OSDN Git Service

android-x86/external-llvm.git
8 years agoRevert "[InstCombine] Enable cast-folding in logic(cast(icmp), cast(icmp))"
Benjamin Kramer [Wed, 20 Jul 2016 11:40:16 +0000 (11:40 +0000)]
Revert "[InstCombine] Enable cast-folding in logic(cast(icmp), cast(icmp))"

Makes InstCombine infloop when compiling v8.

This reverts commit r275989 and r276105.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276106 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[InstCombine] Provide more test cases for cast-folding [NFC]
Tobias Grosser [Wed, 20 Jul 2016 11:24:27 +0000 (11:24 +0000)]
[InstCombine] Provide more test cases for cast-folding [NFC]

Summary: In r275989 we enabled the folding of `logic(cast(icmp), cast(icmp))` to `cast(logic(icmp, icmp))`. Here we add more test cases to assure this folding works for all logical operations `and`/`or`/`xor`.

Reviewers: grosser

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D22561

Contributed-by: Matthias Reisinger
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276105 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86][SSE] Add cost model values for CTPOP of vectors
Simon Pilgrim [Wed, 20 Jul 2016 10:41:28 +0000 (10:41 +0000)]
[X86][SSE] Add cost model values for CTPOP of vectors

This patch adds costs for the vectorized implementations of CTPOP, the default values were seriously underestimating the cost of these and was encouraging vectorization on targets where serialized use of POPCNT would be much better.

Differential Revision: https://reviews.llvm.org/D22456

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276104 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[ARM] Skip inline asm memory operands in DAGToDAGISel
Diana Picus [Wed, 20 Jul 2016 09:48:24 +0000 (09:48 +0000)]
[ARM] Skip inline asm memory operands in DAGToDAGISel

Retry r275776 (no changes, we suspect the issue was with another commit).

The current logic for handling inline asm operands in DAGToDAGISel interprets
the operands by looking for constants, which should represent the flags
describing the kind of operand we're dealing with (immediate, memory, register
def etc). The operands representing actual data are skipped only if they are
non-const, with the exception of immediate operands which are skipped explicitly
when a flag describing an immediate is found.

The oversight is that memory operands may be const too (e.g. for device drivers
reading a fixed address), so we should explicitly skip the operand following a
flag describing a memory operand. If we don't, we risk interpreting that
constant as a flag, which is definitely not intended.

Fixes PR26038

Differential Revision: https://reviews.llvm.org/D22103

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276101 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[docs] fix cmake code-block warning
Renato Golin [Wed, 20 Jul 2016 09:47:09 +0000 (09:47 +0000)]
[docs] fix cmake code-block warning

This will unblock the llvm-sphinx-buildbot, which is currently failing due
to a warning being treated as error.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276100 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[docs] Add proposals to index file
Renato Golin [Wed, 20 Jul 2016 09:38:04 +0000 (09:38 +0000)]
[docs] Add proposals to index file

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276099 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[docs] GitHub Proposal for LLVM
Renato Golin [Wed, 20 Jul 2016 09:09:58 +0000 (09:09 +0000)]
[docs] GitHub Proposal for LLVM

This document was crafted from the various (320+) emails between 2nd June and
20th July regarding the move to GitHub. It tried to consolidate every issue that
was raised and every solution that was presented to have a GitHub repository
with sub-modules.

It *does not* try to argue whether sub-modules are better or worse than any other
Git solution, nor if Git is better than any other VCS, nor if GitHub is better
than any other free code hosting service. This is just the final conclusions of
48 days and 320 emails (plus a lot of IRC discussions) on the LLVM community.

This document will be presented at the survey that the foundation will setup for
us to decide if we move to this solution or not. It reflects what was discussed
on the lists, but it's not authoritative. If something is not clear enough,
please refer to the mailing list discussions (hint: search for "GitHub").

Review: https://reviews.llvm.org/D22463

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276097 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AVX512] Add a missing NoVLX to give priority to the AVX512 version of the pattern.
Craig Topper [Wed, 20 Jul 2016 05:05:50 +0000 (05:05 +0000)]
[AVX512] Add a missing NoVLX to give priority to the AVX512 version of the pattern.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276088 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86] Use 'HasAVX1Only' to properly give priority to the AVX2 version without relying...
Craig Topper [Wed, 20 Jul 2016 05:05:48 +0000 (05:05 +0000)]
[X86] Use 'HasAVX1Only' to properly give priority to the AVX2 version without relying on file ordering.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276087 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86] Create some multiclasses to reduce the repeated patterns for VEXTRACT(F/I)128...
Craig Topper [Wed, 20 Jul 2016 05:05:46 +0000 (05:05 +0000)]
[X86] Create some multiclasses to reduce the repeated patterns for VEXTRACT(F/I)128/VINSERT(I/F)128. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276086 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86] Create some wrapper multiclasses to create AVX and SSE shift instructions with...
Craig Topper [Wed, 20 Jul 2016 05:05:44 +0000 (05:05 +0000)]
[X86] Create some wrapper multiclasses to create AVX and SSE shift instructions with less repeated code. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276085 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoForgot to add a test for r276008.
David Majnemer [Wed, 20 Jul 2016 04:13:05 +0000 (04:13 +0000)]
Forgot to add a test for r276008.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276082 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRevert "Disable this-return argument forwarding on ARM/AArch64"
David Majnemer [Wed, 20 Jul 2016 04:13:01 +0000 (04:13 +0000)]
Revert "Disable this-return argument forwarding on ARM/AArch64"

Inference of the 'returned' attribute was fixed in r276008, lets try
turning the backend support back on.

This reverts commit r275677.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276081 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[LV] Add hotness attribute to missed-optimization remarks
Adam Nemet [Wed, 20 Jul 2016 04:03:43 +0000 (04:03 +0000)]
[LV] Add hotness attribute to missed-optimization remarks

The new OptimizationRemarkEmitter analysis pass is hooked up to both new
and old PM passes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276080 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRevert "Revert r275883 and r275891. They seem to cause PR28608."
Michael Zolotukhin [Wed, 20 Jul 2016 01:55:27 +0000 (01:55 +0000)]
Revert "Revert r275883 and r275891. They seem to cause PR28608."

This reverts commit r276064, and thus reapplies r275891 and r275883 with
a fix for PR28608.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276077 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agollvm-readobj: add some more aliases
Saleem Abdulrasool [Wed, 20 Jul 2016 01:16:28 +0000 (01:16 +0000)]
llvm-readobj: add some more aliases

Alias -d and -t from readelf in llvm-readobj which effectively replaces the
tool.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276075 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[LSV] Don't assume that loads/stores appear in address order in the BB.
Justin Lebar [Wed, 20 Jul 2016 00:55:12 +0000 (00:55 +0000)]
[LSV] Don't assume that loads/stores appear in address order in the BB.

Summary:
getVectorizablePrefix previously didn't work properly in the face of
aliasing loads/stores.  It unwittingly assumed that the loads/stores
appeared in the BB in address order.  If they didn't, it would do the
wrong thing.

Reviewers: asbirlea, tstellarAMD

Subscribers: arsenm, llvm-commits, mzolotukhin

Differential Revision: https://reviews.llvm.org/D22535

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276072 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFixing a few places in this doc which look like obvious typos.
Yunzhong Gao [Wed, 20 Jul 2016 00:40:54 +0000 (00:40 +0000)]
Fixing a few places in this doc which look like obvious typos.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276070 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRevert "RegScavenging: Add scavengeRegisterBackwards()"
Matthias Braun [Wed, 20 Jul 2016 00:21:32 +0000 (00:21 +0000)]
Revert "RegScavenging: Add scavengeRegisterBackwards()"

Reverting this commit for now as it seems to be causing failures on
test-suite tests on the clang-ppc64le-linux-lnt bot.

This reverts commit r276044.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276068 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoCodegen: Tail Duplication: Only duplicate into layout pred if it is a CFG Pred.
Kyle Butt [Wed, 20 Jul 2016 00:01:51 +0000 (00:01 +0000)]
Codegen: Tail Duplication: Only duplicate into layout pred if it is a CFG Pred.

Add a check that the layout predecessor of a block is an actual CFG
predecssor of the block as well. No current code fails this check, but
upcoming patches can trigger this, and it makes sense to separate it
out.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276066 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRevert r275883 and r275891. They seem to cause PR28608.
Sean Silva [Tue, 19 Jul 2016 23:54:29 +0000 (23:54 +0000)]
Revert r275883 and r275891. They seem to cause PR28608.

Revert "[LoopSimplify] Update LCSSA after separating nested loops."

This reverts commit r275891.

Revert "[LCSSA] Post-process PHI-nodes created by SSAUpdate when constructing LCSSA form."

This reverts commit r275883.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276064 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[PM] Port LoopUnroll.
Sean Silva [Tue, 19 Jul 2016 23:54:23 +0000 (23:54 +0000)]
[PM] Port LoopUnroll.

We just set PreserveLCSSA to always true since we don't have an
analogous method `mustPreserveAnalysisID(LCSSA)`.

Also port LoopInfo verifier pass to test LoopUnrollPass.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276063 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoCodegen: Factor out canTailDuplicate
Kyle Butt [Tue, 19 Jul 2016 23:54:21 +0000 (23:54 +0000)]
Codegen: Factor out canTailDuplicate

canTailDuplicate accepts two blocks and returns true if the first can be
duplicated into the second successfully. Use this function to
encapsulate the heuristic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276062 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoThis code block breaks the docs build (http://lab.llvm.org:8011/builders/llvm-sphinx...
Aaron Ballman [Tue, 19 Jul 2016 23:50:11 +0000 (23:50 +0000)]
This code block breaks the docs build (lab.llvm.org:8011/builders/llvm-sphinx-docs/builds/11925/steps/docs-llvm-html/logs/stdio). Setting the code highlighting to none instead of llvm.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276060 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[ADT] Warn on unused results from ArrayRef and StringRef functions that read like...
Justin Lebar [Tue, 19 Jul 2016 23:19:25 +0000 (23:19 +0000)]
[ADT] Warn on unused results from ArrayRef and StringRef functions that read like they might mutate.

Summary:
Functions like "slice" and "drop_front" sound like they might mutate the
underlying object, but they don't.  Warning on unused results would have
saved me an hour yesterday, and I'm sure I'm not the only one.

LLVM and Clang are clean wrt this warning after D22540.

Reviewers: majnemer

Subscribers: sanjoy, chandlerc, llvm-commits

Differential Revision: https://reviews.llvm.org/D22541

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276058 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoGet rid of call to StringRef::substr that's never used.
Justin Lebar [Tue, 19 Jul 2016 23:19:22 +0000 (23:19 +0000)]
Get rid of call to StringRef::substr that's never used.

Summary: substr doesn't modify the string, so this line has no effect.

Reviewers: majnemer

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D22540

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276057 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[LSV] Insert stores at the right point.
Justin Lebar [Tue, 19 Jul 2016 23:19:20 +0000 (23:19 +0000)]
[LSV] Insert stores at the right point.

Summary:
Previously, the insertion point for stores was the last instruction in
Chain *before calling getVectorizablePrefixEndIdx*.  Thus if
getVectorizablePrefixEndIdx didn't return Chain.size(), we still would
insert at the last instruction in Chain.

This patch changes our internal API a bit in an attempt to make it less
prone to this sort of error.  As a result, we end up recalculating the
Chain's boundary instructions, but I think worrying about the speed hit
of this is a premature optimization right now.

Reviewers: asbirlea, tstellarAMD

Subscribers: mzolotukhin, arsenm, llvm-commits

Differential Revision: https://reviews.llvm.org/D22534

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276056 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[LSV] Use make_range, and reformat a DEBUG message. NFC
Justin Lebar [Tue, 19 Jul 2016 23:19:18 +0000 (23:19 +0000)]
[LSV] Use make_range, and reformat a DEBUG message. NFC

Summary:
The DEBUG message was hard to read because two Values were being printed
on the same line with only the delimiter "aliases".  This change makes
us print each Value on its own line.

Reviewers: asbirlea

Subscribers: llvm-commits, arsenm, mzolotukhin

Differential Revision: https://reviews.llvm.org/D22533

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276055 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[LSV] Nix two global (ish) variables in the LoadStoreVectorizer. NFC
Justin Lebar [Tue, 19 Jul 2016 23:19:16 +0000 (23:19 +0000)]
[LSV] Nix two global (ish) variables in the LoadStoreVectorizer.  NFC

Reviewers: asbirlea

Subscribers: mzolotukhin, llvm-commits, arsenm

Differential Revision: https://reviews.llvm.org/D22532

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276054 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[LSV] Add detail to correct-order.ll test.
Justin Lebar [Tue, 19 Jul 2016 23:18:59 +0000 (23:18 +0000)]
[LSV] Add detail to correct-order.ll test.

Summary:
This helps keep us honest -- there were a number of ways we could screw
up and still have passed this test.

Reviewers: asbirlea

Subscribers: llvm-commits, arsenm

Differential Revision: https://reviews.llvm.org/D22531

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276053 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[libFuzzer] extend the messages printed by afl_driver
Kostya Serebryany [Tue, 19 Jul 2016 23:18:28 +0000 (23:18 +0000)]
[libFuzzer] extend the messages printed by afl_driver

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276052 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAMDGPU: Change fdiv lowering based on !fpmath metadata
Matt Arsenault [Tue, 19 Jul 2016 23:16:53 +0000 (23:16 +0000)]
AMDGPU: Change fdiv lowering based on !fpmath metadata

If 2.5 ulp is acceptable, denormals are not required, and
isn't a reciprocal which will already be handled, replace
with a faster fdiv.

Simplify the lowering tests by using per function
subtarget features.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276051 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix unused variable
Daniel Berlin [Tue, 19 Jul 2016 23:08:08 +0000 (23:08 +0000)]
Fix unused variable

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276050 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoMake GVN Hoisting obey optnone/bisect.
Paul Robinson [Tue, 19 Jul 2016 22:57:14 +0000 (22:57 +0000)]
Make GVN Hoisting obey optnone/bisect.

Differential Revision: http://reviews.llvm.org/D22545

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276048 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoMake MemorySSA::dominates/locallydominates constant time
Daniel Berlin [Tue, 19 Jul 2016 22:49:43 +0000 (22:49 +0000)]
Make MemorySSA::dominates/locallydominates constant time

Summary: Make MemorySSA::dominates/locallydominates constant time

Reviewers: george.burgess.iv, gberry

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D22527

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276046 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAdd AIX support to Path.inc, Host.h, and CMake.
Chandler Carruth [Tue, 19 Jul 2016 22:46:39 +0000 (22:46 +0000)]
Add AIX support to Path.inc, Host.h, and CMake.

Patch by Andrew Paprocki!

Differential Revision: https://reviews.llvm.org/D18359

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276045 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRegScavenging: Add scavengeRegisterBackwards()
Matthias Braun [Tue, 19 Jul 2016 22:37:09 +0000 (22:37 +0000)]
RegScavenging: Add scavengeRegisterBackwards()

This is a variant of scavengeRegister() that works for
enterBasicBlockEnd()/backward(). The benefit of the backward mode is
that it is not affected by incomplete kill flags.

This patch also changes
PrologEpilogInserter::doScavengeFrameVirtualRegs() to use the register
scavenger in backwards mode.

Differential Revision: http://reviews.llvm.org/D21885

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276044 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRegisterScavenger: Introduce backward() mode.
Matthias Braun [Tue, 19 Jul 2016 22:37:02 +0000 (22:37 +0000)]
RegisterScavenger: Introduce backward() mode.

This adds two pieces:
- RegisterScavenger:::enterBasicBlockEnd() which behaves similar to
  enterBasicBlock() but starts tracking at the end of the basic block.
- A RegisterScavenger::backward() method. It is subtly different
  from the existing unprocess() method which only considers uses with
  the kill flag set: If a value is dead at the end of a basic block with
  a last use inside the basic block, unprocess() will fail to mark it as
  live. However we cannot change/fix this behaviour because unprocess()
  needs to perform the exact reverse operation of forward().

Differential Revision: http://reviews.llvm.org/D21873

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276043 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoregenerate checks
Sanjay Patel [Tue, 19 Jul 2016 22:32:15 +0000 (22:32 +0000)]
regenerate checks

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276042 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AArch64] Properly validate the reciprocal estimation.
Evandro Menezes [Tue, 19 Jul 2016 22:31:11 +0000 (22:31 +0000)]
[AArch64] Properly validate the reciprocal estimation.

Add check for legal data types when expanding into a Newton series.

Differential Revision: https://reviews.llvm.org/D22267

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276041 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[InstCombine] fold add(zext(xor X, C), C) --> sext X when C is INT_MIN in the source...
Sanjay Patel [Tue, 19 Jul 2016 22:09:34 +0000 (22:09 +0000)]
[InstCombine] fold add(zext(xor X, C), C) --> sext X when C is INT_MIN in the source type

The pattern may look more obviously like a sext if written as:

  define i32 @g(i16 %x) {
    %zext = zext i16 %x to i32
    %xor = xor i32 %zext, 32768
    %add = add i32 %xor, -32768
    ret i32 %add
  }

We already have that fold in visitAdd().

Differential Revision: https://reviews.llvm.org/D22477

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276035 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAttempt to appease MSVC buildbots.
George Burgess IV [Tue, 19 Jul 2016 21:35:47 +0000 (21:35 +0000)]
Attempt to appease MSVC buildbots.

Broken by r276026.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276032 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AMDGPU] Remove spurious line (should've been removed in r276029).
Davide Italiano [Tue, 19 Jul 2016 21:16:30 +0000 (21:16 +0000)]
[AMDGPU] Remove spurious line (should've been removed in r276029).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276030 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AMDGPU] Remove dead code.
Davide Italiano [Tue, 19 Jul 2016 21:10:49 +0000 (21:10 +0000)]
[AMDGPU] Remove dead code.

LGTM'd by Matt Arsenault.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276029 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[CFLAA] Make a test tell the truth. NFC.
George Burgess IV [Tue, 19 Jul 2016 20:56:41 +0000 (20:56 +0000)]
[CFLAA] Make a test tell the truth. NFC.

Dishonesty noted by Jia Chen.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276028 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[CFLAA] Add some interproc. analysis to CFLAnders.
George Burgess IV [Tue, 19 Jul 2016 20:47:15 +0000 (20:47 +0000)]
[CFLAA] Add some interproc. analysis to CFLAnders.

This patch adds function summary support to CFLAnders. It also comes
with a lot of tests! Woohoo!

Patch by Jia Chen.

Differential Revision: https://reviews.llvm.org/D22450

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276026 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoNext step along the way to getting good error messages for bad archives.
Kevin Enderby [Tue, 19 Jul 2016 20:47:07 +0000 (20:47 +0000)]
Next step along the way to getting good error messages for bad archives.

This step builds on Lang Hames work to change Archive::child_iterator
for better interoperation with Error/Expected.  Building on that it is now
possible to return an error message when the size field of an archive
contains non-decimal characters.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276025 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoadd even more missing tests for simplifySelectBitTest()
Sanjay Patel [Tue, 19 Jul 2016 20:47:00 +0000 (20:47 +0000)]
add even more missing tests for simplifySelectBitTest()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276024 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[CFLAA] Teach CFLAnders to distinguish reads from writes.
George Burgess IV [Tue, 19 Jul 2016 20:38:21 +0000 (20:38 +0000)]
[CFLAA] Teach CFLAnders to distinguish reads from writes.

This patch adds more specific edges to CFLAndersAliasAnalysis. The goal
of these edges is to give us more information about *how* two values
that MayAlias alias. With this, we can now tell cases like

a = b; // ergo, a may alias b

apart from

a = c;
b = c;

// so, a may alias b, but only because they were both assigned to c.

...And others.

Patch by Jia Chen.

Differential Revision: https://reviews.llvm.org/D22429

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276023 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoThis code block breaks the docs build (http://lab.llvm.org:8011/builders/llvm-sphinx...
Aaron Ballman [Tue, 19 Jul 2016 20:20:03 +0000 (20:20 +0000)]
This code block breaks the docs build (lab.llvm.org:8011/builders/llvm-sphinx-docs/builds/11921/steps/docs-llvm-html/logs/stdio). Setting the code highlighting to none instead of llvm to hopefully get the bot stumbling back towards green.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276018 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoUse posix_fallocate instead of ftruncate.
Rafael Espindola [Tue, 19 Jul 2016 20:19:56 +0000 (20:19 +0000)]
Use posix_fallocate instead of ftruncate.

This makes sure that space is actually available. With this change
running lld on a full file system causes it to exit with

failed to open foo: No space left on device

instead of crashing with a sigbus.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276017 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[tsan] Don't instrument __llvm_gcov_global_state_pred or __llvm_gcda*
Vedant Kumar [Tue, 19 Jul 2016 20:16:08 +0000 (20:16 +0000)]
[tsan] Don't instrument __llvm_gcov_global_state_pred or __llvm_gcda*

r274801 did not go far enough to allow gcov+tsan to cooperate. With this
commit it's possible to run the following code without false positives:

  std::thread T1(fib), T2(fib);
  T1.join(); T2.join();

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276015 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoARM: move feature for Thumb2 pkhbt/pkhtb onto architectures.
Tim Northover [Tue, 19 Jul 2016 19:49:13 +0000 (19:49 +0000)]
ARM: move feature for Thumb2 pkhbt/pkhtb onto architectures.

There's not much functional change, but it really is an architectural feature
(on v6T2, v7A, v7R and v7EM) rather than something each CPU implements
individually.

The main functional change is the default behaviour you get when specifying
only "-triple".

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276013 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[GlobalISel] Mark newly-created gvregs as having a bank.
Ahmed Bougacha [Tue, 19 Jul 2016 19:48:36 +0000 (19:48 +0000)]
[GlobalISel] Mark newly-created gvregs as having a bank.

Also verify that we never try to set the size of a vreg associated
to a register class.

Report an error when we encounter that in MIR. Fix a testcase that
hit that error and had a size for no reason.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276012 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[GlobalISel] Simplify more RegClassOrRegBank is+get. NFC.
Ahmed Bougacha [Tue, 19 Jul 2016 19:47:06 +0000 (19:47 +0000)]
[GlobalISel] Simplify more RegClassOrRegBank is+get. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276011 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[FunctionAttrs] Correct the safety analysis for inference of 'returned'
David Majnemer [Tue, 19 Jul 2016 18:50:26 +0000 (18:50 +0000)]
[FunctionAttrs] Correct the safety analysis for inference of 'returned'

We skipped over ReturnInsts which didn't return an argument which would
lead us to incorrectly conclude that an argument returned by another
ReturnInst was 'returned'.

This reverts commit r275756.

This fixes PR28610.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276008 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[SCCP] Improve assert messages. NFCI.
Davide Italiano [Tue, 19 Jul 2016 18:31:07 +0000 (18:31 +0000)]
[SCCP] Improve assert messages. NFCI.

I've been hitting those already while working on SCCP and I think
it's be useful to provide a more explanatory diagnostic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276007 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[libFuzzer] properly intercept memmem
Kostya Serebryany [Tue, 19 Jul 2016 18:29:06 +0000 (18:29 +0000)]
[libFuzzer] properly intercept memmem

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276006 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[DSE] Add additional debug output. NFC.
Chad Rosier [Tue, 19 Jul 2016 18:11:11 +0000 (18:11 +0000)]
[DSE] Add additional debug output. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276005 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAdd a testcase for r275581
David Majnemer [Tue, 19 Jul 2016 17:52:41 +0000 (17:52 +0000)]
Add a testcase for r275581

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276002 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[RegionInfo] Some cleanups
David Majnemer [Tue, 19 Jul 2016 17:50:30 +0000 (17:50 +0000)]
[RegionInfo] Some cleanups

- Use unique_ptr instead of managing a container of new'd pointers.
- Use range based for loops.

No functional change is intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276001 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[RegionPass] Some minor cleanups
David Majnemer [Tue, 19 Jul 2016 17:50:27 +0000 (17:50 +0000)]
[RegionPass] Some minor cleanups

No functional change is intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@276000 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[LoopPass] Some minor cleanups
David Majnemer [Tue, 19 Jul 2016 17:50:24 +0000 (17:50 +0000)]
[LoopPass] Some minor cleanups

No functional change is intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275999 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoThis code block breaks the docs build (http://lab.llvm.org:8011/builders/llvm-sphinx...
Aaron Ballman [Tue, 19 Jul 2016 17:46:55 +0000 (17:46 +0000)]
This code block breaks the docs build (lab.llvm.org:8011/builders/llvm-sphinx-docs/builds/11920/steps/docs-llvm-html/logs/stdio), but I cannot see anything immediately wrong with it and cannot reproduce the diagnostic locally. Setting the code highlighting to none instead of nasm to hopefully get the bot stumbling back towards green.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275998 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoadd tests related to PR28466
Sanjay Patel [Tue, 19 Jul 2016 17:07:35 +0000 (17:07 +0000)]
add tests related to PR28466

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275995 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86][AVX512] Added AVX512 subvector broadcast tests
Simon Pilgrim [Tue, 19 Jul 2016 17:04:28 +0000 (17:04 +0000)]
[X86][AVX512] Added AVX512 subvector broadcast tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275994 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86][AVX] Fixed typo in test names
Simon Pilgrim [Tue, 19 Jul 2016 16:52:05 +0000 (16:52 +0000)]
[X86][AVX] Fixed typo in test names

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275992 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[DSE] Add additional debug output. NFC.
Chad Rosier [Tue, 19 Jul 2016 16:50:57 +0000 (16:50 +0000)]
[DSE] Add additional debug output. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275991 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoadd missing test for simplifySelectBitTest()
Sanjay Patel [Tue, 19 Jul 2016 16:49:55 +0000 (16:49 +0000)]
add missing test for simplifySelectBitTest()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275990 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[InstCombine] Enable cast-folding in logic(cast(icmp), cast(icmp))
Tobias Grosser [Tue, 19 Jul 2016 16:39:17 +0000 (16:39 +0000)]
[InstCombine] Enable cast-folding in logic(cast(icmp), cast(icmp))

Summary:
Currently, InstCombine is already able to fold expressions of the form `logic(cast(A), cast(B))` to the simpler form `cast(logic(A, B))`, where logic designates one of `and`/`or`/`xor`. This transformation is implemented in `foldCastedBitwiseLogic()` in InstCombineAndOrXor.cpp. However, this optimization will not be performed if both `A` and `B` are `icmp` instructions. The decision to preclude casts of `icmp` instructions originates in r48715 in combination with r261707, and can be best understood by the title of the former one:

> Transform (zext (or (icmp), (icmp))) to (or (zext (cimp), (zext icmp))) if at least one of the (zext icmp) can be transformed to eliminate an icmp.

Apparently, it introduced a transformation that is a reverse of the transformation that is done in `foldCastedBitwiseLogic()`. Its purpose is to expose pairs of `zext icmp` that would subsequently be optimized by `transformZExtICmp()` in InstCombineCasts.cpp. Therefore, in order to avoid an endless loop of switching back and forth between these two transformations, the one in `foldCastedBitwiseLogic()` has been restricted to exclude `icmp` instructions which is mirrored in the responsible check:

`if ((!isa<ICmpInst>(Cast0Src) || !isa<ICmpInst>(Cast1Src)) && ...`

This check seems to sort out more cases than necessary because:
- the reverse transformation is obviously done for `or` instructions only
- and also not every `zext icmp` pair is necessarily the result of this reverse transformation

Therefore we now remove this check and replace it by a more finegrained one in `shouldOptimizeCast()` that now rejects only those `logic(zext(icmp), zext(icmp))` that would be able to be optimized by `transformZExtICmp()`, which also avoids the mentioned endless loop. That means we are now able to also simplify expressions of the form `logic(cast(icmp), cast(icmp))` to `cast(logic(icmp, icmp))` (`cast` being an arbitrary `CastInst`).

As an example, consider the following IR snippet

```
%1 = icmp sgt i64 %a, %b
%2 = zext i1 %1 to i8
%3 = icmp slt i64 %a, %c
%4 = zext i1 %3 to i8
%5 = and i8 %2, %4
```

which would now be transformed to

```
%1 = icmp sgt i64 %a, %b
%2 = icmp slt i64 %a, %c
%3 = and i1 %1, %2
%4 = zext i1 %3 to i8
```

This issue became apparent when experimenting with the programming language Julia, which makes use of LLVM. Currently, Julia lowers its `Bool` datatype to LLVM's `i8` (also see https://github.com/JuliaLang/julia/pull/17225). In fact, the above IR example is the lowered form of the Julia snippet `(a > b) & (a < c)`. Like shown above, this may introduce `zext` operations, casting between `i1` and `i8`, which could for example hinder ScalarEvolution and Polly on certain code.

Reviewers: grosser, vtjnash, majnemer

Subscribers: majnemer, llvm-commits

Differential Revision: https://reviews.llvm.org/D22511

Contributed-by: Matthias Reisinger
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275989 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAMDGPU: Only use legal inline immediates with kill pseudo
Matt Arsenault [Tue, 19 Jul 2016 16:27:56 +0000 (16:27 +0000)]
AMDGPU: Only use legal inline immediates with kill pseudo

Only if the value is negative or positive is what matters,
so use a constant that doesn't require an instruction to
materialize.

These should really just emit the write exec directly,
but for stick with the kill pseudo-terminator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275988 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86][SSE] Reimplement SSE fp2si conversion intrinsics instead of using generic IR
Simon Pilgrim [Tue, 19 Jul 2016 15:07:43 +0000 (15:07 +0000)]
[X86][SSE] Reimplement SSE fp2si conversion intrinsics instead of using generic IR

D20859 and D20860 attempted to replace the SSE (V)CVTTPS2DQ and VCVTTPD2DQ truncating conversions with generic IR instead.

It turns out that the behaviour of these intrinsics is different enough from generic IR that this will cause problems, INF/NAN/out of range values are guaranteed to result in a 0x80000000 value - which plays havoc with constant folding which converts them to either zero or UNDEF. This is also an issue with the scalar implementations (which were already generic IR and what I was trying to match).

This patch changes both scalar and packed versions back to using x86-specific builtins.

It also deals with the other scalar conversion cases that are runtime rounding mode dependent and can have similar issues with constant folding.

A companion clang patch is at D22105

Differential Revision: https://reviews.llvm.org/D22106

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275981 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[ARM] Refactor Thumb2 Mul and Mla instr descs
Sam Parker [Tue, 19 Jul 2016 14:44:05 +0000 (14:44 +0000)]
[ARM] Refactor Thumb2 Mul and Mla instr descs

Recommitting after r274347 was reverted. This patch introduces some
classes to refactor the 3 and 4 register Thumb2 multiplication
instruction descriptions, plus improved tests for some of those
instructions.

Differential Revision: https://reviews.llvm.org/D21929

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275979 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AArch64] PredictableSelectIsExpensive for Vulcan.
Pankaj Gode [Tue, 19 Jul 2016 14:30:21 +0000 (14:30 +0000)]
[AArch64] PredictableSelectIsExpensive for Vulcan.

Adding PredictableSelectIsExpensive for Vulcan

Differential Revision: https://reviews.llvm.org/D22448

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275978 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAdd support for tlsldm assembler operator to ARM target
Peter Smith [Tue, 19 Jul 2016 14:15:33 +0000 (14:15 +0000)]
Add support for tlsldm assembler operator to ARM target

The standard local dynamic model for TLS on ARM systems needs two
relocations:
- R_ARM_TLS_LDM32 (module idx)
- R_ARM_TLS_LDO32 (offset of object from origin of module TLS block)

In GNU style assembler we use symbol(tlsldm) and symbol(tlsldo) to
produce these relocations.

llvm-mc for ARM supports symbol(tlsldo) but does not support symbol(tlsldm).
This patch wires up the existing symbol(tlsldm) to R_ARM_TLS_LDM32.

TLS for ARM is defined in Addenda to, and Errata in, the ABI for the
ARM Architecture

Differential Revision: https://reviews.llvm.org/D22461

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275977 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AARCH64] Fix linu triple typo
Simon Pilgrim [Tue, 19 Jul 2016 14:12:45 +0000 (14:12 +0000)]
[AARCH64] Fix linu triple typo

As promised in D22191

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275976 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AARCH64] Enable AARCH64 lit tests on windows dev machines
Simon Pilgrim [Tue, 19 Jul 2016 13:35:11 +0000 (13:35 +0000)]
[AARCH64] Enable AARCH64 lit tests on windows dev machines

As discussed on PR27654, this patch fixes the triples of a lot of aarch64 tests and enables lit tests on windows

This will hopefully help stop cases where windows developers break the aarch64 target

Differential Revision: https://reviews.llvm.org/D22191

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275973 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoGet rid of VS2015 operator precedence warning. NFCI.
Simon Pilgrim [Tue, 19 Jul 2016 12:26:51 +0000 (12:26 +0000)]
Get rid of VS2015 operator precedence warning. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275971 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[mips][ias] R_MIPS_GOT_(PAGE|OFST) do not need symbols
Daniel Sanders [Tue, 19 Jul 2016 10:58:06 +0000 (10:58 +0000)]
[mips][ias] R_MIPS_GOT_(PAGE|OFST) do not need symbols

Reviewers: sdardis

Subscribers: dsanders, llvm-commits, sdardis

Differential Revision: https://reviews.llvm.org/D22458

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275968 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[mips] Correct label prefixes for N32 and N64.
Daniel Sanders [Tue, 19 Jul 2016 10:49:03 +0000 (10:49 +0000)]
[mips] Correct label prefixes for N32 and N64.

Summary:
N32 and N64 follow the standard ELF conventions (.L) whereas O32 uses its own
($).

This fixes the majority of object differences between -fintegrated-as and
-fno-integrated-as.

Reviewers: sdardis

Subscribers: dsanders, sdardis, llvm-commits

Differential Revision: https://reviews.llvm.org/D22412

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275967 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[mips] Recognise the triple used by Debian stretch for mips64el.
Daniel Sanders [Tue, 19 Jul 2016 10:22:19 +0000 (10:22 +0000)]
[mips] Recognise the triple used by Debian stretch for mips64el.

Summary:
The triple used for this distribution is mips64el-linux-gnuabi64.

Reviewers: sdardis

Subscribers: sdardis, llvm-commits

Differential Revision: https://reviews.llvm.org/D22406

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275966 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[InstCombine] Minor cleanup of cast simplification code [NFC]
Tobias Grosser [Tue, 19 Jul 2016 09:06:08 +0000 (09:06 +0000)]
[InstCombine] Minor cleanup of cast simplification code [NFC]

Summary:
This patch cleans up parts of InstCombine to raise its compliance with the LLVM coding standards and to increase its readability. The changes and according rationale are summarized in the following:

- Rename `ShouldOptimizeCast()` to `shouldOptimizeCast()` since functions should start with a lower case letter.

- Move `shouldOptimizeCast()` from InstCombineCasts.cpp to InstCombineAndOrXor.cpp since it's only used there.

- Simplify interface of `shouldOptimizeCast()`.

- Minor code style adaptions in `shouldOptimizeCast()`.

- Remove the documentation on the function definition of `shouldOptimizeCast()` since it just repeats the documentation on its declaration. Also enhance the documentation on its declaration with more information describing its intended use and make it doxygen-compliant.

- Change a comment in `foldCastedBitwiseLogic()` from `fold (logic (cast A), (cast B)) -> (cast (logic A, B))` to `fold logic(cast(A), cast(B)) -> cast(logic(A, B))` since the surrounding comments use this format.

- Remove comment `Only do this if the casts both really cause code to be generated.` in `foldCastedBitwiseLogic()` since it just repeats parts of the documentation of `shouldOptimizeCast()` and does not help to improve readability.

- Simplify the interface of `isEliminableCastPair()`.

- Removed the documentation on the function definition of `isEliminableCastPair()` which only contained obvious statements about its implementation. Instead added more general doxygen-compliant documentation to its declaration.

- Renamed parameter `DoXform` of `transformZExtIcmp()` to `DoTransform` to make its intention clearer.

- Moved documentation of `transformZExtIcmp()` from its definition to its declaration and made it doxygen-compliant.

Reviewers: vtjnash, grosser

Subscribers: majnemer, llvm-commits

Differential Revision: https://reviews.llvm.org/D22449

Contributed-by: Matthias Reisinger
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275964 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoStyle: drop some unnecessary ';' [NFC]
Tobias Grosser [Tue, 19 Jul 2016 09:01:46 +0000 (09:01 +0000)]
Style: drop some unnecessary ';' [NFC]

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275963 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAVX-512: Fixed BT instruction selection.
Elena Demikhovsky [Tue, 19 Jul 2016 07:14:21 +0000 (07:14 +0000)]
AVX-512: Fixed BT instruction selection.

The following condition expression ( a >> n) & 1 is converted to "bt a, n" instruction. It works on all intel targets.
But on AVX-512 it was broken because the expression is modified to (truncate (a >>n) to i1).

I added the new sequence (truncate (a >>n) to i1) to the BT pattern.

Differential Revision: https://reviews.llvm.org/D22354

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275950 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[AVX512] Give priority to EVEX encoded PSHUFB over the VEX versions.
Craig Topper [Tue, 19 Jul 2016 02:00:38 +0000 (02:00 +0000)]
[AVX512] Give priority to EVEX encoded PSHUFB over the VEX versions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275942 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86] Remove superfluous parameter from a multiclass. All instantiations passed the...
Craig Topper [Tue, 19 Jul 2016 02:00:35 +0000 (02:00 +0000)]
[X86] Remove superfluous parameter from a multiclass. All instantiations passed the same value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275941 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[MemorySSA] Update to the new shiny walker.
George Burgess IV [Tue, 19 Jul 2016 01:29:15 +0000 (01:29 +0000)]
[MemorySSA] Update to the new shiny walker.

This patch updates MemorySSA's use-optimizing walker to be more
accurate and, in some cases, faster.

Essentially, this changed our core walking algorithm from a
cache-as-you-go DFS to an iteratively expanded DFS, with all of the
caching happening at the end. Said expansion happens when we hit a Phi,
P; we'll try to do the smallest amount of work possible to see if
optimizing above that Phi is legal in the first place. If so, we'll
expand the search to see if we can optimize to the next phi, etc.

An iteratively expanded DFS lets us potentially quit earlier (because we
don't assume that we can optimize above all phis) than our old walker.
Additionally, because we don't cache as we go, we can now optimize above
loops.

As an added bonus, this patch adds a ton of verification (if
EXPENSIVE_CHECKS are enabled), so finding bugs is easier.

Differential Revision: https://reviews.llvm.org/D21777

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275940 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[X86] Rename VINSERTzrr to use a capital Z to match other instructions. NFC
Craig Topper [Tue, 19 Jul 2016 01:26:19 +0000 (01:26 +0000)]
[X86] Rename VINSERTzrr to use a capital Z to match other instructions. NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275939 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRetry: [llvm-profdata] Speed up merging by using a thread pool
Vedant Kumar [Tue, 19 Jul 2016 01:17:20 +0000 (01:17 +0000)]
Retry: [llvm-profdata] Speed up merging by using a thread pool

Add a "-j" option to llvm-profdata to control the number of threads used.
Auto-detect NumThreads when it isn't specified, and avoid spawning threads when
they wouldn't be beneficial.

I tested this patch using a raw profile produced by clang (147MB). Here is the
time taken to merge 4 copies together on my laptop:

  No thread pool: 112.87s user 5.92s system 97% cpu 2:01.08 total
  With 2 threads: 134.99s user 26.54s system 164% cpu 1:33.31 total

Changes since the initial commit:

  - When handling odd-length inputs, call ThreadPool::wait() before merging the
    last profile. Should fix a race/off-by-one (see r275937).

Differential Revision: https://reviews.llvm.org/D22438

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275938 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRevert "[llvm-profdata] Speed up merging by using a thread pool"
Vedant Kumar [Tue, 19 Jul 2016 00:57:09 +0000 (00:57 +0000)]
Revert "[llvm-profdata] Speed up merging by using a thread pool"

This reverts commit r275921. It broke the ppc64be bot:

  http://lab.llvm.org:8011/builders/clang-ppc64be-linux-multistage/builds/3537

I'm not sure why it broke, but based on the output, it looks like an
off-by-one (one profile left un-merged).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275937 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoRecommit the patch "Use uniforms set to populate VecValuesToIgnore".
Wei Mi [Tue, 19 Jul 2016 00:50:43 +0000 (00:50 +0000)]
Recommit the patch "Use uniforms set to populate VecValuesToIgnore".

For instructions in uniform set, they will not have vector versions so
add them to VecValuesToIgnore.
For induction vars, those only used in uniform instructions or consecutive
ptrs instructions have already been added to VecValuesToIgnore above. For
those induction vars which are only used in uniform instructions or
non-consecutive/non-gather scatter ptr instructions, the related phi and
update will also be added into VecValuesToIgnore set.

The change will make the vector RegUsages estimation less conservative.

Differential Revision: https://reviews.llvm.org/D20474

The recommit fixed the testcase global_alias.ll.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275936 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAMDGPU/SI: Fix SI scheduler refcount issue
Matt Arsenault [Tue, 19 Jul 2016 00:35:22 +0000 (00:35 +0000)]
AMDGPU/SI: Fix SI scheduler refcount issue

Without this fix, releaseSuccessors when InOrOutBlock is
false could release SUs outside the schedule BasicBlock.

Patch by Axel Davy

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275935 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAMDGPU: Expand register indexing pseudos in custom inserter
Matt Arsenault [Tue, 19 Jul 2016 00:35:03 +0000 (00:35 +0000)]
AMDGPU: Expand register indexing pseudos in custom inserter

This is to help moveSILowerControlFlow to before regalloc.
There are a couple of tradeoffs with this. The complete CFG
is visible to more passes, the loop body avoids an extra copy of m0,
vcc isn't required, and immediate offsets can be shrunk into s_movk_i32.

The disadvantage is the register allocator doesn't understand that
the single lane's vector is dead within the loop body, so an extra
register is used to outlive the loop block when expanding the
VGPR -> m0 loop. This also now results in worse waitcnt insertion
before the loop instead of after for pending operations at the point
of the indexing, but that should be fixed by future improvements to
cross block waitcnt insertion.

v_movreld_b32's operands are now modeled more correctly since vdst
is not a true output. This is kind of a hack to treat vdst as a
use operand. Extra checking is required in the verifier since
I can't seem to get tablegen to emit an implicit operand for a
virtual register.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275934 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[Kaleidoscope][BuildingAJIT] More work on the text for Chapter 3.
Lang Hames [Tue, 19 Jul 2016 00:25:52 +0000 (00:25 +0000)]
[Kaleidoscope][BuildingAJIT] More work on the text for Chapter 3.

Add an overview of stubs and compile callbacks before the discussion of the
source changes.

-- This line, and those below, will be ignored--

M    docs/tutorial/BuildingAJIT3.rst

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275933 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[LoopReroll] Reroll loops with unordered atomic memory accesses
Sanjoy Das [Tue, 19 Jul 2016 00:23:54 +0000 (00:23 +0000)]
[LoopReroll] Reroll loops with unordered atomic memory accesses

Reviewers: hfinkel, jfb, reames

Subscribers: mcrosier, mzolotukhin, llvm-commits

Differential Revision: https://reviews.llvm.org/D22385

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275932 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoTableGen: Allow custom register operand decoder method
Matt Arsenault [Mon, 18 Jul 2016 23:20:46 +0000 (23:20 +0000)]
TableGen: Allow custom register operand decoder method

This is for a situation where the encoding for a register may be
different depending on the specific operand. For some instructions,
we want to apply additional restrictions beyond the encoding's
constraints.

In AMDGPU some operands are VSrc_32, using the VS_32 pseudo register
class which accept VGPRs, SGPRs, or immediates in the encoding.
Some specific instructions with the same encoding operand do not want
to allow immediates or SGPRs, but the encoding format is different
in this case than a regular VGPR_32 operand.

This allows specifying the encoding should be treated the same
without introducing yet another dummy register class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275929 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoAMDGPU: Fix test name and broken CHECK-LABEL
Matt Arsenault [Mon, 18 Jul 2016 23:09:51 +0000 (23:09 +0000)]
AMDGPU: Fix test name and broken CHECK-LABEL

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275928 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[utils] Generate html reports with the code coverage utility script
Vedant Kumar [Mon, 18 Jul 2016 22:50:10 +0000 (22:50 +0000)]
[utils] Generate html reports with the code coverage utility script

Instead of extracting raw coverage mappings into an artifact directory,
actually generate useful html reports for a given list of binaries with
symbol demangling turned on.

No tests, but this is actively being used to drive the (still nascent)
coverage bot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275927 91177308-0d34-0410-b5e6-96231b3b80d8

8 years agoFix -Wreturn-type with gcc 4.8 and libc++
Matt Arsenault [Mon, 18 Jul 2016 22:12:46 +0000 (22:12 +0000)]
Fix -Wreturn-type with gcc 4.8 and libc++

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275922 91177308-0d34-0410-b5e6-96231b3b80d8

8 years ago[llvm-profdata] Speed up merging by using a thread pool
Vedant Kumar [Mon, 18 Jul 2016 22:02:39 +0000 (22:02 +0000)]
[llvm-profdata] Speed up merging by using a thread pool

Add a "-j" option to llvm-profdata to control the number of threads
used. Auto-detect NumThreads when it isn't specified, and avoid spawning
threads when they wouldn't be beneficial.

I tested this patch using a raw profile produced by clang (147MB). Here is the
time taken to merge 4 copies together on my laptop:

  No thread pool: 112.87s user 5.92s system 97% cpu 2:01.08 total
  With 2 threads: 134.99s user 26.54s system 164% cpu 1:33.31 total

Differential Revision: https://reviews.llvm.org/D22438

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@275921 91177308-0d34-0410-b5e6-96231b3b80d8