases
Kostya Serebryany [Tue, 27 Dec 2016 21:20:06 +0000 (21:20 +0000)]
[sanitizer-coverage] sort the switch cases

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290628 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agollvm-readobj: ELF: Make DT tags machine aware
Hemant Kulkarni [Tue, 27 Dec 2016 19:59:29 +0000 (19:59 +0000)]
llvm-readobj: ELF: Make DT tags machine aware

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290623 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[libFuzzer] fix UB and simplify the computation of the RNG seed (https://llvm.org...
Kostya Serebryany [Tue, 27 Dec 2016 19:51:34 +0000 (19:51 +0000)]
[libFuzzer] fix UB and simplify the computation of the RNG seed (https://llvm.org/bugs/show_bug.cgi?id=31456)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290622 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Teach MemDep to invalidate its result object when its cached
Chandler Carruth [Tue, 27 Dec 2016 19:33:04 +0000 (19:33 +0000)]
[PM] Teach MemDep to invalidate its result object when its cached
analysis handles become invalid.

Add a test case for its invalidation logic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290620 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDebugInfo: add explicit casts for -Wqual-cast
Saleem Abdulrasool [Tue, 27 Dec 2016 18:35:24 +0000 (18:35 +0000)]
DebugInfo: add explicit casts for -Wqual-cast

Fix a warning detected by gcc 6:
  warning: cast from type 'const void*' to type 'uint8_t* {aka unsigned char*}' casts away qualifiers [-Wcast-qual]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290618 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoASMParser: use range-based for loops (NFC)
Saleem Abdulrasool [Tue, 27 Dec 2016 18:35:22 +0000 (18:35 +0000)]
ASMParser: use range-based for loops (NFC)

Convert the verify method to use a few more range based for loops,
converting to const iterators in the process.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290617 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agotest: modernise ARM CodeGen tests
Saleem Abdulrasool [Tue, 27 Dec 2016 18:35:19 +0000 (18:35 +0000)]
test: modernise ARM CodeGen tests

Replace the use of grep with FileCheck.  Tidy up some of the tests.  A
few of the tests have been left as weak as previously, though some have
been made more stringent.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290616 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NewGVN] Simplify a bit removing else after return. NFCI.
Davide Italiano [Tue, 27 Dec 2016 18:15:39 +0000 (18:15 +0000)]
[NewGVN] Simplify a bit removing else after return. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290615 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Remove a pointless optimization.
Chandler Carruth [Tue, 27 Dec 2016 18:04:11 +0000 (18:04 +0000)]
[PM] Remove a pointless optimization.

There is no need to do this within an analysis. That method shouldn't
even be reached if this predicate holds as the actual useful
optimization is in the analysis manager itself.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290614 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAttempt to make the Windows bots green after r290609.
Chad Rosier [Tue, 27 Dec 2016 18:02:27 +0000 (18:02 +0000)]
Attempt to make the Windows bots green after r290609.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290613 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Add more dedicated testing to cover the invalidation logic added to
Chandler Carruth [Tue, 27 Dec 2016 17:59:22 +0000 (17:59 +0000)]
[PM] Add more dedicated testing to cover the invalidation logic added to
BasicAA in r290603.

I've kept the basic testing in the new PM test file as that also covers
the AAManager invalidation logic. If/when there is a good place for
broader AA testing it could move there.

This test is somewhat unsatisfying as I can't get it to fail even with
ASan outside of explicit checks of the invalidation. Apparently we don't
yet have any test coverage of the BasicAA code paths using either the
domtree or loopinfo -- I made both of them always be null and check-llvm
passed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290612 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[MemCpyOpt] Don't sink LoadInst below possible clobber.
Bryant Wong [Tue, 27 Dec 2016 17:58:12 +0000 (17:58 +0000)]
[MemCpyOpt] Don't sink LoadInst below possible clobber.

Differential Revision: https://reviews.llvm.org/D26811

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290611 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ThinLTO] Fix "||" vs "|" mixup.
Teresa Johnson [Tue, 27 Dec 2016 17:45:09 +0000 (17:45 +0000)]
[ThinLTO] Fix "||" vs "|" mixup.

The effect of the bug was that we would incorrectly create summaries
for global and weak values defined in module asm (since we were
essentially testing for bit 1 which is SF_Undefined, and the
RecordStreamer ignores local undefined references). This would have
resulted in conservatively disabling importing of anything referencing
globals and weaks defined in module asm. Added these cases to the test
which now fails without this bug fix.

Fixes PR31459.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290610 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AArch64][AsmParser] Add support for parsing shift/extend operands with symbols.
Chad Rosier [Tue, 27 Dec 2016 16:58:09 +0000 (16:58 +0000)]
[AArch64][AsmParser] Add support for parsing shift/extend operands with symbols.

Differential Revision: https://reviews.llvm.org/D27953

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290609 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AMDGPU][llvm-mc] Predefined symbols to access register counts (.kernel.{v|s}gpr_count)
Artem Tamazov [Tue, 27 Dec 2016 16:00:11 +0000 (16:00 +0000)]
[AMDGPU][llvm-mc] Predefined symbols to access register counts (.kernel.{v|s}gpr_count)

The feature allows for conditional assembly, filling the entries
of .amd_kernel_code_t etc.

Symbols are defined with value 0 at the beginning of each kernel scope.
After each register usage, the respective symbol is set to:
value = max( value, ( register index + 1 ) )
Thus, at the end of scope the value represents a count of used registers.

Kernel scopes begin at .amdgpu_hsa_kernel directive, end at the
next .amdgpu_hsa_kernel (or EOF, whichever comes first). There is also
dummy scope that lies from the beginning of source file til the
first .amdgpu_hsa_kernel.

Test added.

Differential Revision: https://reviews.llvm.org/D27859

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290608 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[MemDep] Operand visited twice bugfix
Piotr Padlewski [Tue, 27 Dec 2016 15:06:07 +0000 (15:06 +0000)]
[MemDep] Operand visited twice bugfix

Because operand was not marked as seen it was visited twice.
It doesn't change behavior of optimization, it just saves redudant
visit, so no test changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290607 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRuntimeDyldELF: refactor AArch64 relocations. NFC.
Eugene Leviant [Tue, 27 Dec 2016 13:33:32 +0000 (13:33 +0000)]
RuntimeDyldELF: refactor AArch64 relocations. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290606 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix unit test in NDEBUG build
Eugene Leviant [Tue, 27 Dec 2016 11:07:53 +0000 (11:07 +0000)]
Fix unit test in NDEBUG build

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290604 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Teach BasicAA how to invalidate its result object.
Chandler Carruth [Tue, 27 Dec 2016 10:30:45 +0000 (10:30 +0000)]
[PM] Teach BasicAA how to invalidate its result object.

This requires custom handling because BasicAA caches handles to other
analyses and so it needs to trigger indirect invalidation.

This fixes one of the common crashes when using the new PM in real
pipelines. I've also tweaked a regression test to check that we are at
least handling the most immediate case.

I'm going to work at re-structuring this test some to both scale better
(rather than all being in one file) and check more invalidation paths in
a follow-up commit, but I wanted to get the basic bug fix in place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290603 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAttempt to fix build bot after r290597
Eugene Leviant [Tue, 27 Dec 2016 10:24:58 +0000 (10:24 +0000)]
Attempt to fix build bot after r290597

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290602 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Disable more of the loop passes -- LCSSA and LoopSimplify are also
Chandler Carruth [Tue, 27 Dec 2016 10:16:46 +0000 (10:16 +0000)]
[PM] Disable more of the loop passes -- LCSSA and LoopSimplify are also
not really wired into the loop pass manager in a way that will let us
productively use these passes yet.

This lets the new PM get farther in basic testing which is useful for
establishing a good baseline of "doesn't explode". There are still
plenty of crashers in basic testing though, this just gets rid of some
noise that is well understood and not representing a specific or narrow
bug.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290601 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AMDGPU] Assembler: support SDWA and DPP for VOP2b instructions
Sam Kolton [Tue, 27 Dec 2016 10:06:42 +0000 (10:06 +0000)]
[AMDGPU] Assembler: support SDWA and DPP for VOP2b instructions

Reviewers: nhaustov, artem.tamazov, vpykhtin, tstellarAMD

Subscribers: arsenm, kzhuravl, wdng, nhaehnle, yaxunl, tony-tye

Differential Revision: https://reviews.llvm.org/D28051

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290599 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRuntimeDyldELF: add R_AARCH64_ADD_ABS_LO12_NC reloc
Eugene Leviant [Tue, 27 Dec 2016 09:51:38 +0000 (09:51 +0000)]
RuntimeDyldELF: add R_AARCH64_ADD_ABS_LO12_NC reloc

Differential revision: https://reviews.llvm.org/D28115

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290598 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAllow setting multiple debug types
Eugene Leviant [Tue, 27 Dec 2016 09:31:20 +0000 (09:31 +0000)]
Allow setting multiple debug types

Differential revision: https://reviews.llvm.org/D28109

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290597 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoChange a std::vector to SmallVector in NewGVN
Daniel Berlin [Tue, 27 Dec 2016 09:20:36 +0000 (09:20 +0000)]
Change a std::vector to SmallVector in NewGVN

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290596 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Teach the AAManager and AAResults layer (the worst offender for
Chandler Carruth [Tue, 27 Dec 2016 08:44:39 +0000 (08:44 +0000)]
[PM] Teach the AAManager and AAResults layer (the worst offender for
inter-analysis dependencies) to use the new invalidation infrastructure.

This teaches it to invalidate itself when any of the peer function
AA results that it uses become invalid. We do this by just tracking the
originating IDs. I've kept it in a somewhat clunky API since some users
of AAResults are outside the new PM right now. We can clean this API up
if/when those users go away.

Secondly, it uses the registration on the outer analysis manager proxy
to trigger deferred invalidation when a module analysis result becomes
invalid.

I've included test cases that specifically try to trigger use-after-free
in both of these cases and they would crash or hang pretty horribly for
me even without ASan. Now they work nicely.

The `InvalidateAnalysis` utility pass required some tweaking to be
useful in this context and it still is pretty garbage. I'd like to
switch it back to the previous implementation and teach the explicit
invalidate method on the AnalysisManager to take care of correctly
triggering indirect invalidation, but I wanted to go ahead and send this
out so folks could see how all of this stuff works together in practice.
And, you know, that it does actually work. =]

Differential Revision: https://reviews.llvm.org/D27205

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290595 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Introduce the facilities for registering cross-IR-unit dependencies
Chandler Carruth [Tue, 27 Dec 2016 08:40:39 +0000 (08:40 +0000)]
[PM] Introduce the facilities for registering cross-IR-unit dependencies
that require deferred invalidation.

This handles the other real-world invalidation scenario that we have
cases of: a function analysis which caches references to a module
analysis. We currently do this in the AA aggregation layer and might
well do this in other places as well.

Since this is relative rare, the technique is somewhat more cumbersome.
Analyses need to register themselves when accessing the outer analysis
manager's proxy. This proxy is already necessarily present to allow
access to the outer IR unit's analyses. By registering here we can track
and trigger invalidation when that outer analysis goes away.

To make this work we need to enhance the PreservedAnalyses
infrastructure to support a (slightly) more explicit model for "sets" of
analyses, and allow abandoning a single specific analyses even when
a set covering that analysis is preserved. That allows us to describe
the scenario of preserving all Function analyses *except* for the one
where deferred invalidation has triggered.

We also need to teach the invalidator API to support direct ID calls
instead of always going through a template to dispatch so that we can
just record the ID mapping.

I've introduced testing of all of this both for simple module<->function
cases as well as for more complex cases involving a CGSCC layer.

Much like the previous patch I've not tried to fully update the loop
pass management layer because that layer is due to be heavily reworked
to use similar techniques to the CGSCC to handle updates. As that
happens, we'll have a better testing basis for adding support like this.

Many thanks to both Justin and Sean for the extensive reviews on this to
help bring the API design and documentation into a better state.

Differential Revision: https://reviews.llvm.org/D27198

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290594 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Turn on the new PM's inliner in addition to the current one for
Chandler Carruth [Tue, 27 Dec 2016 07:18:43 +0000 (07:18 +0000)]
[PM] Turn on the new PM's inliner in addition to the current one for
most of the inliner test cases.

The inliner involves a bunch of interesting code and tends to be where
most of the issues I've seen experimenting with the new PM lie. All of
these test cases pass, but I'd like to keep some more thorough coverage
here so doing a fairly blanket enabling.

There are a handful of interesting tests I've not enabled yet because
they're focused on the always inliner, or on functionality that doesn't
(yet) exist in the inliner.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290592 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512] Add all forms of VPALIGNR, VALIGND, and VALIGNQ to the load folding tables.
Craig Topper [Tue, 27 Dec 2016 06:51:09 +0000 (06:51 +0000)]
[AVX-512] Add all forms of VPALIGNR, VALIGND, and VALIGNQ to the load folding tables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290591 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Add one of the features left out of the initial inliner patch:
Chandler Carruth [Tue, 27 Dec 2016 06:46:20 +0000 (06:46 +0000)]
[PM] Add one of the features left out of the initial inliner patch:
skipping indirectly recursive inline chains.

To do this, we implicitly build an inline stack for each callsite and
check prior to inlining that doing so would not form a cycle. This uses
the exact same technique and even shares some code with the legacy PM
inliner.

This solution remains deeply unsatisfying to me because it means we
cannot actually iterate the inliner externally. Doing so would not be
able to easily detect and avoid such cycles. Some day I would very much
like to have a solution that works without this internal state to detect
cycles, but this is not that day.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290590 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Wire up another test to the new pass manager.
Chandler Carruth [Tue, 27 Dec 2016 06:46:16 +0000 (06:46 +0000)]
[PM] Wire up another test to the new pass manager.

Nothing really interesting here, but I had to improve the test to use
variables rather than hard coding value names as we happen to end up
with different value names in the new PM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290589 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Analysis] Ignore `nobuiltin` on `allocsize` function calls.
George Burgess IV [Tue, 27 Dec 2016 06:32:14 +0000 (06:32 +0000)]
[Analysis] Ignore `nobuiltin` on `allocsize` function calls.

We currently ignore the `allocsize` attribute on functions calls with
the `nobuiltin` attribute when trying to lower `@llvm.objectsize`. We
shouldn't care about `nobuiltin` here: `allocsize` is explicitly added
by the user, not inferred based on a function's symbol.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290588 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Analysis] Refactor as promised in r290397.
George Burgess IV [Tue, 27 Dec 2016 06:10:50 +0000 (06:10 +0000)]
[Analysis] Refactor as promised in r290397.

This also makes us no longer check for `allocsize` on intrinsic calls.
This shouldn't matter, since intrinsics should provide the information
we get from `allocsize` on their own.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290585 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512] Remove masked pmuldq and pmuludq intrinsics and autoupgrade them to unmaske...
Craig Topper [Tue, 27 Dec 2016 05:30:14 +0000 (05:30 +0000)]
[AVX-512] Remove masked pmuldq and pmuludq intrinsics and autoupgrade them to unmasked intrinsics plus a select.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290583 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine][X86] Add DemandedElts support for 512-bit PMULDQ/PMULUDQ instructions
Craig Topper [Tue, 27 Dec 2016 05:30:09 +0000 (05:30 +0000)]
[InstCombine][X86] Add DemandedElts support for 512-bit PMULDQ/PMULUDQ instructions

PMULDQ/PMULUDQ vXi64 instructions only use the even numbered v2Xi32 input elements which SimplifyDemandedVectorElts should try and use.

This builds on r290554 which added supported for 128 and 256-bit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290582 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LCG] Teach the LazyCallGraph to handle visiting the blockaddress
Chandler Carruth [Tue, 27 Dec 2016 05:00:45 +0000 (05:00 +0000)]
[LCG] Teach the LazyCallGraph to handle visiting the blockaddress
constant expression and to correctly form function reference edges
through them without crashing because one of the operands (the
`BasicBlock` isn't actually a constant despite being an operand of
a constant).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290581 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512] Add 512-bit unmasked intrinsics for pmuldq and pmuludq so we can add them...
Craig Topper [Tue, 27 Dec 2016 03:46:05 +0000 (03:46 +0000)]
[AVX-512] Add 512-bit unmasked intrinsics for pmuldq and pmuludq so we can add them to InstCombine with the 128 and 256 bit versions.

The 128 and 256 bit masked intrinsics are currently unused by clang. The sse and avx2 unmasked intrinsics are used instead. The new 512-bit intrinsic will be used to do the same. Then all masked versions will removed and autoupgraded.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290573 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Teach the inliner in the new PM to merge attributes after inlining.
Chandler Carruth [Tue, 27 Dec 2016 03:39:54 +0000 (03:39 +0000)]
[PM] Teach the inliner in the new PM to merge attributes after inlining.

Also enable the new PM in the attributes test case which caught this
issue.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290572 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Inliner] Modernize all of the inliner tests that were using grep.
Chandler Carruth [Tue, 27 Dec 2016 02:47:37 +0000 (02:47 +0000)]
[Inliner] Modernize all of the inliner tests that were using grep.

This mostly involved converting from grep to FileCheck and tidying up
the IR used.

In one case (invoke_test-3.ll) the test had become completely pointless
as we use 'resume' rather than 'unwind' now, and even then it did not
occur at the end of the line.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290570 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512][InstCombine] Teach InstCombine to turn masked scalar add/sub/mul/div with...
Craig Topper [Tue, 27 Dec 2016 01:56:30 +0000 (01:56 +0000)]
[AVX-512][InstCombine] Teach InstCombine to turn masked scalar add/sub/mul/div with rounding intrinsics into normal IR operations if the rounding mode is CUR_DIRECTION.

An earlier commit added support for unmasked scalar operations. At that time isel wouldn't generate an optimal sequence for masked operations, but that has now been fixed.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290566 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine][AVX-512] Add masked scalar add/sub/mul/div intrinsic test cases that...
Craig Topper [Tue, 27 Dec 2016 01:56:27 +0000 (01:56 +0000)]
[InstCombine][AVX-512] Add masked scalar add/sub/mul/div intrinsic test cases that don't have a CUR_DIRECTION rounding mode.

The CUR_DIRECTION case will be optimized in a future commit so this provides coverage for the other cases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290565 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512] Add isel patterns to turn native masked scalar add/sub/mul/div into masked...
Craig Topper [Tue, 27 Dec 2016 01:56:24 +0000 (01:56 +0000)]
[AVX-512] Add isel patterns to turn native masked scalar add/sub/mul/div into masked instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290564 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512] Add tests to show missed opportunities for combining masking with scalar...
Craig Topper [Tue, 27 Dec 2016 01:56:22 +0000 (01:56 +0000)]
[AVX-512] Add tests to show missed opportunities for combining masking with scalar arithmetic operations.

These particular sequences will be generated after a future change to teach InstCombine to turn masked scalar arithmetic intrinsics into native IR.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290563 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Move the collection of call sites to a more appropriate place
Chandler Carruth [Tue, 27 Dec 2016 01:24:50 +0000 (01:24 +0000)]
[PM] Move the collection of call sites to a more appropriate place
inside of `InlineFunction`. Prior to this, call instructions are
specifically being rewritten and replaced within the inlined region,
invalidating some of the call sites.

Several of these regions are using the same technique to walk the
inlined region so this seems clearly safe up to this point.

I've also added a short circuit to the scan for call sites based on what
other code is doing.

With this, the most common crash I've found in the new inliner code is
fixed. I've turned it on for another test case that covers this
scenario.

I'll make my way through most of the other inliner test cases
just to get some easy coverage next.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290562 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512][InstCombine] Teach InstCombine to turn packed add/sub/mul/div with rounding...
Craig Topper [Tue, 27 Dec 2016 00:23:16 +0000 (00:23 +0000)]
[AVX-512][InstCombine] Teach InstCombine to turn packed add/sub/mul/div with rounding intrinsics into normal IR operations if the rounding mode is CUR_DIRECTION.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290559 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[PM] Teach the always inliner in the new pass manager to support
Chandler Carruth [Mon, 26 Dec 2016 23:43:27 +0000 (23:43 +0000)]
[PM] Teach the always inliner in the new pass manager to support
removing fully-dead comdats without removing dead entries in comdats
with live members.

This factors the core logic out of the current inliner's internals to
a reusable utility and leverages that in both places. The factored out
code should also be (minorly) more efficient in cases where we have very
few dead functions or dead comdats to consider.

I've added a test case to cover this behavior of the always inliner.
This is the last significant bug in the new PM's always inliner I've
found (so far).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290557 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[doc] Add mention of the difference in optimization level between Release and RelWith...
Mehdi Amini [Mon, 26 Dec 2016 23:42:12 +0000 (23:42 +0000)]
[doc] Add mention of the difference in optimization level between Release and RelWithDebInfo in Cmake.rst

This is surprising to many people.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290556 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ADT] Add an llvm::erase_if utility to make the standard erase+remove_if
Chandler Carruth [Mon, 26 Dec 2016 23:30:44 +0000 (23:30 +0000)]
[ADT] Add an llvm::erase_if utility to make the standard erase+remove_if
pattern easier to write.

Differential Revision: https://reviews.llvm.org/D28120

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290555 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine][X86] Add DemandedElts support for PMULDQ/PMULUDQ instructions
Simon Pilgrim [Mon, 26 Dec 2016 23:28:17 +0000 (23:28 +0000)]
[InstCombine][X86] Add DemandedElts support for PMULDQ/PMULUDQ instructions

PMULDQ/PMULUDQ vXi64 instructions only use the even numbered v2Xi32 input elements which SimplifyDemandedVectorElts should try and use.

Differential Revision: https://reviews.llvm.org/D28119

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290554 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ADT] Add a boring std::partition wrapper similar to our std::remove_if
Chandler Carruth [Mon, 26 Dec 2016 23:10:40 +0000 (23:10 +0000)]
[ADT] Add a boring std::partition wrapper similar to our std::remove_if
wrapper.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290553 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoclang-format NewGVN files
Daniel Berlin [Mon, 26 Dec 2016 20:06:58 +0000 (20:06 +0000)]
clang-format NewGVN files

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290551 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMisc cleanups and simplifications for NewGVN.
Daniel Berlin [Mon, 26 Dec 2016 19:57:25 +0000 (19:57 +0000)]
Misc cleanups and simplifications for NewGVN.
Mostly use a bit more idiomatic C++ where we can,
so we can combine some things later.

Reviewers: davide

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D28111

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290550 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoDon't use our own incorrect version of isTriviallyDeadInstruction in NewGVN. Fixes...
Daniel Berlin [Mon, 26 Dec 2016 18:44:36 +0000 (18:44 +0000)]
Don't use our own incorrect version of isTriviallyDeadInstruction in NewGVN. Fixes PR/31472

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290549 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NewGVN] Add a flag to enable the pass via `-mllvm`.
Davide Italiano [Mon, 26 Dec 2016 18:26:19 +0000 (18:26 +0000)]
[NewGVN] Add a flag to enable the pass via `-mllvm`.

NewGVN can be tested passing `-mllvm -enable-newgvn` to clang.

Differential Revision:  https://reviews.llvm.org/D28059

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290548 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NewGVN] Change test to reflect difference between GVN and NewGVN.
Davide Italiano [Mon, 26 Dec 2016 18:10:09 +0000 (18:10 +0000)]
[NewGVN] Change test to reflect difference between GVN and NewGVN.

The current GVN algorithm folds unconditional branches to, it claims,
expose more PRE oportunities. The folding, if really needed,
(which is not sure, as it's not really proved it improves analysis)
can be done by an earlier cleanup pass instead of GVN itself.
Ack'ed/SGTM'd by Daniel Berlin.

Differential Revision:  https://reviews.llvm.org/D28117

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290546 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoWdocumentation fix
Simon Pilgrim [Mon, 26 Dec 2016 17:48:19 +0000 (17:48 +0000)]
Wdocumentation fix

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290545 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][AVX512] Added v64i8 reverse shuffle test (PR31470)
Simon Pilgrim [Mon, 26 Dec 2016 17:38:58 +0000 (17:38 +0000)]
[X86][AVX512] Added v64i8 reverse shuffle test (PR31470)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290544 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NewGVN] Fold lookupOperandLeader() when there's only one use. NFCI.
Davide Italiano [Mon, 26 Dec 2016 16:19:34 +0000 (16:19 +0000)]
[NewGVN] Fold lookupOperandLeader() when there's only one use. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290543 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombiner] Simplify lib calls to `round{,f}`
Bryant Wong [Mon, 26 Dec 2016 14:29:29 +0000 (14:29 +0000)]
[InstCombiner] Simplify lib calls to `round{,f}`

Differential Revision: https://reviews.llvm.org/D28110

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290542 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoTest the different scenarios of GlobalDCE and comdats more
Chandler Carruth [Mon, 26 Dec 2016 08:54:01 +0000 (08:54 +0000)]
Test the different scenarios of GlobalDCE and comdats more
systematically and document in the test what all is going on.

This replaces the PR-named test that was the only coverage for GlobalDCE
and comdats previously. I wrote this because I wasn't certain how
comdat DCE was supposed to work and wanted to step through what
GlobalDCE did to fully understand it. After talking to folks and reading
the code and really staring at things it all makes sense but it seemed
good to help write down some of this in a more explicit and fully
covering test case.

For example, it seemed like a bug that GlobalDCE didn't consider comdat
participation of ifuncs. Specifically it seemed like an accident because
testing didn't really cover that case. But in fact, ifuncs specifically
cannot participate in a comdat despite having that API. The new test
case covers this and explicitly documents that DCE gets to fire here
even though there are comdats involved.

Also, we didn't have any positive tests for the challenging cases such
as usage cycles between comdat participants that might make them seem
alive except that there is no external edge into the cycle.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290537 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512] Fix some patterns to use extended register classes.
Craig Topper [Mon, 26 Dec 2016 07:26:07 +0000 (07:26 +0000)]
[AVX-512] Fix some patterns to use extended register classes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290536 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512][InstCombine] Teach InstCombine to turn scalar add/sub/mul/div with rounding...
Craig Topper [Mon, 26 Dec 2016 06:33:19 +0000 (06:33 +0000)]
[AVX-512][InstCombine] Teach InstCombine to turn scalar add/sub/mul/div with rounding intrinsics into normal IR operations if the rounding mode is CUR_DIRECTION.

Summary:
I only do this for unmasked cases for now because isel is failing to fold the mask. I'll try to fix that soon.

I'll do the same thing for packed add/sub/mul/div in a future patch.

Reviewers: delena, RKSimon, zvi, craig.topper

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D27879

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290535 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512] Don't assume that the rounding mode argument to intrinsics is a constant...
Craig Topper [Mon, 26 Dec 2016 01:40:17 +0000 (01:40 +0000)]
[AVX-512] Don't assume that the rounding mode argument to intrinsics is a constant. While clang will guarantee this, nothing in the backend will.

A non-constant value will now result in an isel error instead of just asserting or crashing due to a bad cast during lowering.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290532 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix some bad indentation that I or another introduced somehow.
Chandler Carruth [Mon, 26 Dec 2016 01:20:59 +0000 (01:20 +0000)]
Fix some bad indentation that I or another introduced somehow.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290531 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512][InstCombine] Teach InstCombine to converted masked vpermv intrinsics into...
Craig Topper [Sun, 25 Dec 2016 23:58:57 +0000 (23:58 +0000)]
[AVX-512][InstCombine] Teach InstCombine to converted masked vpermv intrinsics into shufflevector instructions

Summary:
This patch adds support for converting the masked vpermv intrinsics into shufflevector instructions if the indices are constants.

We also need to wrap a select instruction around the shuffle to take care of the masking part. InstCombine will take care of optimizing the select if the mask is constant so I didn't bother checking for that.

Reviewers: zvi, delena, spatel, RKSimon

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D27825

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290530 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix `update_test_checks.py` bug that incorrectly truncates IR body.
Bryant Wong [Sun, 25 Dec 2016 23:46:55 +0000 (23:46 +0000)]
Fix `update_test_checks.py` bug that incorrectly truncates IR body.

Differential Revision: https://reviews.llvm.org/D26619

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290529 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ADT] Add a generic concatenating iterator and range (take 2).
Chandler Carruth [Sun, 25 Dec 2016 23:41:14 +0000 (23:41 +0000)]
[ADT] Add a generic concatenating iterator and range (take 2).

This recommits r290512 that was reverted when MSVC failed to compile it. Since
then I've played with various approaches using rextester.com (where I was able
to reproduce the failure) and think that I have a solution thanks in part to
the help of Dave Blaikie! It seems MSVC just has a defective `decltype` in this
version. Manually writing out the type seems to do the trick, even though it is
.... quite complicated.

Original commit message:
This allows both defining convenience iterator/range accessors on types
which walk across N different independent ranges within the object, and
more direct and simple usages with range based for loops such as shown
in the unittest. The same facilities are used for both. They end up
quite small and simple as it happens.

I've also switched an iterator on `Module` to use this. I would like to
add another convenience iterator that includes even more sequences as
part of it and seeing this one already present motivated me to actually
abstract it away and introduce a general utility.

Differential Revision: https://reviews.llvm.org/D28093

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290528 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[MemorySSA] Define a restricted upward AccessList splice.
Bryant Wong [Sun, 25 Dec 2016 23:34:07 +0000 (23:34 +0000)]
[MemorySSA] Define a restricted upward AccessList splice.

Differential Revision: https://reviews.llvm.org/D26661

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290527 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AliasAnalysis] Teach BasicAA about memcpy.
Bryant Wong [Sun, 25 Dec 2016 22:42:27 +0000 (22:42 +0000)]
[AliasAnalysis] Teach BasicAA about memcpy.

Differential Revision: https://reviews.llvm.org/D27034

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290526 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoValue number stores and memory states so we can detect when memory states are equival...
Daniel Berlin [Sun, 25 Dec 2016 22:23:49 +0000 (22:23 +0000)]
Value number stores and memory states so we can detect when memory states are equivalent (IE store of same value to memory).

Reviewers: davide

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D28084

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290525 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRename GVNExpression *ops_ members to *op_* to match conventions in the rest of LLVM
Daniel Berlin [Sun, 25 Dec 2016 22:10:37 +0000 (22:10 +0000)]
Rename GVNExpression *ops_ members to *op_* to match conventions in the rest of LLVM

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290524 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Orc][RPC] Add a ParallelCallGroup utility for dispatching and waiting on
Lang Hames [Sun, 25 Dec 2016 21:55:05 +0000 (21:55 +0000)]
[Orc][RPC] Add a ParallelCallGroup utility for dispatching and waiting on
multiple asynchronous RPC calls.

ParallelCallGroup allows multiple asynchronous calls to be dispatched,
and provides a wait method that blocks until all asynchronous calls have
been executed on the remote and all return value handlers run on the
local machine.

This will allow, for example, the JIT client to issue memory allocation calls
for all sections in parallel, then block until all memory has been allocated
on the remote and the allocated addresses registered with the client, at which
point the JIT client can proceed to applying relocations.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290523 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Orc][RPC] Clang-format RPCUtils header.
Lang Hames [Sun, 25 Dec 2016 19:55:59 +0000 (19:55 +0000)]
[Orc][RPC] Clang-format RPCUtils header.

Some of the recent RPC call type-checking changes weren't formatted prior to
commit.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290520 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd newline to end of file to quiet warnings.
Greg Clayton [Sun, 25 Dec 2016 18:41:47 +0000 (18:41 +0000)]
Add newline to end of file to quiet warnings.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290519 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agorevert commit 290516
Michael Zuckerman [Sun, 25 Dec 2016 12:45:18 +0000 (12:45 +0000)]
revert commit 290516

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290517 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoCommit try added new empty line
Michael Zuckerman [Sun, 25 Dec 2016 12:01:34 +0000 (12:01 +0000)]
Commit try added new empty line

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@290516 91177308-0d34-0410-b5e6-96231b3b80d8