OSDN Git Service

android-x86/external-llvm.git
5 years ago[llvm-mca][BtVer2] Modify ANDN tests in zero-idioms-avx-256.s. NFC
Andrea Di Biagio [Thu, 20 Sep 2018 15:48:23 +0000 (15:48 +0000)]
[llvm-mca][BtVer2] Modify ANDN tests in zero-idioms-avx-256.s. NFC

Two test cases should have tested 256-bit variants of VANDN zero-idioms instead
of the 128-bit variants.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342655 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[IR] add shuffle query for vector concatenation
Sanjay Patel [Thu, 20 Sep 2018 15:21:52 +0000 (15:21 +0000)]
[IR] add shuffle query for vector concatenation

This can be used for combining and in the vectorizers/cost models.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342653 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[IR] reduce duplication in unittest for shuffles; NFC
Sanjay Patel [Thu, 20 Sep 2018 14:36:09 +0000 (14:36 +0000)]
[IR] reduce duplication in unittest for shuffles; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342650 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] Handle vector compares in foldGEPIcmp()
Jesper Antonsson [Thu, 20 Sep 2018 13:37:28 +0000 (13:37 +0000)]
[InstCombine] Handle vector compares in foldGEPIcmp()

Summary:
This is to fix PR38984 "InstCombine assertion at vector gep/icmp folding":
https://bugs.llvm.org/show_bug.cgi?id=38984

Reviewers: majnemer, spatel, lattner, lebedev.ri

Reviewed By: lebedev.ri

Subscribers: lebedev.ri, llvm-commits

Differential Revision: https://reviews.llvm.org/D52263

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342647 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-exegesis] Fix broken build bots.
Guillaume Chatelet [Thu, 20 Sep 2018 13:37:04 +0000 (13:37 +0000)]
[llvm-exegesis] Fix broken build bots.

Reviewers: javed.absar

Subscribers: tschuett, courbet, llvm-commits

Differential Revision: https://reviews.llvm.org/D52302

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342646 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86][SSE] Remove UNPCKL(SHUFFLE)->UNPCKH custom combine
Simon Pilgrim [Thu, 20 Sep 2018 13:10:22 +0000 (13:10 +0000)]
[X86][SSE] Remove UNPCKL(SHUFFLE)->UNPCKH custom combine

This can be achieved more generally by combineX86ShufflesRecursively.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342645 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-exegesis] Improve Register Setup (roll forward of D51856).
Guillaume Chatelet [Thu, 20 Sep 2018 12:22:18 +0000 (12:22 +0000)]
[llvm-exegesis] Improve Register Setup (roll forward of D51856).

Summary:
Added function to set a register to a particular value + tests.
Add EFLAGS test, use new setRegTo instead of setRegToConstant.

Reviewers: courbet, javed.absar

Subscribers: llvm-commits, tschuett, mgorny

Differential Revision: https://reviews.llvm.org/D52297

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342644 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ADT] Bring back memmove to make GCC 5.4 happy
Benjamin Kramer [Thu, 20 Sep 2018 12:21:24 +0000 (12:21 +0000)]
[ADT] Bring back memmove to make GCC 5.4 happy

All other GCCs look good so far. GCC 5.4 complains about strict
aliasing, so fix that.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342643 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86][SSE] Remove PSHUFLW/PSHUFHW combineRedundantHalfShuffle combine
Simon Pilgrim [Thu, 20 Sep 2018 12:11:38 +0000 (12:11 +0000)]
[X86][SSE] Remove PSHUFLW/PSHUFHW combineRedundantHalfShuffle combine

This can be achieved more generally by combineX86ShufflesRecursively and was causing a fuzz test failure found by Mikael Holmén.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342642 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[RISCV][MC] Modify evaluateConstantImm interface to allow reuse from addExpr
Alex Bradbury [Thu, 20 Sep 2018 11:40:43 +0000 (11:40 +0000)]
[RISCV][MC] Modify evaluateConstantImm interface to allow reuse from addExpr

This is a trivial refactoring that I'm committing now as it makes a patch I'm
about to post for review easier to follow. There is some overlap between
evaluateConstantImm and addExpr in RISCVAsmParser. This patch allows
evaluateConstantImm to be reused from addExpr to remove this overlap. The
benefit will be greater when a future patch adds extra code to allows
immediates to be evaluated from constant symbols (e.g. `.equ CONST, 0x1234`).

No functional change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342641 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix -Wdocumentation warnings introduced by r342555. NFC
Andrea Di Biagio [Thu, 20 Sep 2018 11:07:57 +0000 (11:07 +0000)]
Fix -Wdocumentation warnings introduced by r342555. NFC

These were reported by buildbot llvm-clang-lld-x86_64-scei-ps4-ubuntu-fast (see
build #36798).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342640 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix line-endings. NFCI.
Simon Pilgrim [Thu, 20 Sep 2018 10:59:08 +0000 (10:59 +0000)]
Fix line-endings. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342639 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ADT] Try again to use the same version of llvm::Optional on all compilers
Benjamin Kramer [Thu, 20 Sep 2018 10:02:06 +0000 (10:02 +0000)]
[ADT] Try again to use the same version of llvm::Optional on all compilers

The miscompile doesn't reproduce for me anymore with GCC 7.3. I'll watch
the buildbots closely.

Having different versions of Optional is an ABI violation when linking
GCC- and clang-built code together.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342637 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DWARF] - Emit the correct value for DW_AT_addr_base.
George Rimar [Thu, 20 Sep 2018 09:17:36 +0000 (09:17 +0000)]
[DWARF] - Emit the correct value for DW_AT_addr_base.

Currently, we emit DW_AT_addr_base that points to the beginning of
the .debug_addr section. That is not correct for the DWARF5 case because address
table contains the header and the attribute should point to the first entry
following the header.

This is currently the reason why LLDB does not work with such executables correctly.
Patch fixes the issue.

Differential revision: https://reviews.llvm.org/D52168

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342635 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[IPSCCP] Fix a problem with removing labels in a switch with undef condition
Bjorn Pettersson [Thu, 20 Sep 2018 09:00:17 +0000 (09:00 +0000)]
[IPSCCP] Fix a problem with removing labels in a switch with undef condition

Summary:
Before removing basic blocks that ipsccp has considered as dead
all uses of the basic block label must be removed. That is done
by calling ConstantFoldTerminator on the users. An exception
is when the branch condition is an undef value. In such
scenarios ipsccp is using some internal assumptions regarding
which edge in the control flow that should remain, while
ConstantFoldTerminator don't know how to fold the terminator.

The problem addressed here is related to ConstantFoldTerminator's
ability to rewrite a 'switch' into a conditional 'br'. In such
situations ConstantFoldTerminator returns true indicating that
the terminator has been rewritten. However, ipsccp treated the
true value as if the edge to the dead basic block had been
removed. So the code for resolving an undef branch condition
did not trigger, and we ended up with assertion that there were
uses remaining when deleting the basic block.

The solution is to resolve indeterminate branches before the
call to ConstantFoldTerminator.

Reviewers: efriedma, fhahn, davide

Reviewed By: fhahn

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D52232

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342632 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[IR] Add a boolean field in DILocation to know if a line must covered or not
Calixte Denizet [Thu, 20 Sep 2018 08:53:06 +0000 (08:53 +0000)]
[IR] Add a boolean field in DILocation to know if a line must covered or not

Summary:
Some lines have a hit counter where they should not have one.
For example, in C++, some cleanup is adding at the end of a scope represented by a '}'.
So such a line has a hit counter where a user expects to not have one.
The goal of the patch is to add this information in DILocation which is used to get the covered lines in GCOVProfiling.cpp.
A following patch in clang will add this information when generating IR (https://reviews.llvm.org/D49916).

Reviewers: marco-c, davidxl, vsk, javed.absar, rnk

Reviewed By: rnk

Subscribers: eraman, xur, danielcdh, aprantl, rnk, dblaikie, #debug-info, vsk, llvm-commits, sylvestre.ledru

Tags: #debug-info

Differential Revision: https://reviews.llvm.org/D49915

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342631 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[RISCV][MC] Improve parsing of jal/j operands
Alex Bradbury [Thu, 20 Sep 2018 08:10:35 +0000 (08:10 +0000)]
[RISCV][MC] Improve parsing of jal/j operands

Examples such as `jal a3`, `j a3` and `jal a3, a3` are accepted by gas
but rejected by LLVM MC. This patch rectifies this. I introduce
RISCVAsmParser::parseJALOffset to ensure that symbol names that coincide with
register names can safely be parsed. This is made a somewhat fiddly due to the
single-operand alias form (see the comment in parseJALOffset for more info).

Differential Revision: https://reviews.llvm.org/D52029

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342629 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC][x86][AArch64] Add BEXTR-like test patterns.
Roman Lebedev [Thu, 20 Sep 2018 07:54:49 +0000 (07:54 +0000)]
[NFC][x86][AArch64] Add BEXTR-like test patterns.

Summary: Also, adjust the check prefixes so that we actually get to check the BMI1-only-case.

Reviewers: craig.topper, RKSimon, spatel, javed.absar

Reviewed By: RKSimon

Subscribers: kristof.beyls, llvm-commits

Differential Revision: https://reviews.llvm.org/D48490

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342623 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoImprove the doc about the initial commit email sent to the ML
Sylvestre Ledru [Thu, 20 Sep 2018 07:43:24 +0000 (07:43 +0000)]
Improve the doc about the initial commit email sent to the ML

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342622 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[MachineVerifier] Relax checkLivenessAtDef regarding dead subreg defs
Bjorn Pettersson [Thu, 20 Sep 2018 06:59:18 +0000 (06:59 +0000)]
[MachineVerifier] Relax checkLivenessAtDef regarding dead subreg defs

Summary:
Consider an instruction that has multiple defs of the same
vreg, but defining different subregs:
  %7.sub1:rc, dead %7.sub2:rc = inst

Calling checkLivenessAtDef for the live interval associated
with %7 incorrectly reported "live range continues after a
dead def". The live range for %7 has a dead def at the slot
index for "inst" even if the live range continues (given that
there are later uses of %7.sub1).

This patch adjusts MachineVerifier::checkLivenessAtDef
to allow dead subregister definitions, unless we are checking
a subrange (when tracking subregister liveness).

A limitation is that we do not detect the situation when the
live range continues past an instruction that defines the
full virtual register by multiple dead subreg defines.

I also removed some dead code related to physical register
in checkLivenessAtDef. Wwe only call that method for virtual
registers, so I added an assertion instead.

Reviewers: kparzysz

Reviewed By: kparzysz

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D52237

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342618 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[XRay][compiler-rt] FDRLogWriter Abstraction
Dean Michael Berris [Thu, 20 Sep 2018 05:22:37 +0000 (05:22 +0000)]
[XRay][compiler-rt] FDRLogWriter Abstraction

Summary:
This change introduces an `FDRLogWriter` type which is responsible for
serialising metadata and function records to character buffers. This is
the first step in a refactoring of the implementation of the FDR runtime
to allow for more granular testing of the individual components of the
implementation.

The main contribution of this change is a means of hiding the details of
how specific records are written to a buffer, and for managing the
extents of these buffers. We make use of C++ features (templates and
some metaprogramming) to reduce repetition in the act of writing out
specific kinds of records to the buffer.

In this process, we make a number of changes across both LLVM and
compiler-rt to allow us to use the `Trace` abstraction defined in the
LLVM project in the testing of the runtime implementation. This gives us
a closer end-to-end test which version-locks the runtime implementation
with the loading implementation in LLVM.

We also allow using gmock in compiler-rt unit tests, by adding the
requisite definitions in the `AddCompilerRT.cmake` module. We also add
the terminfo library detection along with inclusion of the appropriate
compiler flags for header include lookup.

Finally, we've gone ahead and updated the FDR logging implementation to
use the FDRLogWriter for the lowest-level record-writing details.

Following patches will isolate the state machine transitions which
manage the set-up and tear-down of the buffers we're using in multiple
threads.

Reviewers: mboerger, eizan

Subscribers: mgorny, jfb, llvm-commits

Differential Revision: https://reviews.llvm.org/D52220

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342617 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoTemporarily Revert "[New PM] Introducing PassInstrumentation framework"
Eric Christopher [Thu, 20 Sep 2018 05:16:29 +0000 (05:16 +0000)]
Temporarily Revert "[New PM] Introducing PassInstrumentation framework"
as it was causing failures in the asan buildbot.

This reverts commit r342597.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342616 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix for bug 34002 - label generated before it block is finalized. Differential Revisi...
Maya Madhavan [Thu, 20 Sep 2018 05:11:42 +0000 (05:11 +0000)]
Fix for bug 34002 - label generated before it block is finalized. Differential Revision: https://reviews.llvm.org/D52258

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342615 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[unittests] Do not use llvm::sort in googlemock
Dean Michael Berris [Thu, 20 Sep 2018 04:27:32 +0000 (04:27 +0000)]
[unittests] Do not use llvm::sort in googlemock

Summary:
This reverts r329475 which applied to googlemock. This change makes the
googlemock implementation in LLVM dependent on LLVM unnecessarily.

Reviewers: echristo, mgrang

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D52287

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342612 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[PowerPC] Fix the assert of combineBVOfConsecutiveLoads when element num is 1
QingShan Zhang [Thu, 20 Sep 2018 03:09:15 +0000 (03:09 +0000)]
[PowerPC] Fix the assert of combineBVOfConsecutiveLoads when element num is 1
Building a vector out of multiple loads can be converted to a load of the vector type if the loads are consecutive.
But the special condition is that the element number is 1, such as <1 x i128>. So just early exit to fix the assert.

Patch By: wuzish (Zixuan Wu)
Differential Revision: https://reviews.llvm.org/D52072

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342611 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly] Renumber SIMD ops
Thomas Lively [Thu, 20 Sep 2018 02:55:28 +0000 (02:55 +0000)]
[WebAssembly] Renumber SIMD ops

Summary:
This change leaves holes in the opcode space where missing
instructions could logically be added later if they were found to be
useful.

Reviewers: aheejin, dschuff

Subscribers: sbc100, jgravelle-google, sunfish, llvm-commits

Differential Revision: https://reviews.llvm.org/D52282

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342610 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WEB] add new flags to a DebugInfo lit test
Michael Berg [Wed, 19 Sep 2018 22:56:14 +0000 (22:56 +0000)]
[WEB] add new flags to a DebugInfo lit test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342598 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[New PM] Introducing PassInstrumentation framework
Fedor Sergeev [Wed, 19 Sep 2018 22:42:57 +0000 (22:42 +0000)]
[New PM] Introducing PassInstrumentation framework

Pass Execution Instrumentation interface enables customizable instrumentation
of pass execution, as per "RFC: Pass Execution Instrumentation interface"
posted 06/07/2018 on llvm-dev@

The intent is to provide a common machinery to implement all
the pass-execution-debugging features like print-before/after,
opt-bisect, time-passes etc.

Here we get a basic implementation consisting of:
* PassInstrumentationCallbacks class that handles registration of callbacks
  and access to them.

* PassInstrumentation class that handles instrumentation-point interfaces
  that call into PassInstrumentationCallbacks.

* Callbacks accept StringRef which is just a name of the Pass right now.
  There were some ideas to pass an opaque wrapper for the pointer to pass instance,
  however it appears that pointer does not actually identify the instance
  (adaptors and managers might have the same address with the pass they govern).
  Hence it was decided to go simple for now and then later decide on what the proper
  mental model of identifying a "pass in a phase of pipeline" is.

* Callbacks accept llvm::Any serving as a wrapper for const IRUnit*, to remove direct dependencies
  on different IRUnits (e.g. Analyses).

* PassInstrumentationAnalysis analysis is explicitly requested from PassManager through
  usual AnalysisManager::getResult. All pass managers were updated to run that
  to get PassInstrumentation object for instrumentation calls.

* Using tuples/index_sequence getAnalysisResult helper to extract generic AnalysisManager's extra
  args out of a generic PassManager's extra args. This is the only way I was able to explicitly
  run getResult for PassInstrumentationAnalysis out of a generic code like PassManager::run or
  RepeatedPass::run.
  TODO: Upon lengthy discussions we agreed to accept this as an initial implementation
  and then get rid of getAnalysisResult by improving RepeatedPass implementation.

* PassBuilder takes PassInstrumentationCallbacks object to pass it further into
  PassInstrumentationAnalysis. Callbacks registration should be performed directly
  through PassInstrumentationCallbacks.

* new-pm tests updated to account for PassInstrumentationAnalysis being run

* Added PassInstrumentation tests to PassBuilderCallbacks unit tests.
  Other unit tests updated with registration of the now-required PassInstrumentationAnalysis.

Reviewers: chandlerc, philip.pfaffe
Differential Revision: https://reviews.llvm.org/D47858

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342597 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert "[XRay][compiler-rt] FDRLogWriter Abstraction" and 1 more.
Evgeniy Stepanov [Wed, 19 Sep 2018 22:29:56 +0000 (22:29 +0000)]
Revert "[XRay][compiler-rt] FDRLogWriter Abstraction" and 1 more.

Revert the following 2 commits to fix standalone compiler-rt build:
* r342523 [XRay] Detect terminfo library
* r342518 [XRay][compiler-rt] FDRLogWriter Abstraction

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342596 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[x86] add test for 256-bit andn (PR37749); NFC
Sanjay Patel [Wed, 19 Sep 2018 22:00:56 +0000 (22:00 +0000)]
[x86] add test for 256-bit andn (PR37749); NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342595 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[SelectionDAG] allow vector types with isBitwiseNot()
Sanjay Patel [Wed, 19 Sep 2018 21:48:30 +0000 (21:48 +0000)]
[SelectionDAG] allow vector types with isBitwiseNot()

The test diff in not-and-simplify.ll is from a use in SimplifyDemandedBits,
and the test diff in add.ll is from a DAGCombiner transform.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342594 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix the build when LLVM_TARGETS_TO_BUILD is empty.
Juergen Ributzka [Wed, 19 Sep 2018 21:36:26 +0000 (21:36 +0000)]
Fix the build when LLVM_TARGETS_TO_BUILD is empty.

Using LLVMTestingSupport in the LLVM_LINK_COMPONENTS breaks the build when
LLVM_TARGETS_TO_BUILD is set to empty.

Libraries that depend on LLVMTestingSupport need to use
target_link_libraries(<target> PRIVATE LLVMTestingSupport) instead.

This required change was already commited by r341899 to fix another build
issue.

This fixes rdar://problem/44615064.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342593 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[x86] add test for add+not vector fold; NFC
Sanjay Patel [Wed, 19 Sep 2018 21:28:39 +0000 (21:28 +0000)]
[x86] add test for add+not vector fold; NFC

The fold uses 'isBitwiseNot()', but that's not vector-friendly currently.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342592 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAArch64: Add FuseCryptoEOR fusion rules
Matthias Braun [Wed, 19 Sep 2018 20:50:51 +0000 (20:50 +0000)]
AArch64: Add FuseCryptoEOR fusion rules

There's some additional rules available on newer apple CPUs.

rdar://41235346

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342590 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoMachineScheduler: Add -misched-print-dags flag
Matthias Braun [Wed, 19 Sep 2018 20:50:49 +0000 (20:50 +0000)]
MachineScheduler: Add -misched-print-dags flag

Add a flag to dump the schedule DAG to the debug stream. This will be
used in upcoming commits to test schedule DAG mutations such as macro
fusion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342589 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoWorkaround a limitation of llvm::Any when used with types that have
Chandler Carruth [Wed, 19 Sep 2018 20:45:09 +0000 (20:45 +0000)]
Workaround a limitation of llvm::Any when used with types that have
a converting constructor from llvm::Any like gmock matchers. This issue
has come up elsewhere as well and the workaround here is being
considered for use in the standard long-term, but we can pretty cheaply
experiment with it to see if anything ends up going wrong.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342588 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ARM] Adjust the feature set for Exynos
Evandro Menezes [Wed, 19 Sep 2018 19:51:29 +0000 (19:51 +0000)]
[ARM] Adjust the feature set for Exynos

Fine tune the cost model for all Exynos processors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342585 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ARM] Refactor Exynos feature set (NFC)
Evandro Menezes [Wed, 19 Sep 2018 19:43:23 +0000 (19:43 +0000)]
[ARM] Refactor Exynos feature set (NFC)

Since all Exynos processors share the same feature set, fold them in the
implied fatures list for the subtarget.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342583 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] Disable strcmp->memcmp transform for MSan.
Matt Morehouse [Wed, 19 Sep 2018 19:37:24 +0000 (19:37 +0000)]
[InstCombine] Disable strcmp->memcmp transform for MSan.

Summary:
The strcmp->memcmp transform can make the resulting memcmp read
uninitialized data, which MSan doesn't like.

Resolves https://github.com/google/sanitizers/issues/993.

Reviewers: eugenis, xbolva00

Reviewed By: eugenis

Subscribers: hiraditya, llvm-commits

Differential Revision: https://reviews.llvm.org/D52272

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342582 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Handle COPYs of physregs better (regalloc hints)
Simon Pilgrim [Wed, 19 Sep 2018 18:59:08 +0000 (18:59 +0000)]
[X86] Handle COPYs of physregs better (regalloc hints)

Enable enableMultipleCopyHints() on X86.

Original Patch by @jonpa:

While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling.

Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates.

Differential Revision: https://reviews.llvm.org/D38128

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342578 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoCopy utilities updated and added for MI flags
Michael Berg [Wed, 19 Sep 2018 18:52:08 +0000 (18:52 +0000)]
Copy utilities updated and added for MI flags

Summary: This patch adds a GlobalIsel copy utility into MI for flags and updates the instruction emitter for the SDAG path.  Some tests show new behavior and I added one for GlobalIsel which mirrors an SDAG test for handling nsw/nuw.

Reviewers: spatel, wristow, arsenm

Reviewed By: arsenm

Subscribers: wdng

Differential Revision: https://reviews.llvm.org/D52006

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342576 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[x86] change names of vector splitting helper functions; NFC
Sanjay Patel [Wed, 19 Sep 2018 18:52:00 +0000 (18:52 +0000)]
[x86] change names of vector splitting helper functions; NFC

As the code comments suggest, these are about splitting, and they
are not necessarily limited to lowering, so that misled me.

There's nothing that's actually x86-specific in these either, so
they might be better placed in a common header so any target can
use them.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342575 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ThinLTO] Write TYPE_IDs for types used in functions imported by aliases
Vitaly Buka [Wed, 19 Sep 2018 18:51:42 +0000 (18:51 +0000)]
[ThinLTO] Write TYPE_IDs for types used in functions imported by aliases

Summary:
ThinLTO imports alias as a copy of a aliasee, so when we import such functions with type tests we will
need type ids used by function. However after D49565 we pick types only during processing of
FunctionSummary which is not happening for such aliesees.

Example:
Unit U1 with a type, a functions F with the type check, and an alias A to the function.
Unit U2 with only call to the alias A.

In particular, this happens when we use -mconstructor-aliases, which is default.
So if c++ unit only creates instance of the class, without calling any other methods it will lack of
necessary type ids, which will result in false CFI reports.

Reviewers: tejohnson, eugenis

Subscribers: pcc, mehdi_amini, inglorion, eraman, hiraditya, steven_wu, dexonsmith, llvm-commits

Differential Revision: https://reviews.llvm.org/D52201

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342574 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ThinLTO] Extract getReferencedTypeIds from [NFC]
Vitaly Buka [Wed, 19 Sep 2018 18:51:25 +0000 (18:51 +0000)]
[ThinLTO] Extract getReferencedTypeIds from [NFC]

Summary: Refactoring before D52201

Reviewers: eugenis, tejohnson

Subscribers: pcc, inglorion, hiraditya, dexonsmith, llvm-commits

Differential Revision: https://reviews.llvm.org/D52203

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342573 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[mips][microMIPS] Extending size reduction pass with MOVEP
Simon Atanasyan [Wed, 19 Sep 2018 18:46:29 +0000 (18:46 +0000)]
[mips][microMIPS] Extending size reduction pass with MOVEP

The patch extends size reduction pass for MicroMIPS. Two MOVE
instructions are transformed into one MOVEP instrucition.

Patch by Milena Vujosevic Janicic.

Differential revision: https://reviews.llvm.org/D52037

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342572 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[mips][microMIPS] Fix the definition of MOVEP instruction
Simon Atanasyan [Wed, 19 Sep 2018 18:46:21 +0000 (18:46 +0000)]
[mips][microMIPS] Fix the definition of MOVEP instruction

The patch fixes definition of MOVEP instruction. Two registers are used
instead of register pairs. This is necessary as machine verifier cannot
handle register pairs.

Patch by Milena Vujosevic Janicic.

Differential revision: https://reviews.llvm.org/D52035

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342571 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[LLJIT] Add API to expose linking layer from LLJIT classes
Chris Bieneman [Wed, 19 Sep 2018 18:29:58 +0000 (18:29 +0000)]
[LLJIT] Add API to expose linking layer from LLJIT classes

Summary: This patch just surfaces the object linking layer from the LLJIT classes so that clients can take advantage of the changes implemented in r341154.

Reviewers: lhames

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D51551

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342567 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86] Add initial SimplifyDemandedVectorEltsForTargetNode support
Simon Pilgrim [Wed, 19 Sep 2018 18:11:34 +0000 (18:11 +0000)]
[X86] Add initial SimplifyDemandedVectorEltsForTargetNode support

This patch adds an initial x86 SimplifyDemandedVectorEltsForTargetNode implementation to handle target shuffles.

Currently the patch only decodes a target shuffle, calls SimplifyDemandedVectorElts on its input operands and removes any shuffle that reduces to undef/zero/identity.

Future work will need to integrate this with combineX86ShufflesRecursively, add support for other x86 ops, etc.

NOTE: There is a minor regression that appears to be affecting further (extractelement?) combines which I haven't been able to solve yet - possibly something to do with how nodes are added to the worklist after simplification.

Differential Revision: https://reviews.llvm.org/D52140

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342564 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAttempt to unbreak buidlbot lld-x86_64-darwin13 after r342555.
Andrea Di Biagio [Wed, 19 Sep 2018 17:54:01 +0000 (17:54 +0000)]
Attempt to unbreak buidlbot lld-x86_64-darwin13 after r342555.

The reason why build #25777 might have failed is because the SmallVector move
constructor is _not_ noexcept, and the stl implementation used by that buildbot
calls _VSTD::move_if_noexcept() (according to the backtrace).

OpcodeInfo has a default move constructor, and the copy constructor is deleted.
However, as far as I can see, SmallVector doesn't declare a noexcept move
constructor. So, what I believe it is happening here is that,
_VSTD::move_if_noexcept() returns an lvalue reference and not an rvalue
reference.
This eventually triggers a copy that fails to compile.

Hopefully, using a std::vector instead of SmallVector (as it was originally
suggested by Simon in the code review) should be enough to unbreak the buildbot.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342561 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AMDGPU] Add instruction selection for i1 to f16 conversion
Carl Ritson [Wed, 19 Sep 2018 16:32:12 +0000 (16:32 +0000)]
[AMDGPU] Add instruction selection for i1 to f16 conversion

Summary:
This is required for GPUs with 16 bit instructions where f16 is a
legal register type and hence int_to_fp i1 to f16 is not lowered
by legalizing.

Reviewers: arsenm, nhaehnle

Reviewed By: nhaehnle

Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, llvm-commits

Differential Revision: https://reviews.llvm.org/D52018

Change-Id: Ie4c0fd6ced7cf10ad612023c6879724d9ded5851

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342558 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[bpf] Symbol sizes and types in object file
Yonghong Song [Wed, 19 Sep 2018 16:04:13 +0000 (16:04 +0000)]
[bpf] Symbol sizes and types in object file

Clang-compiled object files currently don't include the symbol sizes and
types.  Some tools however need that information.  For example, ctfconvert
uses that information to generate FreeBSD's CTF representation from ELF
files.
With this patch, symbol sizes and types are included in object files.

Signed-off-by: Paul Chaignon <paul.chaignon@orange.com>
Reported-by: Yutaro Hayakawa <yhayakawa3720@gmail.com>
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342556 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[TableGen][SubtargetEmitter] Add the ability for processor models to describe depende...
Andrea Di Biagio [Wed, 19 Sep 2018 15:57:45 +0000 (15:57 +0000)]
[TableGen][SubtargetEmitter] Add the ability for processor models to describe dependency breaking instructions.

This patch adds the ability for processor models to describe dependency breaking
instructions.

Different processors may specify a different set of dependency-breaking
instructions.
That means, we cannot assume that all processors of the same target would use
the same rules to classify dependency breaking instructions.

The main goal of this patch is to provide the means to describe dependency
breaking instructions directly via tablegen, and have the following
TargetSubtargetInfo hooks redefined in overrides by tabegen'd
XXXGenSubtargetInfo classes (here, XXX is a Target name).

```
virtual bool isZeroIdiom(const MachineInstr *MI, APInt &Mask) const {
  return false;
}

virtual bool isDependencyBreaking(const MachineInstr *MI, APInt &Mask) const {
  return isZeroIdiom(MI);
}
```

An instruction MI is a dependency-breaking instruction if a call to method
isDependencyBreaking(MI) on the STI (TargetSubtargetInfo object) evaluates to
true. Similarly, an instruction MI is a special case of zero-idiom dependency
breaking instruction if a call to STI.isZeroIdiom(MI) returns true.
The extra APInt is used for those targets that may want to select which machine
operands have their dependency broken (see comments in code).
Note that by default, subtargets don't know about the existence of
dependency-breaking. In the absence of external information, those method calls
would always return false.

A new tablegen class named STIPredicate has been added by this patch to let
processor models classify instructions that have properties in common. The idea
is that, a MCInstrPredicate definition can be used to "generate" an instruction
equivalence class, with the idea that instructions of a same class all have a
property in common.

STIPredicate definitions are essentially a collection of instruction equivalence
classes.
Also, different processor models can specify a different variant of the same
STIPredicate with different rules (i.e. predicates) to classify instructions.
Tablegen backends (in this particular case, the SubtargetEmitter) will be able
to process STIPredicate definitions, and automatically generate functions in
XXXGenSubtargetInfo.

This patch introduces two special kind of STIPredicate classes named
IsZeroIdiomFunction and IsDepBreakingFunction in tablegen. It also adds a
definition for those in the BtVer2 scheduling model only.

This patch supersedes the one committed at r338372 (phabricator review: D49310).

The main advantages are:
 - We can describe subtarget predicates via tablegen using STIPredicates.
 - We can describe zero-idioms / dep-breaking instructions directly via
   tablegen in the scheduling models.

In future, the STIPredicates framework can be used for solving other problems.
Examples of future developments are:
 - Teach how to identify optimizable register-register moves
 - Teach how to identify slow LEA instructions (each subtarget defining its own
   concept of "slow" LEA).
 - Teach how to identify instructions that have undocumented false dependencies
   on the output registers on some processors only.

It is also (in my opinion) an elegant way to expose knowledge to both external
tools like llvm-mca, and codegen passes.
For example, machine schedulers in LLVM could reuse that information when
internally constructing the data dependency graph for a code region.

This new design feature is also an "opt-in" feature. Processor models don't have
to use the new STIPredicates. It has all been designed to be as unintrusive as
possible.

Differential Revision: https://reviews.llvm.org/D52174

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342555 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DAGCombiner][x86] add transform/hook to decompose integer multiply into shift/add
Sanjay Patel [Wed, 19 Sep 2018 15:57:40 +0000 (15:57 +0000)]
[DAGCombiner][x86] add transform/hook to decompose integer multiply into shift/add

This is an alternative to D37896. I don't see a way to decompose multiplies
generically without a target hook to tell us when it's profitable.

ARM and AArch64 may be able to remove some duplicate code that overlaps with
this transform.

As a first step, we're only getting the most clear wins on the vector examples
requested in PR34474:
https://bugs.llvm.org/show_bug.cgi?id=34474

As noted in the code comment, it's likely that the x86 constraints are tighter
than necessary, but it may not always be a win to replace a pmullw/pmulld.

Differential Revision: https://reviews.llvm.org/D52195

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342554 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert rL342544: [New PM] Introducing PassInstrumentation framework
Fedor Sergeev [Wed, 19 Sep 2018 14:54:48 +0000 (14:54 +0000)]
Revert rL342544: [New PM] Introducing PassInstrumentation framework

A bunch of bots fail to compile unittests. Reverting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342552 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix -Wsign-compare warnings. NFCI
Simon Pilgrim [Wed, 19 Sep 2018 14:52:34 +0000 (14:52 +0000)]
Fix -Wsign-compare warnings. NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342551 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AtomicExpandPass]: Add a hook for custom cmpxchg expansion in IR
Alex Bradbury [Wed, 19 Sep 2018 14:51:42 +0000 (14:51 +0000)]
[AtomicExpandPass]: Add a hook for custom cmpxchg expansion in IR

This involves changing the shouldExpandAtomicCmpXchgInIR interface, but I have
updated the in-tree backends using this hook (ARM, AArch64, Hexagon) so they
will see no functional change. Previously this hook returned bool, but it now
returns AtomicExpansionKind.

This hook allows targets to select how a given cmpxchg is to be expanded.
D48131 uses this to expand part-word cmpxchg to a target-specific intrinsic.

See my associated RFC for more info on the motivation for this change
<http://lists.llvm.org/pipermail/llvm-dev/2018-June/123993.html>.

Differential Revision: https://reviews.llvm.org/D48130

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342550 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[benchmark] Cherrypick fix for MinGW/ARM from upstream
Martin Storsjo [Wed, 19 Sep 2018 14:30:45 +0000 (14:30 +0000)]
[benchmark] Cherrypick fix for MinGW/ARM from upstream

This fixes building for Windows on ARM, with MinGW headers.
(Building for Windows on ARM with Windows SDK still is unsupported
by the benchmark library.)

Differential Revision: https://reviews.llvm.org/D52262

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342549 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] foldICmpWithLowBitMaskedVal(): handle uncanonical ((-1 << y) >> y)...
Roman Lebedev [Wed, 19 Sep 2018 13:35:46 +0000 (13:35 +0000)]
[InstCombine] foldICmpWithLowBitMaskedVal(): handle uncanonical ((-1 << y) >> y) mask

Summary:
The last low-bit-mask-pattern-producing-pattern i can think of.

https://rise4fun.com/Alive/UGzE <- non-canonical
But we can not canonicalize it because of extra uses.

https://bugs.llvm.org/show_bug.cgi?id=38123

Reviewers: spatel, craig.topper, RKSimon

Reviewed By: spatel

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D52148

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342548 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] foldICmpWithLowBitMaskedVal(): handle uncanonical ((1 << y)+(-1)) mask
Roman Lebedev [Wed, 19 Sep 2018 13:35:40 +0000 (13:35 +0000)]
[InstCombine] foldICmpWithLowBitMaskedVal(): handle uncanonical ((1 << y)+(-1)) mask

Summary:
Same as to D52146.
`((1 << y)+(-1))` is simply non-canoniacal version of `~(-1 << y)`: https://rise4fun.com/Alive/0vl
We can not canonicalize it due to the extra uses. But we can handle it here.

Reviewers: spatel, craig.topper, RKSimon

Reviewed By: spatel

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D52147

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342547 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] foldICmpWithLowBitMaskedVal(): handle ~(-1 << y) mask
Roman Lebedev [Wed, 19 Sep 2018 13:35:27 +0000 (13:35 +0000)]
[InstCombine] foldICmpWithLowBitMaskedVal(): handle  ~(-1 << y)  mask

Summary:
Two folds are happening here:
1. https://rise4fun.com/Alive/oaFX
2. And then `foldICmpWithHighBitMask()` (D52001): https://rise4fun.com/Alive/wsP4

This change doesn't just add the handling for eq/ne predicates,
it actually builds upon the previous `foldICmpWithLowBitMaskedVal()` work,
so **all** the 16 fold variants* are immediately supported.

I'm indeed only testing these two predicates.
I do not feel like re-proving all 16 folds*, because they were already proven
for the general case of constant with all-ones in low bits. So as long as
the mask produces all-ones in low bits, i'm pretty sure the fold is valid.

But required, i can re-prove, let me know.

* eq/ne are commutative - 4 folds; ult/ule/ugt/uge - are not commutative (the commuted variant is InstSimplified), 4 folds; slt/sle/sgt/sge are not commutative - 4 folds. 12 folds in total.

https://bugs.llvm.org/show_bug.cgi?id=38123
https://bugs.llvm.org/show_bug.cgi?id=38708

Reviewers: spatel, craig.topper, RKSimon

Reviewed By: spatel

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D52146

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342546 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ARM] Fix unwind information for floating point registers
Oliver Stannard [Wed, 19 Sep 2018 13:25:31 +0000 (13:25 +0000)]
[ARM] Fix unwind information for floating point registers

Fixes the unwind information generated for floating-point registers.
Previously, all padding registers were assumed to be four bytes wide. Now, the
width of the register is used to specify the amount of padding.

Patch by Jackson Woodruff!

Differential revision: https://reviews.llvm.org/D51494

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342545 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[New PM] Introducing PassInstrumentation framework
Fedor Sergeev [Wed, 19 Sep 2018 12:25:52 +0000 (12:25 +0000)]
[New PM] Introducing PassInstrumentation framework

Summary:
Pass Execution Instrumentation interface enables customizable instrumentation
of pass execution, as per "RFC: Pass Execution Instrumentation interface"
posted 06/07/2018 on llvm-dev@

The intent is to provide a common machinery to implement all
the pass-execution-debugging features like print-before/after,
opt-bisect, time-passes etc.

Here we get a basic implementation consisting of:
* PassInstrumentationCallbacks class that handles registration of callbacks
  and access to them.

* PassInstrumentation class that handles instrumentation-point interfaces
  that call into PassInstrumentationCallbacks.

* Callbacks accept StringRef which is just a name of the Pass right now.
  There were some ideas to pass an opaque wrapper for the pointer to pass instance,
  however it appears that pointer does not actually identify the instance
  (adaptors and managers might have the same address with the pass they govern).
  Hence it was decided to go simple for now and then later decide on what the proper
  mental model of identifying a "pass in a phase of pipeline" is.

* Callbacks accept llvm::Any serving as a wrapper for const IRUnit*, to remove direct dependencies
  on different IRUnits (e.g. Analyses).

* PassInstrumentationAnalysis analysis is explicitly requested from PassManager through
  usual AnalysisManager::getResult. All pass managers were updated to run that
  to get PassInstrumentation object for instrumentation calls.

* Using tuples/index_sequence getAnalysisResult helper to extract generic AnalysisManager's extra
  args out of a generic PassManager's extra args. This is the only way I was able to explicitly
  run getResult for PassInstrumentationAnalysis out of a generic code like PassManager::run or
  RepeatedPass::run.
  TODO: Upon lengthy discussions we agreed to accept this as an initial implementation
  and then get rid of getAnalysisResult by improving RepeatedPass implementation.

* PassBuilder takes PassInstrumentationCallbacks object to pass it further into
  PassInstrumentationAnalysis. Callbacks registration should be performed directly
  through PassInstrumentationCallbacks.

* new-pm tests updated to account for PassInstrumentationAnalysis being run

* Added PassInstrumentation tests to PassBuilderCallbacks unit tests.
  Other unit tests updated with registration of the now-required PassInstrumentationAnalysis.

Reviewers: chandlerc, philip.pfaffe
Differential Revision: https://reviews.llvm.org/D47858

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342544 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[TableGen] CodeGenDAGPatterns::GenerateVariants - use BitVector::set() where possible...
Simon Pilgrim [Wed, 19 Sep 2018 12:23:50 +0000 (12:23 +0000)]
[TableGen] CodeGenDAGPatterns::GenerateVariants - use BitVector::set() where possible. NFCI.

In these cases we are always setting to true.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342543 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[InstCombine] Don't transform sin/cos -> tanl if for half types
Benjamin Kramer [Wed, 19 Sep 2018 12:01:38 +0000 (12:01 +0000)]
[InstCombine] Don't transform sin/cos -> tanl if for half types

This is still unsafe for long double, we will transform things into tanl
even if tanl is for another type. But that's for someone else to fix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342542 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoVerify commit access in fixing typo
Calixte Denizet [Wed, 19 Sep 2018 11:26:20 +0000 (11:26 +0000)]
Verify commit access in fixing typo

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342538 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[TableGen] CodeGenDAGPatterns::GenerateVariants - use BitVector::push_back()
Simon Pilgrim [Wed, 19 Sep 2018 11:18:49 +0000 (11:18 +0000)]
[TableGen] CodeGenDAGPatterns::GenerateVariants - use BitVector::push_back()

As mentioned on D52236, this saves 10secs in debug builds of x86 -gen-dag-isel

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342536 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[ADT][BitVector] Add push_back()
Simon Pilgrim [Wed, 19 Sep 2018 11:08:54 +0000 (11:08 +0000)]
[ADT][BitVector] Add push_back()

Add a higher performance alternative to calling resize() every time which performs a lot of clearing to zero - when we're adding a single bit most of the time this will be completely unnecessary.

Differential Revision: https://reviews.llvm.org/D52236

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342535 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[RISCV] Codegen for i8, i16, and i32 atomicrmw with RV32A
Alex Bradbury [Wed, 19 Sep 2018 10:54:22 +0000 (10:54 +0000)]
[RISCV] Codegen for i8, i16, and i32 atomicrmw with RV32A

Introduce a new RISCVExpandPseudoInsts pass to expand atomic
pseudo-instructions after register allocation. This is necessary in order to
ensure that register spills aren't introduced between LL and SC, thus breaking
the forward progress guarantee for the operation. AArch64 does something
similar for CmpXchg (though only at O0), and Mips is moving towards this
approach (see D31287). See also [this mailing list
post](http://lists.llvm.org/pipermail/llvm-dev/2016-May/099490.html) from
James Knight, which summarises the issues with lowering to ll/sc in IR or
pre-RA.

See the [accompanying RFC
thread](http://lists.llvm.org/pipermail/llvm-dev/2018-June/123993.html) for an
overview of the lowering strategy.

Differential Revision: https://reviews.llvm.org/D47882

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342534 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[COFF] Emit @feat.00 on 64-bit and set the CFG bit when emitting guardcf tables
Hans Wennborg [Wed, 19 Sep 2018 09:58:30 +0000 (09:58 +0000)]
[COFF] Emit @feat.00 on 64-bit and set the CFG bit when emitting guardcf tables

The 0x800 bit in @feat.00 needs to be set in order to make LLD pick up
the .gfid$y table. I believe this is fine to set even if we don't emit
the instrumentation.

We haven't emitted @feat.00 on 64-bit before. I see that MSVC does emit
it, but I'm not entirely sure what the default value should be. I went
with zero since that seems as safe as not emitting the symbol in the
first place.

Differential Revision: https://reviews.llvm.org/D52235

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342532 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[X86][SSE] Update extractelement test in preparation for D52140
Simon Pilgrim [Wed, 19 Sep 2018 09:50:32 +0000 (09:50 +0000)]
[X86][SSE] Update extractelement test in preparation for D52140

SimplifyDemandedVectorEltsForTargetNode will remove most of this test unless get rid of the undefs - still testing for align 1 which was the point of the test

Removed out of date comment as well

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342531 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DebugInfo][Dexter] Speculated BB presents illegal variable value to debugger.
Carlos Alberto Enciso [Wed, 19 Sep 2018 08:16:56 +0000 (08:16 +0000)]
[DebugInfo][Dexter] Speculated BB presents illegal variable value to debugger.

When SimplifyCFG changes the PHI node into a select instruction, the debug information becomes ambiguous. It causes the debugger to display wrong variable value.

Differential Revision: https://reviews.llvm.org/D51976

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342527 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[DWARF Verifier] Add helper function to dump DIEs. [NFC]
Jonas Devlieghere [Wed, 19 Sep 2018 08:08:13 +0000 (08:08 +0000)]
[DWARF Verifier] Add helper function to dump DIEs. [NFC]

It's pretty common for the verifier to dump the relevant DIE when it
finds an issue. This tends to be relatively verbose and error prone
because we have to pass the DIDumpOptions to the DIE's dump method. This
patch adds a helper function to the verifier to make this easier.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342526 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly][NFC] Remove extra space in WebAssemblyInstrSIMD.td
Thomas Lively [Wed, 19 Sep 2018 00:54:20 +0000 (00:54 +0000)]
[WebAssembly][NFC] Remove extra space in WebAssemblyInstrSIMD.td

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342522 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoAArch64MacroFusion: Factor out some opcode handling code; NFC
Matthias Braun [Wed, 19 Sep 2018 00:23:37 +0000 (00:23 +0000)]
AArch64MacroFusion: Factor out some opcode handling code; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342521 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoScheduleDAG: Cleanup dumping code; NFC
Matthias Braun [Wed, 19 Sep 2018 00:23:35 +0000 (00:23 +0000)]
ScheduleDAG: Cleanup dumping code; NFC

- Instead of having both `SUnit::dump(ScheduleDAG*)` and
  `ScheduleDAG::dumpNode(ScheduleDAG*)`, just keep the latter around.
- Add `ScheduleDAG::dump()` and avoid code duplication in several
  places. Implement it for different ScheduleDAG variants.
- Add `ScheduleDAG::dumpNodeName()` in favor of the `SUnit::print()`
  functions. They were only ever used for debug dumping and putting the
  function into ScheduleDAG is consistent with the `dumpNode()` change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342520 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[XRay][compiler-rt] FDRLogWriter Abstraction
Dean Michael Berris [Tue, 18 Sep 2018 23:59:32 +0000 (23:59 +0000)]
[XRay][compiler-rt] FDRLogWriter Abstraction

Summary:
This change introduces an `FDRLogWriter` type which is responsible for
serialising metadata and function records to character buffers. This is
the first step in a refactoring of the implementation of the FDR runtime
to allow for more granular testing of the individual components of the
implementation.

The main contribution of this change is a means of hiding the details of
how specific records are written to a buffer, and for managing the
extents of these buffers. We make use of C++ features (templates and
some metaprogramming) to reduce repetition in the act of writing out
specific kinds of records to the buffer.

In this process, we make a number of changes across both LLVM and
compiler-rt to allow us to use the `Trace` abstraction defined in the
LLVM project in the testing of the runtime implementation. This gives us
a closer end-to-end test which version-locks the runtime implementation
with the loading implementation in LLVM.

We also allow using gmock in compiler-rt unit tests, by adding the
requisite definitions in the `AddCompilerRT.cmake` module.

Finally, we've gone ahead and updated the FDR logging implementation to
use the FDRLogWriter for the lowest-level record-writing details.

Following patches will isolate the state machine transitions which
manage the set-up and tear-down of the buffers we're using in multiple
threads.

Reviewers: mboerger, eizan

Subscribers: mgorny, jfb, llvm-commits

Differential Revision: https://reviews.llvm.org/D52220

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342518 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[WebAssembly] v4f32.abs and v2f64.abs
Thomas Lively [Tue, 18 Sep 2018 21:45:12 +0000 (21:45 +0000)]
[WebAssembly] v4f32.abs and v2f64.abs

Summary: implement lowering of @llvm.fabs for vector types.

Reviewers: aheejin, dschuff

Subscribers:

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342513 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert r342494 as it was failing on a bot and the author cannot look at it until...
Douglas Yung [Tue, 18 Sep 2018 19:34:05 +0000 (19:34 +0000)]
Revert r342494 as it was failing on a bot and the author cannot look at it until tomorrow.

Failing bot: http://lab.llvm.org:8011/builders/llvm-clang-lld-x86_64-scei-ps4-ubuntu-fast/builds/36708

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342509 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[benchmark] Mention another cherry-picked change in README.LLVM. NFC.
Martin Storsjo [Tue, 18 Sep 2018 19:31:09 +0000 (19:31 +0000)]
[benchmark] Mention another cherry-picked change in README.LLVM. NFC.

This was cherry-picked in SVN r342450.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342506 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[bugpoint] Revert r318459
Don Hinton [Tue, 18 Sep 2018 18:39:27 +0000 (18:39 +0000)]
[bugpoint] Revert r318459

Summary: Revert r318459 which introduced a TempFile scoping bug.

Differential Revision: https://reviews.llvm.org/D51836

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342503 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRemove svn:executable from a documentation file.
Kristina Brooks [Tue, 18 Sep 2018 18:37:16 +0000 (18:37 +0000)]
Remove svn:executable from a documentation file.

Should fix the buildbot failure caused by it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342502 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[Docs] [Support] System Library to Support Library transition along with minor correc...
Kristina Brooks [Tue, 18 Sep 2018 18:05:38 +0000 (18:05 +0000)]
[Docs] [Support] System Library to Support Library transition along with minor corrections to reflect it.

System Library has been a long deprecated term along with the path lib/System, having been superseded/renamed
to the Support Library a long time ago. These patches reflect those changes in documentation as well as
update some outdated examples and provide context to the origin of the Support Library.

Differential Revision: https://reviews.llvm.org/D52107

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342500 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoDo not optimize atomic load to non-atomic memcmp
Christy Lee [Tue, 18 Sep 2018 17:02:42 +0000 (17:02 +0000)]
Do not optimize atomic load to non-atomic memcmp

Differential Revision: https://reviews.llvm.org/D51998

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342498 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[AMDGPU] Match udot8 pattern
Farhana Aleen [Tue, 18 Sep 2018 16:59:48 +0000 (16:59 +0000)]
[AMDGPU] Match udot8 pattern

Summary: D.u32 = S0.u4[0] * S1.u4[0] +

         S0.u4[1] * S1.u4[1] +
         S0.u4[2] * S1.u4[2] +
         S0.u4[3] * S1.u4[3] +
         S0.u4[4] * S1.u4[4] +
         S0.u4[5] * S1.u4[5] +
         S0.u4[6] * S1.u4[6] +
         S0.u4[7] * S1.u4[7] +
         S2.u32

Author: FarhanaAleen

Reviewed By: arsenm, nhaehnle

Differential Revision: https://reviews.llvm.org/D51947

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342497 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoFix function signature to actually be an override.
Zachary Turner [Tue, 18 Sep 2018 16:58:17 +0000 (16:58 +0000)]
Fix function signature to actually be an override.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342496 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[PGO][CHR] Add opt remarks.
Hiroshi Yamauchi [Tue, 18 Sep 2018 16:50:10 +0000 (16:50 +0000)]
[PGO][CHR] Add opt remarks.

Reviewers: davidxl

Reviewed By: davidxl

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D52056

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342495 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoCheck lines before using alias analysis to check for interference
Christy Lee [Tue, 18 Sep 2018 16:43:44 +0000 (16:43 +0000)]
Check lines before using alias analysis to check for interference

This diff is to show the difference before and after D51550

Differential Revision: https://reviews.llvm.org/D52044

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342494 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[PDB] Better support for enumerating pointer types.
Zachary Turner [Tue, 18 Sep 2018 16:35:05 +0000 (16:35 +0000)]
[PDB] Better support for enumerating pointer types.

There were several issues with the previous implementation.

1) There were no tests.
2) We didn't support creating PDBSymbolTypePointer records for
   builtin types since those aren't described by LF_POINTER
   records.
3) We didn't support a wide enough variety of builtin types even
   ignoring pointers.

This patch fixes all of these issues.  In order to add tests,
it's helpful to be able to ignore the symbol index id hierarchy
because it makes the golden output from the DIA version not match
our output, so I've extended the dumper to disable dumping of id
fields.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342493 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[PostRASink] Make sure to remove subregisters from live-ins as well
Krzysztof Parzyszek [Tue, 18 Sep 2018 16:10:51 +0000 (16:10 +0000)]
[PostRASink] Make sure to remove subregisters from live-ins as well

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342492 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[NFC] Update comments regarding BufferSize for ProcResources
Jinsong Ji [Tue, 18 Sep 2018 15:38:56 +0000 (15:38 +0000)]
[NFC] Update comments regarding BufferSize for ProcResources

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342491 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert rL342465: Added function to set a register to a particular value + tests.
Simon Pilgrim [Tue, 18 Sep 2018 15:38:16 +0000 (15:38 +0000)]
Revert rL342465: Added function to set a register to a particular value + tests.

rL342465 is breaking the MSVC buildbots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342490 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert rL342466: [llvm-exegesis] Improve Register Setup.
Simon Pilgrim [Tue, 18 Sep 2018 15:35:49 +0000 (15:35 +0000)]
Revert rL342466: [llvm-exegesis] Improve Register Setup.

rL342465 is breaking the MSVC buildbots, but I need to revert this dependent revision as well.

Summary:
Added function to set a register to a particular value + tests.
Add EFLAGS test, use new setRegTo instead of setRegToConstant.

Reviewers: courbet, javed.absar

Subscribers: mgorny, tschuett, llvm-commits

Differential Revision: https://reviews.llvm.org/D51856

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342489 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[RISCV][MC] Use a custom ParserMethod for the bare_symbol operand type
Alex Bradbury [Tue, 18 Sep 2018 15:18:16 +0000 (15:18 +0000)]
[RISCV][MC] Use a custom ParserMethod for the bare_symbol operand type

This allows the hard-coded shouldForceImmediate logic to be removed because
the generated MatchOperandParserImpl makes use of the current context (i.e.
the current mnemonic) to determine parsing behaviour, and so won't first try
to parse a register before parsing a symbol name.

No functional change is intended. gas accepts immediate arguments for call,
tail and lla. This patch doesn't address this discrepancy.

Differential Revision: https://reviews.llvm.org/D51733

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342488 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[RISCV][MC] Reject bare symbols for the simm12 operand type
Alex Bradbury [Tue, 18 Sep 2018 15:13:29 +0000 (15:13 +0000)]
[RISCV][MC] Reject bare symbols for the simm12 operand type

addi a0, a0, foo and lw a0, foo(a0) and similar are now rejected. An explicit
%lo and %pcrel_lo modifier is required. This matches gas behaviour.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342487 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[RISCV][MC] Tighten up checking of sybol operands to lui and auipc
Alex Bradbury [Tue, 18 Sep 2018 15:08:35 +0000 (15:08 +0000)]
[RISCV][MC] Tighten up checking of sybol operands to lui and auipc

Reject bare symbols and accept only %pcrel_hi(sym) for auipc and %hi(sym) for
lui. Also test valid operand modifiers in rv32i-valid.s.

Note this is slightly stricter than gas, which will accept either %pcrel_hi or
%hi for both lui and auipc.

Differential Revision: https://reviews.llvm.org/D51731

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342486 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRemove dead function user_cache_directory()
Nico Weber [Tue, 18 Sep 2018 15:06:16 +0000 (15:06 +0000)]
Remove dead function user_cache_directory()

It's been unused since it was added almost 3 years ago in
https://reviews.llvm.org/D13801

Motivated by https://reviews.llvm.org/rL342002 since it removes one of the
functions keeping a ref to SHGetKnownFolderPath.

Differential Revision: https://reviews.llvm.org/D52184

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342485 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-mca] Add the ability to mark register reads/writes associated with dep-breaking...
Andrea Di Biagio [Tue, 18 Sep 2018 15:00:06 +0000 (15:00 +0000)]
[llvm-mca] Add the ability to mark register reads/writes associated with dep-breaking instructions. NFCI

This patch adds two new boolean fields:
  - Field `ReadState::IndependentFromDef`.
  - Field `WriteState::WritesZero`.

Field `IndependentFromDef` is set for ReadState objects associated with
dependency-breaking instructions. It is used by the simulator when updating data
dependencies between registers.

Field `WritesZero` is set by WriteState objects associated with dependency
breaking zero-idiom instructions. It helps the PRF identify which writes don't
consume any physical registers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342483 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoRevert r342457 "Fixes removal of dead elements from PressureDiff (PR37252)."
Hans Wennborg [Tue, 18 Sep 2018 14:12:54 +0000 (14:12 +0000)]
Revert r342457 "Fixes removal of dead elements from PressureDiff (PR37252)."

This broke the lit tests on a bunch of buildbots, e.g.
http://lab.llvm.org:8011/builders/llvm-clang-lld-x86_64-scei-ps4-ubuntu-fast/builds/36679

> Reviewed By: MatzeB
>
> Differential Revision: https://reviews.llvm.org/D51495

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342482 91177308-0d34-0410-b5e6-96231b3b80d8

5 years agoUse pass-by-reference for-range loop. NFCI.
Simon Pilgrim [Tue, 18 Sep 2018 14:05:07 +0000 (14:05 +0000)]
Use pass-by-reference for-range loop. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342481 91177308-0d34-0410-b5e6-96231b3b80d8

5 years ago[llvm-mca] Slightly refactor class InstRef. NFC.
Andrea Di Biagio [Tue, 18 Sep 2018 14:03:46 +0000 (14:03 +0000)]
[llvm-mca] Slightly refactor class InstRef. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@342480 91177308-0d34-0410-b5e6-96231b3b80d8