OSDN Git Service

android-x86/external-llvm.git
4 years ago[llvm-readobj] Allow --hex-dump/--string-dump to dump multiple sections
Fangrui Song [Tue, 18 Jun 2019 14:01:03 +0000 (14:01 +0000)]
[llvm-readobj] Allow --hex-dump/--string-dump to dump multiple sections

1) `-x foo` currently dumps one `foo`. This change makes it dump all `foo`.
2) `-x foo -x foo` currently dumps `foo` twice. This change makes it dump `foo` once.
   In addition, if foo has section index 9, `-x foo -x 9` dumps `foo` once.
3) Give a warning instead of an error if `foo` does not exist.

The new behaviors match GNU readelf.

Also, print a new line as a separator between two section dumps.
GNU readelf uses two lines, but one seems good enough.

Reviewed By: grimar, jhenderson

Differential Revision: https://reviews.llvm.org/D63475

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363683 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU: Add ds_gws_init / ds_gws_barrier intrinsics
Matt Arsenault [Tue, 18 Jun 2019 13:19:57 +0000 (13:19 +0000)]
AMDGPU: Add ds_gws_init / ds_gws_barrier intrinsics

There may or may not be additional work to handle this correctly on
SI/CI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363678 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[MCA] Slightly refactor the bottleneck analysis view. NFCI
Andrea Di Biagio [Tue, 18 Jun 2019 12:59:46 +0000 (12:59 +0000)]
[MCA] Slightly refactor the bottleneck analysis view. NFCI

This patch slightly refactors data structures internally used by the bottleneck
analysis to track data and resource dependencies.
This patch also updates methods used to print out information about dependency
edges when in debug mode.
This is the last of a sequence of commits done in preparation for an upcoming
patch that fixes PR37494. No functional change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363677 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU: Change API for checking for exec modification
Matt Arsenault [Tue, 18 Jun 2019 12:48:36 +0000 (12:48 +0000)]
AMDGPU: Change API for checking for exec modification

Invert the name and return value to better reflect the imprecise
nature.

Force passing in the DefMI, since it's known in the 2 users and could
possibly fail for an arbitrary vreg.

Allow specifying a specific user instruction. Scan through use
instructions, instead of use operands. Add scan thresholds instead of
searching infinitely.

Stop using a set to track seen uses. I didn't understand this usage,
or why it would not check the last use. I don't think the use list has
any particular order.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363675 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoMCContext: Delete unused functions
Fangrui Song [Tue, 18 Jun 2019 12:30:06 +0000 (12:30 +0000)]
MCContext: Delete unused functions

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363674 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agogn build: Merge r363658
Nico Weber [Tue, 18 Jun 2019 12:29:04 +0000 (12:29 +0000)]
gn build: Merge r363658

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363673 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agogn build: Merge r363649
Nico Weber [Tue, 18 Jun 2019 12:26:31 +0000 (12:26 +0000)]
gn build: Merge r363649

This reverts commit "gn build: Merge r363626" because r363626
was reverted in r363649.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363672 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[SelectionDAG] Legalize vaargs that require vector splitting
Simon Pilgrim [Tue, 18 Jun 2019 12:24:02 +0000 (12:24 +0000)]
[SelectionDAG] Legalize vaargs that require vector splitting

This adds vector splitting for vaarg instructions during type legalization

Committed on behalf of @luke (Luke Lau)

Differential Revision: https://reviews.llvm.org/D60762

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363671 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU: Fold readlane from copy of SGPR or imm
Matt Arsenault [Tue, 18 Jun 2019 12:23:46 +0000 (12:23 +0000)]
AMDGPU: Fold readlane from copy of SGPR or imm

These may be inserted to assert uniformity somewhere.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363670 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU: Remove unnecessary check for virtual register
Matt Arsenault [Tue, 18 Jun 2019 12:23:45 +0000 (12:23 +0000)]
AMDGPU: Remove unnecessary check for virtual register

The copy was found by searching the uses of a virtual register, so
it's already known to be virtual.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363669 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU: Fix iterator crash in AMDGPUPromoteAlloca
Matt Arsenault [Tue, 18 Jun 2019 12:23:44 +0000 (12:23 +0000)]
AMDGPU: Fix iterator crash in AMDGPUPromoteAlloca

The lifetime intrinsic was erased, which was the next iterator.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363668 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU/GlobalISel: RegBankSelect for amdgcn.div.scale
Matt Arsenault [Tue, 18 Jun 2019 12:23:42 +0000 (12:23 +0000)]
AMDGPU/GlobalISel: RegBankSelect for amdgcn.div.scale

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363667 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[ARM] Some Thumb2ITBlock clean ups. NFC
Sjoerd Meijer [Tue, 18 Jun 2019 12:13:11 +0000 (12:13 +0000)]
[ARM] Some Thumb2ITBlock clean ups. NFC

Some more refactoring, like registering the IT Block pass, less cryptic
variable names, and some simplification of loops.

Differential Revision: https://reviews.llvm.org/D63419

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363666 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[SystemZ] Fix AHIMuxK pseudo expansion.
Jonas Paulsson [Tue, 18 Jun 2019 12:10:02 +0000 (12:10 +0000)]
[SystemZ]  Fix AHIMuxK pseudo expansion.

Do not emit a copy if the source and destination registers are the same.

Review: Ulrich Weigand

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363665 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[AMDGPU] Speed up live-in virtual register set computaion in GCNScheduleDAGMILive.
Valery Pykhtin [Tue, 18 Jun 2019 11:43:17 +0000 (11:43 +0000)]
[AMDGPU] Speed up live-in virtual register set computaion in GCNScheduleDAGMILive.

Differential revision: https://reviews.llvm.org/D62401

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363661 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[SVE][IR] Scalable Vector IR Type with pr42210 fix
Graham Hunter [Tue, 18 Jun 2019 10:11:56 +0000 (10:11 +0000)]
[SVE][IR] Scalable Vector IR Type with pr42210 fix

Recommit of D32530 with a few small changes:
  - Stopped recursively walking through aggregates in
    the verifier, so that we don't impose too much
    overhead on large modules under LTO (see PR42210).
  - Changed tests to match; the errors are slightly
    different since they only report the array or
    struct that actually contains a scalable vector,
    rather than all aggregates which contain one in
    a nested member.
  - Corrected an older comment

Reviewers: thakis, rengolin, sdesmalen

Reviewed By: sdesmalen

Differential Revision: https://reviews.llvm.org/D63321

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363658 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86] Regenerate promote.ll. NFC.
Simon Pilgrim [Tue, 18 Jun 2019 10:10:53 +0000 (10:10 +0000)]
[X86] Regenerate promote.ll. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363657 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[NFC] Improve triple match of scripts that update tests
Diogo N. Sampaio [Tue, 18 Jun 2019 10:04:36 +0000 (10:04 +0000)]
[NFC] Improve triple match of scripts that update tests

Summary:
The prior behavior of the triple matcher would stop
in the first matched triple. It was not possible to
create specific matches for sub-sets of a triple
(e.g aarch64-apple-darwin would never be used after
aarch64 was matched).

This patch:
1) Allows that specialized triples take priority,
considering that the string lenght of the triple
indentifies how specialized a triple is. If two
triples of same lenght match, the one matched first
prevails, preserving the old behavior.

2) Remove 20 duplicated triples of arm, thumb,
aarch64 options with same arguments, matching
the common prefix (aarch64, arm, thumb) of them.

3) Creates three new function matching regexes and
five triple options for arm64-apple-ios,
(arm|thumb)-apple-ios and thumb(v5)?-macho

Reviewers: lebedev.ri, RKSimon, MaskRay, gbedwell

Reviewed By: MaskRay

Subscribers: javed.absar, kristof.beyls, llvm-commits, carwil

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63145

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363656 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86] Replace any_extend* vector extensions with zero_extend* equivalents
Simon Pilgrim [Tue, 18 Jun 2019 09:50:13 +0000 (09:50 +0000)]
[X86] Replace any_extend* vector extensions with zero_extend* equivalents

First step toward addressing the vector-reduce-mul-widen.ll regression in D63281 - we should replace ANY_EXTEND/ANY_EXTEND_VECTOR_INREG in X86ISelDAGToDAG to avoid having to add duplicate patterns when treating any extensions as legal.

In future patches this will also allow us to keep any extension nodes around a lot longer in the DAG, which should mean that we can keep better track of undef elements that otherwise become zeros that we think we have to keep......

Differential Revision: https://reviews.llvm.org/D63326

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363655 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[DebugInfo][Docs] Document that prologue/epilogue variable location changes are ignored
Jeremy Morse [Tue, 18 Jun 2019 08:52:38 +0000 (08:52 +0000)]
[DebugInfo][Docs] Document that prologue/epilogue variable location changes are ignored

This patch documents that LLVM does not describe all changes in variable
locations during the prologue and the epilogue. The debugger doesn't /
shouldn't step through that portion of the function anyway, and describing
every location through such stages would bloat location lists.

Perform some minor cleanup at the same time,
 * Fix an enumerated list
 * Document that dbg.declare intrinsics have their variable location recorded
   in a MachineFunction table, not with DBG_VALUE meta-insts
 * Adds frame-indexes to the list of things that can be operands to
   DBG_VALUEs.

Differential Revision: https://reviews.llvm.org/D63083

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363654 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[SimplifyCFG] NFC, prof branch_weighs handling is simplified
Yevgeny Rouban [Tue, 18 Jun 2019 06:50:52 +0000 (06:50 +0000)]
[SimplifyCFG] NFC, prof branch_weighs handling is simplified

Using the new SwitchInstProfUpdateWrapper this patch
simplifies 3 places of prof branch_weights handling.

Differential Revision: https://reviews.llvm.org/D62123

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363652 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[llvm-objdump] Tidy up AMDGCNPrettyPrinter
Fangrui Song [Tue, 18 Jun 2019 06:35:18 +0000 (06:35 +0000)]
[llvm-objdump] Tidy up AMDGCNPrettyPrinter

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363650 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86] Add i128 ctpop and i32/i64/i128 optsize test cases to popcnt.ll
Craig Topper [Tue, 18 Jun 2019 04:52:49 +0000 (04:52 +0000)]
[X86] Add i128 ctpop and i32/i64/i128 optsize test cases to popcnt.ll

Test cases for PR41151 and D59909.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363647 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86] Move code that shrinks immediates for ((x << C1) op C2) into a helper function...
Craig Topper [Tue, 18 Jun 2019 04:23:58 +0000 (04:23 +0000)]
[X86] Move code that shrinks immediates for ((x << C1) op C2) into a helper function. NFCI

Preliminary step for D59909

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363645 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86] Remove MOVDI2SSrm/MOV64toSDrm/MOVSS2DImr/MOVSDto64mr CodeGenOnly instructions.
Craig Topper [Tue, 18 Jun 2019 03:23:15 +0000 (03:23 +0000)]
[X86] Remove MOVDI2SSrm/MOV64toSDrm/MOVSS2DImr/MOVSDto64mr CodeGenOnly instructions.

The isel patterns for these use a bitcast and load/store, but
DAG combine should have canonicalized those away.

For the purposes of the memory folding table these opcodes can be
replaced by the MOVSSrm_alt/MOVSDrm_alt and MOVSSmr/MOVSDmr opcodes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363644 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86] Introduce new MOVSSrm/MOVSDrm opcodes that use VR128 register class.
Craig Topper [Tue, 18 Jun 2019 03:23:11 +0000 (03:23 +0000)]
[X86] Introduce new MOVSSrm/MOVSDrm opcodes that use VR128 register class.

Rename the old versions that use FR32/FR64 to MOVSSrm_alt/MOVSDrm_alt.

Use the new versions in patterns that previously used a COPY_TO_REGCLASS
to VR128. These patterns expect the upper bits to be zero. The
current set up appears to work, but I'm not sure we should be
enforcing upper bits being zero through a COPY_TO_REGCLASS.

I wanted to flip the arrangement and use a COPY_TO_REGCLASS to
FR32/FR64 for the patterns that need an f32/f64 result, but that
complicated fastisel and globalisel.

I've been doing some experiments with reducing some isel patterns
and ended up in a situation where I had a
(SUBREG_TO_REG (COPY_TO_RECLASS (VMOVSSrm), VR128)) and our
post-isel peephole was unable to avoid using an instruction for
the SUBREG_TO_REG due to the COPY_TO_REGCLASS. Having a VR128
instruction removes the COPY_TO_REGCLASS that was breaking this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363643 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoGlobalISel: Remove redundant pass initialization
Tom Stellard [Tue, 18 Jun 2019 02:05:06 +0000 (02:05 +0000)]
GlobalISel: Remove redundant pass initialization

Summary:
All the GlobalISel passes are initialized when the target calls
initializeGlobalISel(), so we don't need to call the initializers
from the pass constructors.

Reviewers: qcolombet, t.p.northover, paquette, dsanders, aemerson, aditya_nandakumar

Reviewed By: aemerson

Subscribers: rovka, kristof.beyls, hiraditya, volkan, Petar.Avramovic, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63235

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363642 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[llvm-strip] Error when using stdin twice
Alex Brachet [Tue, 18 Jun 2019 00:39:10 +0000 (00:39 +0000)]
[llvm-strip] Error when using stdin twice

Summary: Implements bug [[ https://bugs.llvm.org/show_bug.cgi?id=42204 | 42204 ]]. llvm-strip now warns when the same input file is used more than once, and errors when stdin is used more than once.

Reviewers: jhenderson, rupprecht, espindola, alexshap

Reviewed By: jhenderson, rupprecht

Subscribers: emaste, arichardson, jakehehrlich, MaskRay, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63122

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363638 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoGlobalISel: Use the original flags when lowering fneg to fsub
Matt Arsenault [Mon, 17 Jun 2019 23:48:43 +0000 (23:48 +0000)]
GlobalISel: Use the original flags when lowering fneg to fsub

This was ignoring the flag on fneg, and using the source instruction's
flags. Also fixes tests missing from r358702.

Note the expansion itself isn't correct without nnan, but that should
be fixed separately.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363637 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agohwasan: Use bits [3..11) of the ring buffer entry address as the base stack tag.
Peter Collingbourne [Mon, 17 Jun 2019 23:39:51 +0000 (23:39 +0000)]
hwasan: Use bits [3..11) of the ring buffer entry address as the base stack tag.

This saves roughly 32 bytes of instructions per function with stack objects
and causes us to preserve enough information that we can recover the original
tags of all stack variables.

Now that stack tags are deterministic, we no longer need to pass
-hwasan-generate-tags-with-calls during check-hwasan. This also means that
the new stack tag generation mechanism is exercised by check-hwasan.

Differential Revision: https://reviews.llvm.org/D63360

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363636 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agohwasan: Add a tag_offset DWARF attribute to instrumented stack variables.
Peter Collingbourne [Mon, 17 Jun 2019 23:39:41 +0000 (23:39 +0000)]
hwasan: Add a tag_offset DWARF attribute to instrumented stack variables.

The goal is to improve hwasan's error reporting for stack use-after-return by
recording enough information to allow the specific variable that was accessed
to be identified based on the pointer's tag. Currently we record the PC and
lower bits of SP for each stack frame we create (which will eventually be
enough to derive the base tag used by the stack frame) but that's not enough
to determine the specific tag for each variable, which is the stack frame's
base tag XOR a value (the "tag offset") that is unique for each variable in
a function.

In IR, the tag offset is most naturally represented as part of a location
expression on the llvm.dbg.declare instruction. However, the presence of the
tag offset in the variable's actual location expression is likely to confuse
debuggers which won't know about tag offsets, and moreover the tag offset
is not required for a debugger to determine the location of the variable on
the stack, so at the DWARF level it is represented as an attribute so that
it will be ignored by debuggers that don't know about it.

Differential Revision: https://reviews.llvm.org/D63119

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363635 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agogn build: Merge r363626.
Peter Collingbourne [Mon, 17 Jun 2019 23:39:31 +0000 (23:39 +0000)]
gn build: Merge r363626.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363634 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[GlobalISel][Localizer] Rewrite localizer to run in 2 phases, inter & intra block.
Amara Emerson [Mon, 17 Jun 2019 23:20:29 +0000 (23:20 +0000)]
[GlobalISel][Localizer] Rewrite localizer to run in 2 phases, inter & intra block.

Inter-block localization is the same as what currently happens, except now it
only runs on the entry block because that's where the problematic constants with
long live ranges come from.

The second phase is a new intra-block localization phase which attempts to
re-sink the already localized instructions further right before one of the
multiple uses.

One additional change is to also localize G_GLOBAL_VALUE as they're constants
too. However, on some targets like arm64 it takes multiple instructions to
materialize the value, so some additional heuristics with a TTI hook have been
introduced attempt to prevent code size regressions when localizing these.

Overall, these changes improve CTMark code size on arm64 by 1.2%.

Full code size results:

Program                                         baseline       new       diff
------------------------------------------------------------------------------
 test-suite...-typeset/consumer-typeset.test    1249984      1217216     -2.6%
 test-suite...:: CTMark/ClamAV/clamscan.test    1264928      1232152     -2.6%
 test-suite :: CTMark/SPASS/SPASS.test          1394092      1361316     -2.4%
 test-suite...Mark/mafft/pairlocalalign.test    731320       714928      -2.2%
 test-suite :: CTMark/lencod/lencod.test        1340592      1324200     -1.2%
 test-suite :: CTMark/kimwitu++/kc.test         3853512      3820420     -0.9%
 test-suite :: CTMark/Bullet/bullet.test        3406036      3389652     -0.5%
 test-suite...ark/tramp3d-v4/tramp3d-v4.test    8017000      8016992     -0.0%
 test-suite...TMark/7zip/7zip-benchmark.test    2856588      2856588      0.0%
 test-suite...:: CTMark/sqlite3/sqlite3.test    765704       765704       0.0%
 Geomean difference                                                      -1.2%

Differential Revision: https://reviews.llvm.org/D63303

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363632 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoPropagate fmf in IRTranslate for fneg
Michael Berg [Mon, 17 Jun 2019 23:19:40 +0000 (23:19 +0000)]
Propagate fmf in IRTranslate for fneg

Summary: This case is related to D63405 in that we need to be propagating FMF on negates.

Reviewers: volkan, spatel, arsenm

Reviewed By: arsenm

Subscribers: wdng, javed.absar

Differential Revision: https://reviews.llvm.org/D63458

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363631 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoUse VR128X instead of FR32X/FR64X for the register class in VMOVSSZmrk/VMOVSDZmrk.
Craig Topper [Mon, 17 Jun 2019 23:08:29 +0000 (23:08 +0000)]
Use VR128X instead of FR32X/FR64X for the register class in VMOVSSZmrk/VMOVSDZmrk.

Removes COPY_TO_REGCLASS from some patterns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363630 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86] Make an assert in LowerSCALAR_TO_VECTOR stricter to make it clear what types...
Craig Topper [Mon, 17 Jun 2019 23:08:09 +0000 (23:08 +0000)]
[X86] Make an assert in LowerSCALAR_TO_VECTOR stricter to make it clear what types are allowed here. NFC

Make it clear that only integer type with i32 or smaller elements shoudl get to this part of the code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363629 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[AMDGPU] Use custom inserter for gfx10 VOP2b
Stanislav Mekhanoshin [Mon, 17 Jun 2019 22:37:37 +0000 (22:37 +0000)]
[AMDGPU] Use custom inserter for gfx10 VOP2b

This is part of the approved D63204 pending parent revision.
This small change is in fact a part of the VOP2b legalization which
does not technically belong to wave32 support, so extracted
separately.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363625 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[AMDGPU] gfx1010 subvector test. NFC.
Stanislav Mekhanoshin [Mon, 17 Jun 2019 21:55:06 +0000 (21:55 +0000)]
[AMDGPU] gfx1010 subvector test. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363623 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[test][AArch64] Relax the check line for G_BRJT in legalizer-info-validation.mir
Volkan Keles [Mon, 17 Jun 2019 21:25:25 +0000 (21:25 +0000)]
[test][AArch64] Relax the check line for G_BRJT in legalizer-info-validation.mir

Replace the specific number with a pattern to relax the test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363621 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoTeach getSCEVAtScope how to handle loop phis w/invariant operands in loops w/taken...
Philip Reames [Mon, 17 Jun 2019 21:06:17 +0000 (21:06 +0000)]
Teach getSCEVAtScope how to handle loop phis w/invariant operands in loops w/taken backedges

This patch really contains two pieces:
    Teach SCEV how to fold a phi in the header of a loop to the value on the backedge when a) the backedge is known to execute at least once, and b) the value is safe to use globally within the scope dominated by the original phi.
    Teach IndVarSimplify's rewriteLoopExitValues to allow loop invariant expressions which already exist (and thus don't need new computation inserted) even in loops where we can't optimize away other uses.

Differential Revision: https://reviews.llvm.org/D63224

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363619 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAdd convenience utility for replacing a range within a container with a
Richard Smith [Mon, 17 Jun 2019 21:01:09 +0000 (21:01 +0000)]
Add convenience utility for replacing a range within a container with a
different range, in preparation for use in Clang.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363617 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[globalisel] Fix iterator invalidation in the extload combines
Daniel Sanders [Mon, 17 Jun 2019 20:56:31 +0000 (20:56 +0000)]
[globalisel] Fix iterator invalidation in the extload combines

Summary:
Change the way we deal with iterator invalidation in the extload combines as it
was still possible to neglect to visit a use. Even worse, it happened in the
in-tree test cases and the checks weren't good enough to detect it.

We now take a cheap copy of the use list before iterating over it. This
prevents iterator invalidation from occurring and has the nice side effect
of making the existing schedule-for-erase/schedule-for-insert mechanism
moot.

Reviewers: aditya_nandakumar

Reviewed By: aditya_nandakumar

Subscribers: rovka, kristof.beyls, javed.absar, volkan, Petar.Avramovic, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D61813

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363616 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[AMDGPU] Propagate function attributes thru bitcasts
Stanislav Mekhanoshin [Mon, 17 Jun 2019 20:42:48 +0000 (20:42 +0000)]
[AMDGPU] Propagate function attributes thru bitcasts

AMDGPUPropagateAttributes will not work on function bitcatsts,
so move AMDGPUFixFunctionBitcasts before it.

Differential Revision: https://reviews.llvm.org/D63455

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363614 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoFix a bug w/inbounds invalidation in LFTR (recommit)
Philip Reames [Mon, 17 Jun 2019 20:32:22 +0000 (20:32 +0000)]
Fix a bug w/inbounds invalidation in LFTR (recommit)

Recommit r363289 with a bug fix for crash identified in pr42279.  Issue was that a loop exit test does not have to be an icmp, leading to a null dereference crash when new logic was exercised for that case.  Test case previously committed in r363601.

Original commit comment follows:

This contains fixes for two cases where we might invalidate inbounds and leave it stale in the IR (a miscompile). Case 1 is when switching to an IV with no dynamically live uses, and case 2 is when doing pre-to-post conversion on the same pointer type IV.

The basic scheme used is to prove that using the given IV (pre or post increment forms) would have to already trigger UB on the path to the test we're modifying. As such, our potential UB triggering use does not change the semantics of the original program.

As was pointed out in the review thread by Nikita, this is defending against a separate issue from the hasConcreteDef case. This is about poison, that's about undef. Unfortunately, the two are different, see Nikita's comment for a fuller explanation, he explains it well.

(Note: I'm going to address Nikita's last style comment in a separate commit just to minimize chance of subtle bugs being introduced due to typos.)

Differential Revision: https://reviews.llvm.org/D62939

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363613 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agogn build: Merge r363483.
Peter Collingbourne [Mon, 17 Jun 2019 20:03:11 +0000 (20:03 +0000)]
gn build: Merge r363483.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363610 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agogn build: Merge r363584.
Peter Collingbourne [Mon, 17 Jun 2019 19:59:16 +0000 (19:59 +0000)]
gn build: Merge r363584.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363609 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU/GFX10: Don't generate s_code_end padding in the asm-printer
Nicolai Haehnle [Mon, 17 Jun 2019 19:28:43 +0000 (19:28 +0000)]
AMDGPU/GFX10: Don't generate s_code_end padding in the asm-printer

Summary:
The purpose of the padding is to guard against stale code being
fetched into the instruction cache by the lowest level prefetching.
We're generating relocatable ELF here, and so the padding should
arguably be added by the linker. This is in fact what Mesa does.

This also fixes multi-part shaders for Mesa.

Change-Id: I6bfede58f20e9f337762ccf39ef9e0e263e69e82

Reviewers: arsenm, rampitec, t-tye

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

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63427

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363602 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoReduced test case for pr42279 in advance of the relevant re-commit + fix
Philip Reames [Mon, 17 Jun 2019 19:27:45 +0000 (19:27 +0000)]
Reduced test case for pr42279 in advance of the relevant re-commit + fix

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363601 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU: Explicitly define a triple for some tests
Nicolai Haehnle [Mon, 17 Jun 2019 19:25:57 +0000 (19:25 +0000)]
AMDGPU: Explicitly define a triple for some tests

Summary:
This is related to the changes to the groupstaticsize intrinsic in
D61494 which would otherwise make the related tests in these files
fail or much less useful.

Note that for some reason, SOPK generation is less effective in the
amdhsa OS, which is why I chose PAL. I haven't investigated this
deeper.

Change-Id: I6bb99569338f7a433c28b4c9eb1e3e036b00d166

Reviewers: arsenm

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

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63392

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363600 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[EarlyCSE] Fix hashing of self-compares
Joseph Tremoulet [Mon, 17 Jun 2019 19:11:28 +0000 (19:11 +0000)]
[EarlyCSE] Fix hashing of self-compares

Summary:
Update compare normalization in SimpleValue hashing to break ties (when
the same value is being compared to itself) by switching to the swapped
predicate if it has a lower numerical value.  This brings the hashing in
line with isEqual, which already recognizes the self-compares with
swapped predicates as equal.

Fixes PR 42280.

Reviewers: spatel, efriedma, nikic, fhahn, uabelho

Reviewed By: nikic

Subscribers: hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63349

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363598 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[MemorySSA] Don't use template when the clone is a simplified instruction.
Alina Sbirlea [Mon, 17 Jun 2019 18:58:40 +0000 (18:58 +0000)]
[MemorySSA] Don't use template when the clone is a simplified instruction.

Summary:
LoopRotate doesn't create a faithful clone of an instruction, it may
simplify it beforehand. Hence the clone of an instruction that has a
MemoryDef associated may not be a definition, but a use or not a memory
alternig instruction.
Don't rely on the template when the clone may be simplified.

Reviewers: george.burgess.iv

Subscribers: jlebar, Prazek, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63355

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363597 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[GlobalISel][AArch64] Fold G_SUB into G_ICMP when it's safe to do so
Jessica Paquette [Mon, 17 Jun 2019 18:40:06 +0000 (18:40 +0000)]
[GlobalISel][AArch64] Fold G_SUB into G_ICMP when it's safe to do so

Basically porting over the behaviour in AArch64ISelLowering to GISel. See
emitComparison for reference.

When we have something like this:

```
  lhs = G_SUB 0, y
  ...
  G_ICMP lhs, rhs
```

We can fold away the G_SUB and produce a cmn instead, given that we produce
the same value in NZCV.

Add a test showing that the transformation works, and also showing that we
don't perform the transformation when it's unsafe.

Also factor out the CSet emission into emitCSetForICMP.

Differential Revision: https://reviews.llvm.org/D63163

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363596 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86] Add TB_NO_REVERSE to some memory folding table entries where the register form...
Craig Topper [Mon, 17 Jun 2019 18:38:07 +0000 (18:38 +0000)]
[X86] Add TB_NO_REVERSE to some memory folding table entries where the register form requires 64-bit mode, but the memory form does not.

We don't know if its safe to unfold if we're in 32-bit mode.

This is simlar to what was done to some load opcodes in r363523.

I think its pretty unlikely we will try to unfold these anyway so
I don't think this is testable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363595 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoLiveInterval.h: add LiveRange::findIndexesLiveAt function - return a list of SlotInde...
Valery Pykhtin [Mon, 17 Jun 2019 18:23:39 +0000 (18:23 +0000)]
LiveInterval.h: add LiveRange::findIndexesLiveAt function - return a list of SlotIndexes the LiveRange live at.

Differential revision: https://reviews.llvm.org/D62411

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363593 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86][SSE] Scalarize under-aligned XMM vector nt-stores (PR42026)
Simon Pilgrim [Mon, 17 Jun 2019 18:20:04 +0000 (18:20 +0000)]
[X86][SSE] Scalarize under-aligned XMM vector nt-stores (PR42026)

If a XMM non-temporal store has less than natural alignment, scalarize the vector - with SSE4A we can stay on the vector and use MOVNTSD(f64), else we must move to GPRs and use MOVNTI(i32/i64).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363592 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU: Make getreg intrinsic inaccessiblememonly
Matt Arsenault [Mon, 17 Jun 2019 18:17:25 +0000 (18:17 +0000)]
AMDGPU: Make getreg intrinsic inaccessiblememonly

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363591 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[MemorySSA] Add all MemoryPhis before filling their values.
Alina Sbirlea [Mon, 17 Jun 2019 18:16:53 +0000 (18:16 +0000)]
[MemorySSA] Add all MemoryPhis before filling their values.

Summary:
Add all MemoryPhis in IDF before filling in their incomign values.
Otherwise, a new Phi can be added that needs to become the incoming
value of another Phi.
Test fails the verification in verifyPrevDefInPhis.

Reviewers: george.burgess.iv

Subscribers: jlebar, Prazek, zzheng, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63353

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363590 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[AMDGPU] gfx1010 wavefrontsize intrinsic folding
Stanislav Mekhanoshin [Mon, 17 Jun 2019 17:57:50 +0000 (17:57 +0000)]
[AMDGPU] gfx1010 wavefrontsize intrinsic folding

Differential Revision: https://reviews.llvm.org/D63206

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363588 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU: Fold readlane/readfirstlane calls
Matt Arsenault [Mon, 17 Jun 2019 17:52:35 +0000 (17:52 +0000)]
AMDGPU: Fold readlane/readfirstlane calls

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363587 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[AMDGPU] Pass to propagate ABI attributes from kernels to the functions
Stanislav Mekhanoshin [Mon, 17 Jun 2019 17:47:28 +0000 (17:47 +0000)]
[AMDGPU] Pass to propagate ABI attributes from kernels to the functions

The pass works in two modes:

Mode 1: Just set attributes starting from kernels. This can work at
the very beginning of opt and llc pipeline, but cannot clone functions
because it must be a function pass.

Mode 2: Actually clone functions for new attributes. This can only work
after all function passes in the opt pipeline because it has to be a
module pass.

Differential Revision: https://reviews.llvm.org/D63208

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363586 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agogn build: Merge r363541
Nico Weber [Mon, 17 Jun 2019 17:45:12 +0000 (17:45 +0000)]
gn build: Merge r363541

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363583 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86][AVX] Split under-aligned vector nt-stores.
Simon Pilgrim [Mon, 17 Jun 2019 17:22:38 +0000 (17:22 +0000)]
[X86][AVX] Split under-aligned vector nt-stores.

If a YMM/ZMM non-temporal store has less than natural alignment, split the vector - either they will be satisfactorily aligned or will continue to be split until they are XMMs - at which point the legalizer will scalarize it.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363582 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[LV] Suppress vectorization in some nontemporal cases
Warren Ristow [Mon, 17 Jun 2019 17:20:08 +0000 (17:20 +0000)]
[LV] Suppress vectorization in some nontemporal cases

When considering a loop containing nontemporal stores or loads for
vectorization, suppress the vectorization if the corresponding
vectorized store or load with the aligment of the original scaler
memory op is not supported with the nontemporal hint on the target.

This adds two new functions:
  bool isLegalNTStore(Type *DataType, unsigned Alignment) const;
  bool isLegalNTLoad(Type *DataType, unsigned Alignment) const;

to TTI, leaving the target independent default implementation as
returning true, but with overriding implementations for X86 that
check the legality based on available Subtarget features.

This fixes https://llvm.org/PR40759

Differential Revision: https://reviews.llvm.org/D61764

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363581 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoGlobalISel: Ignore callsite attributes when picking intrinsic type
Matt Arsenault [Mon, 17 Jun 2019 17:01:35 +0000 (17:01 +0000)]
GlobalISel: Ignore callsite attributes when picking intrinsic type

A target intrinsic may be defined as possibly reading memory, but the
call site may have additional knowledge that it doesn't read
memory. The intrinsic lowering will expect the pessimistic assumption
of the intrinsic definition, so the chain should still be used.

I fixed the same bug in SelectionDAG in r287593.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363580 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoGlobalISel: Verify intrinsics
Matt Arsenault [Mon, 17 Jun 2019 17:01:32 +0000 (17:01 +0000)]
GlobalISel: Verify intrinsics

I keep using the wrong instruction when manually writing tests. This
really needs to check the number of operands, but I don't see an easy
way to do that right now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363579 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU/GlobalISel: Account for multiple defs when finding intrinsic ID
Matt Arsenault [Mon, 17 Jun 2019 17:01:27 +0000 (17:01 +0000)]
AMDGPU/GlobalISel: Account for multiple defs when finding intrinsic ID

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363578 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[AMDGPU] gfx1010 wave32 metadata
Stanislav Mekhanoshin [Mon, 17 Jun 2019 16:48:56 +0000 (16:48 +0000)]
[AMDGPU] gfx1010 wave32 metadata

Differential Revision: https://reviews.llvm.org/D63207

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363577 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU/GlobalISel: Implement select for G_ICMP and G_SELECT
Tom Stellard [Mon, 17 Jun 2019 16:27:43 +0000 (16:27 +0000)]
AMDGPU/GlobalISel: Implement select for G_ICMP and G_SELECT

Reviewers: arsenm

Subscribers: kzhuravl, jvesely, wdng, nhaehnle, yaxunl, rovka, kristof.beyls, dstuttard, tpr, t-tye, hiraditya, Petar.Avramovic, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D60640

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363576 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[Remarks] Extend -fsave-optimization-record to specify the format
Francis Visoiu Mistrih [Mon, 17 Jun 2019 16:06:00 +0000 (16:06 +0000)]
[Remarks] Extend -fsave-optimization-record to specify the format

Use -fsave-optimization-record=<format> to specify a different format
than the default, which is YAML.

For now, only YAML is supported.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363573 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86] combineLoad - begun making the load split code more generic. NFCI.
Simon Pilgrim [Mon, 17 Jun 2019 15:54:36 +0000 (15:54 +0000)]
[X86] combineLoad - begun making the load split code more generic. NFCI.

This is currently only used for ymm->xmm splitting but we shouldn't hardcode the offsets/alignment.

This is necessary for an upcoming patch to split under-aligned non-temporal vector loads.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363570 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoPHINode: introduce setIncomingValueForBlock() function, and use it.
Whitney Tsang [Mon, 17 Jun 2019 14:38:56 +0000 (14:38 +0000)]
PHINode: introduce setIncomingValueForBlock() function, and use it.

Summary:
There is PHINode::getBasicBlockIndex() and PHINode::setIncomingValue()
but no function to replace incoming value for a specified BasicBlock*
predecessor.
Clearly, there are a lot of places that could use that functionality.

Reviewer: craig.topper, lebedev.ri, Meinersbur, kbarton, fhahn
Reviewed By: Meinersbur, fhahn
Subscribers: fhahn, hiraditya, zzheng, jsji, llvm-commits
Tag: LLVM
Differential Revision: https://reviews.llvm.org/D63338

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363566 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86][SSE] Add tests for underaligned nt loads
Simon Pilgrim [Mon, 17 Jun 2019 14:38:17 +0000 (14:38 +0000)]
[X86][SSE] Add tests for underaligned nt loads

Test both 'unaligned' (which we should just use regular unaligned loads) and 'subvector aligned' (which we should split)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363565 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86][SSE] Prevent misaligned non-temporal vector load/store combines
Simon Pilgrim [Mon, 17 Jun 2019 14:26:10 +0000 (14:26 +0000)]
[X86][SSE] Prevent misaligned non-temporal vector load/store combines

For loads, pre-SSE41 we can't perform NT loads at all, and after that we can only perform vector aligned loads, so if the alignment is less than for a xmm we'll just end up using the regular unaligned vector loads anyway.

First step towards fixing PR42026 - the next step for stores will be to use SSE4A movntsd where possible and to avoid the stack spill on SSE2 targets.

Differential Revision: https://reviews.llvm.org/D63246

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363564 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoInferAddressSpaces: Fix cloning original addrspacecast
Matt Arsenault [Mon, 17 Jun 2019 14:13:29 +0000 (14:13 +0000)]
InferAddressSpaces: Fix cloning original addrspacecast

If an addrspacecast needed to be inserted again, this was creating a
clone of the original cast for each user. Just use the original, which
also saves losing the value name.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363562 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU: Ignore subtarget for InferAddressSpaces
Matt Arsenault [Mon, 17 Jun 2019 14:13:24 +0000 (14:13 +0000)]
AMDGPU: Ignore subtarget for InferAddressSpaces

Even if the target doesn't have flat instructions, addrspace(0) is
still flat. It just happens to not work.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363561 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU: Mark exp/exp.compr as inaccessiblememonly
Matt Arsenault [Mon, 17 Jun 2019 13:52:24 +0000 (13:52 +0000)]
AMDGPU: Mark exp/exp.compr as inaccessiblememonly

Should also be marked writeonly, but I think that would require
splitting the version with done set to a separate intrinsic

Test change is only from renumbering the attribute group numbers,
which for some reason the generated check lines consider.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363560 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU/GlobalISel: Fix default mapping for non-register operands
Matt Arsenault [Mon, 17 Jun 2019 13:52:19 +0000 (13:52 +0000)]
AMDGPU/GlobalISel: Fix default mapping for non-register operands

Tests will be in future commits when new intrinsics are handled here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363559 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAMDGPU: Cleanup custom PseudoSourceValue definitions
Matt Arsenault [Mon, 17 Jun 2019 13:52:15 +0000 (13:52 +0000)]
AMDGPU: Cleanup custom PseudoSourceValue definitions

Use separate enums for each kind, avoid repeating overloads, and add
missing classof implementation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363558 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[CodeGen] Check for HardwareLoop Latch ExitBlock
Sam Parker [Mon, 17 Jun 2019 13:39:28 +0000 (13:39 +0000)]
[CodeGen] Check for HardwareLoop Latch ExitBlock

The HardwareLoops pass finds exit blocks with a scevable exit count.
If the target specifies to update the loop counter in a register,
through a phi, we need to ensure that the exit block is a latch so
that we can insert the phi with the correct value for the incoming
edge.

Differential Revision: https://reviews.llvm.org/D63336

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363556 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[X86][SSE] Avoid unnecessary stack codegen in NT store codegen tests.
Simon Pilgrim [Mon, 17 Jun 2019 12:35:26 +0000 (12:35 +0000)]
[X86][SSE] Avoid unnecessary stack codegen in NT store codegen tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363552 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoAsmPrinter: add doc-string for EmitLinkage
Nicolai Haehnle [Mon, 17 Jun 2019 12:24:04 +0000 (12:24 +0000)]
AsmPrinter: add doc-string for EmitLinkage

Change-Id: I376fcbd58f84a2aac6aaf744bc1665c92d312b25

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363550 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agogn build: Merge r363530
Nico Weber [Mon, 17 Jun 2019 12:18:27 +0000 (12:18 +0000)]
gn build: Merge r363530

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363549 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[LV] Deny irregular types in interleavedAccessCanBeWidened
Bjorn Pettersson [Mon, 17 Jun 2019 12:02:24 +0000 (12:02 +0000)]
[LV] Deny irregular types in interleavedAccessCanBeWidened

Summary:
Avoid that loop vectorizer creates loads/stores of vectors
with "irregular" types when interleaving. An example of
an irregular type is x86_fp80 that is 80 bits, but that
may have an allocation size that is 96 bits. So an array
of x86_fp80 is not bitcast compatible with a vector
of the same type.

Not sure if interleavedAccessCanBeWidened is the best
place for this check, but it solves the problem seen
in the added test case. And it is the same kind of check
that already exists in memoryInstructionCanBeWidened.

Reviewers: fhahn, Ayal, craig.topper

Reviewed By: fhahn

Subscribers: hiraditya, rkruppe, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63386

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363547 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoTest forward references in IntrinsicEmitter on Neon LD(2|3|4)
Sander de Smalen [Mon, 17 Jun 2019 12:01:53 +0000 (12:01 +0000)]
Test forward references in IntrinsicEmitter on Neon LD(2|3|4)

This patch tests the forward-referencing added in D62995 by changing
some existing intrinsics to use forward referencing of overloadable
parameters, rather than backward referencing.

This patch changes the TableGen definition/implementation of
llvm.aarch64.neon.ld2lane and llvm.aarch64.neon.ld2lane intrinsics
(and similar for ld3 and ld4). This change is intended to be
non-functional, since the behaviour of the intrinsics is
expected to be the same.

Reviewers: arsenm, dmgreen, RKSimon, greened, rnk

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D63189

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363546 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[DAGCombiner] [CodeGenPrepare] More comprehensive GEP splitting
Luis Marques [Mon, 17 Jun 2019 10:54:12 +0000 (10:54 +0000)]
[DAGCombiner] [CodeGenPrepare] More comprehensive GEP splitting

Some GEPs were not being split, presumably because that split would just be
undone by the DAGCombiner. Not performing those splits can prevent important
optimizations, such as preventing the element indices / member offsets from
being (partially) folded into load/store instruction immediates. This patch:

- Makes the splits also occur in the cases where the base address and the GEP
  are in the same BB.
- Ensures that the DAGCombiner doesn't reassociate them back again.

Differential Revision: https://reviews.llvm.org/D60294

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363544 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoFix clang -Wcovered-switch-default after stack-id change by D60137
Fangrui Song [Mon, 17 Jun 2019 10:20:20 +0000 (10:20 +0000)]
Fix clang -Wcovered-switch-default after stack-id change by D60137

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363543 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[SelectionDAG] Fold insert_subvector(undef, extract_subvector(v, c), c) -> v in getNode
Simon Pilgrim [Mon, 17 Jun 2019 10:14:52 +0000 (10:14 +0000)]
[SelectionDAG] Fold insert_subvector(undef, extract_subvector(v, c), c) -> v in getNode

This is already done in DAGCombiner::visitINSERT_SUBVECTOR, but this helps a number of shuffles across different vector widths recognise when they come from the same source.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363542 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[SCEV] Use NoWrapFlags when expanding a simple mul
Sam Parker [Mon, 17 Jun 2019 10:05:18 +0000 (10:05 +0000)]
[SCEV] Use NoWrapFlags when expanding a simple mul

Second functional change following on from rL362687. Pass the
NoWrapFlags from the MulExpr to InsertBinop when we're generating a
shl or mul.

Differential Revision: https://reviews.llvm.org/D61934

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363540 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[llvm-objdump] Use %08 instead of %016 to print leading addresses for 32-bit binaries
Fangrui Song [Mon, 17 Jun 2019 09:59:55 +0000 (09:59 +0000)]
[llvm-objdump] Use %08 instead of %016 to print leading addresses for 32-bit binaries

Reviewed By: grimar

Differential Revision: https://reviews.llvm.org/D63398

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363539 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[lit] Delete empty lines at the end of lit.local.cfg NFC
Fangrui Song [Mon, 17 Jun 2019 09:51:07 +0000 (09:51 +0000)]
[lit] Delete empty lines at the end of lit.local.cfg NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363538 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[NFC][Codegen] Standalone tests for icmp eq/ne (urem %x, C), 0 -> icmp eq/ne %x,...
Roman Lebedev [Mon, 17 Jun 2019 09:50:50 +0000 (09:50 +0000)]
[NFC][Codegen] Standalone tests for icmp eq/ne (urem %x, C), 0 -> icmp eq/ne %x, 0 fold (D63390)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363537 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[ARM] Fix another -Wunused-variable in -DLLVM_ENABLE_ASSERTIONS=off builds after...
Fangrui Song [Mon, 17 Jun 2019 09:29:50 +0000 (09:29 +0000)]
[ARM] Fix another -Wunused-variable in -DLLVM_ENABLE_ASSERTIONS=off builds after D63265

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363535 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[ARM] Fix -Wunused-variable in -DLLVM_ENABLE_ASSERTIONS=off builds after D63265
Fangrui Song [Mon, 17 Jun 2019 09:26:50 +0000 (09:26 +0000)]
[ARM] Fix -Wunused-variable in -DLLVM_ENABLE_ASSERTIONS=off builds after D63265

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363534 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoDescribe stack-id as an enum
Sander de Smalen [Mon, 17 Jun 2019 09:13:29 +0000 (09:13 +0000)]
Describe stack-id as an enum

This patch changes MIR stack-id from an integer to an enum,
and adds printing/parsing support for this in MIR files. The default
stack-id '0' is now renamed to 'default'.

This should make MIR tests that have stack objects with different stack-ids
more descriptive. It also clarifies code operating on StackID.

Reviewers: arsenm, thegameg, qcolombet

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D60137

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363533 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[ARM] Remove ARMComputeBlockSize
Sam Parker [Mon, 17 Jun 2019 09:13:10 +0000 (09:13 +0000)]
[ARM] Remove ARMComputeBlockSize

Forgot to remove file!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363532 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[ARM] Add ARMBasicBlockInfo.cpp
Sam Parker [Mon, 17 Jun 2019 09:05:43 +0000 (09:05 +0000)]
[ARM] Add ARMBasicBlockInfo.cpp

Forgot to add file!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363531 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[ARM] Extract some code from ARMConstantIslandPass
Sam Parker [Mon, 17 Jun 2019 08:49:09 +0000 (08:49 +0000)]
[ARM] Extract some code from ARMConstantIslandPass

Create the ARMBasicBlockUtils class for tracking and querying basic
blocks sizes so we can use them when generating low-overhead loops.

Differential Revision: https://reviews.llvm.org/D63265

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363530 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoRe-commit r357452 (take 3): "SimplifyCFG SinkCommonCodeFromPredecessors: Also sink...
Hans Wennborg [Mon, 17 Jun 2019 07:47:28 +0000 (07:47 +0000)]
Re-commit r357452 (take 3): "SimplifyCFG SinkCommonCodeFromPredecessors: Also sink function calls without used results (PR41259)"

Third time's the charm.

This was reverted in r363220 due to being suspected of an internal benchmark
regression and a test failure, none of which turned out to be caused by this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363529 91177308-0d34-0410-b5e6-96231b3b80d8

4 years ago[SimplifyCFG] Fix prof branch_weights MD while removing unreachable switch cases
Yevgeny Rouban [Mon, 17 Jun 2019 05:55:12 +0000 (05:55 +0000)]
[SimplifyCFG] Fix prof branch_weights MD while removing unreachable switch cases

SimplifyCFG has a bug that results in inconsistent prof branch_weights metadata
if unreachable switch cases are removed. This patch fixes this bug by making use
of the newly introduced SwitchInstProfUpdateWrapper class (see patch D62122).
A new test is created.

Differential Revision: https://reviews.llvm.org/D62186

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363527 91177308-0d34-0410-b5e6-96231b3b80d8

4 years agoPowerPC: Optimize SPE double parameter calling setup
Justin Hibbits [Mon, 17 Jun 2019 03:15:23 +0000 (03:15 +0000)]
PowerPC: Optimize SPE double parameter calling setup

Summary:
SPE passes doubles the same as soft-float, in register pairs as i32
types.  This is all handled by the target-independent layer.  However,
this is not optimal when splitting or reforming the doubles, as it
pushes to the stack and loads from, on either side.

For instance, to pass a double argument to a function, assuming the
double value is in r5, the sequence currently looks like this:

    evstdd      5, X(1)
    lwz         3, X(1)
    lwz         4, X+4(1)

Likewise, to form a double into r5 from args in r3 and r4:

    stw         3, X(1)
    stw         4, X+4(1)
    evldd       5, X(1)

This optimizes the fence to use SPE instructions.  Now, to pass a double
to a function:

    mr          4, 5
    evmergehi   3, 5, 5

And to form a double into r5 from args in r3 and r4:

    evmergelo   5, 3, 4

This is comparable to the way that gcc generates the double splits.

This also fixes a bug with expanding builtins to libcalls, where the
LowerCallTo() code path was generating intermediate illegal type nodes.

Reviewers: nemanjai, hfinkel, joerg

Subscribers: kbarton, jfb, jsji, llvm-commits

Differential Revision: https://reviews.llvm.org/D54583

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@363526 91177308-0d34-0410-b5e6-96231b3b80d8