platform/upstream/llvm.git
4 years ago[X86] Use carry flag from add for (seteq (add X, -1), -1).
Craig Topper [Tue, 31 Dec 2019 23:05:21 +0000 (15:05 -0800)]
[X86] Use carry flag from add for (seteq (add X, -1), -1).

If we just subtracted 1 and are checking if the result is -1. We can use the carry flag from the ADD instead of an explicit CMP. I'm using the same checks for the add users as EmitTest.

Fixes one case from PR44412

Differential Revision: https://reviews.llvm.org/D72019

4 years ago[LegalizeVectorOps][AArch64] Stop asking for v4f16 fp_round and fp_extend to be promoted.
Craig Topper [Tue, 31 Dec 2019 22:04:09 +0000 (14:04 -0800)]
[LegalizeVectorOps][AArch64] Stop asking for v4f16 fp_round and fp_extend to be promoted.

These operations are needed as building blocks for promoting so they
can't be promoted themselves.

This appeared to work because the fp_extend query type for operation
actions is the result type, not the input type so it never triggered
in the legalizer.

For fp_round, the vector op legalizer just ended up creating a
nop fp_extend that was elided by getNode, followed by a nop
fp_round that was also elided by getNode. This was followed by
a final fp_round from v4f32 back to vf416 which was CSEd to the
original node. Then legalize vector ops just believed that node
legalized to itself. LegalizeDAG took another crack at promoting
it, but didn't have a handler so just skipped it with a debug
message saying it wasn't promoted.

This patch just removes the operation actions to avoid this
non-sense. Found while trying to refactor LegalizeVectorOps to
handle multiple result nodes better.

4 years agoAMDGPU: Precommit test showing extra instructions are introduced
Matt Arsenault [Tue, 31 Dec 2019 18:51:52 +0000 (13:51 -0500)]
AMDGPU: Precommit test showing extra instructions are introduced

4 years agoRevert "[Diagnostic] Add ftabstop to -Wmisleading-indentation"
Martin Storsjö [Tue, 31 Dec 2019 19:38:43 +0000 (21:38 +0200)]
Revert "[Diagnostic] Add ftabstop to -Wmisleading-indentation"

This reverts commit b47b35ff51b355a446483777155290541ab64fae.

This caused failed asserts (Assertion `FIDAndOffset.second >
ColNo && "Column number smaller than file offset?"' failed.)
on a source file with a single line containing
"int main (void) { for( int i = 0; i < 9; i++ ); return 0; }".

4 years ago[amdgpu] Fix scoreboard updating on `s_waitcnt_vscnt`.
Michael Liao [Wed, 25 Dec 2019 05:47:18 +0000 (00:47 -0500)]
[amdgpu] Fix scoreboard updating on `s_waitcnt_vscnt`.

Summary: - Other counters are accidentally cleared.

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

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D71866

4 years ago[OPENMP]Emit artificial threprivate vars as threadlocal, if possible.
Alexey Bataev [Tue, 31 Dec 2019 17:41:57 +0000 (12:41 -0500)]
[OPENMP]Emit artificial threprivate vars as threadlocal, if possible.

It may improve performance for declare reduction constructs.

4 years ago[mlir][docs] Remove redundant path prefix
Jacques Pienaar [Tue, 31 Dec 2019 18:56:24 +0000 (10:56 -0800)]
[mlir][docs] Remove redundant path prefix

./ is not needed.

4 years ago[X86] Add test case for opposite branch condition for PR44412. NFC
Craig Topper [Tue, 31 Dec 2019 18:57:58 +0000 (10:57 -0800)]
[X86] Add test case for opposite branch condition for PR44412. NFC

4 years ago[CodeGen] Emit conj/conjf/confjl libcalls as fneg instructions if possible.
Craig Topper [Tue, 31 Dec 2019 18:33:30 +0000 (10:33 -0800)]
[CodeGen] Emit conj/conjf/confjl libcalls as fneg instructions if possible.

We already recognize the __builtin versions of these, might as well
recognize the libcall version.

Differential Revision: https://reviews.llvm.org/D72028

4 years ago[mlir] Make code blocks more consistent
Jacques Pienaar [Tue, 31 Dec 2019 17:52:19 +0000 (09:52 -0800)]
[mlir] Make code blocks more consistent

Use the same form specification for the same type of code.

4 years ago[InstCombine] fold zext of masked bit set/clear
Sanjay Patel [Tue, 31 Dec 2019 17:29:53 +0000 (12:29 -0500)]
[InstCombine] fold zext of masked bit set/clear

This does not solve PR17101, but it is one of the
underlying diffs noted here:
https://bugs.llvm.org/show_bug.cgi?id=17101#c8

We could ease the one-use checks for the 'clear'
(no 'not' op) half of the transform, but I do not
know if that asymmetry would make things better
or worse.

Proofs:
https://rise4fun.com/Alive/uVB

  Name: masked bit set
  %sh1 = shl i32 1, %y
  %and = and i32 %sh1, %x
  %cmp = icmp ne i32 %and, 0
  %r = zext i1 %cmp to i32
  =>
  %s = lshr i32 %x, %y
  %r = and i32 %s, 1

  Name: masked bit clear
  %sh1 = shl i32 1, %y
  %and = and i32 %sh1, %x
  %cmp = icmp eq i32 %and, 0
  %r = zext i1 %cmp to i32
  =>
  %xn = xor i32 %x, -1
  %s = lshr i32 %xn, %y
  %r = and i32 %s, 1

4 years ago[InstCombine] add/adjust tests for masked bit; NFC
Sanjay Patel [Tue, 31 Dec 2019 17:00:07 +0000 (12:00 -0500)]
[InstCombine] add/adjust tests for masked bit; NFC

4 years agoImplement additional traverse() overloads
Stephen Kelly [Sun, 29 Dec 2019 19:21:57 +0000 (19:21 +0000)]
Implement additional traverse() overloads

Summary:
These overloads make it possible to wrap unless(), anyOf(), has() etc
with the traverse matcher.

Reviewers: aaron.ballman

Subscribers: cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D71977

4 years agoUnnest struct in Matcher implementation
Stephen Kelly [Sun, 29 Dec 2019 19:19:55 +0000 (19:19 +0000)]
Unnest struct in Matcher implementation

This allows implementation of the traverse() matcher to surround
matchers like unless().

4 years agoMatch code following lambdas when ignoring invisible nodes
Stephen Kelly [Sun, 29 Dec 2019 19:19:35 +0000 (19:19 +0000)]
Match code following lambdas when ignoring invisible nodes

Reviewers: aaron.ballman

Subscribers: cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D71976

4 years ago[Attributor][Fix] Avoid leaking memory after D68765
Johannes Doerfert [Tue, 31 Dec 2019 16:55:07 +0000 (10:55 -0600)]
[Attributor][Fix] Avoid leaking memory after D68765

4 years agoRevert "[InstCombine] Fix infinite loop due to bitcast <-> phi transforms"
Nikita Popov [Tue, 31 Dec 2019 16:42:38 +0000 (17:42 +0100)]
Revert "[InstCombine] Fix infinite loop due to bitcast <-> phi transforms"

This reverts commit 27a0795943fee0f30b995fe5165428afc2dfd402.

Seems to break test-suite.

4 years ago[PowerPC][NFC] Fix clang-tidy warning
Jinsong Ji [Tue, 31 Dec 2019 16:24:35 +0000 (16:24 +0000)]
[PowerPC][NFC] Fix clang-tidy warning

Reported by
https://results.llvm-merge-guard.org/amd64_debian_testing_clang8-726/clang-tidy.txt

/mnt/disks/ssd0/agent/workspace/amd64_debian_testing_clang8/llvm/lib/Target/PowerPC/PPCISelLowering.cpp:11672:10:
warning: invalid case style for variable 'isEQ'
[readability-identifier-naming]
    bool isEQ = (MI.getOpcode() == PPC::ANDI_rec_1_EQ_BIT ||
         ^~~~
         IsEq
/mnt/disks/ssd0/agent/workspace/amd64_debian_testing_clang8/llvm/lib/Target/PowerPC/PPCISelLowering.cpp:11679:14:
warning: invalid case style for variable 'dl'
[readability-identifier-naming]
    DebugLoc dl = MI.getDebugLoc();
             ^~
             Dl

4 years ago[OpenCL] Remove redundant foreach in OpenCLBuiltins.td; NFC
Sven van Haastregt [Tue, 31 Dec 2019 15:30:02 +0000 (15:30 +0000)]
[OpenCL] Remove redundant foreach in OpenCLBuiltins.td; NFC

Remove various `foreach` declarations where the iterator is used only
once.  This makes the .td file more compact.

4 years ago[AArch64] add test for fsub+fneg; NFC
Sanjay Patel [Tue, 31 Dec 2019 15:25:41 +0000 (10:25 -0500)]
[AArch64] add test for fsub+fneg; NFC

D72015 proposes to restrict the current behavior.

4 years ago[InstCombine] add tests for masked bit set/clear; NFC
Sanjay Patel [Mon, 30 Dec 2019 19:56:57 +0000 (14:56 -0500)]
[InstCombine] add tests for masked bit set/clear; NFC

4 years ago[InstCombine] Fix infinite loop due to bitcast <-> phi transforms
Nikita Popov [Sat, 7 Dec 2019 11:21:22 +0000 (12:21 +0100)]
[InstCombine] Fix infinite loop due to bitcast <-> phi transforms

Fix for https://bugs.llvm.org/show_bug.cgi?id=44245.

The optimizeBitCastFromPhi() and FoldPHIArgOpIntoPHI() end up
fighting against each other, because optimizeBitCastFromPhi()
assumes that bitcasts of loads will get folded. This doesn't happen
here, because a dangling phi node prevents the one-use fold in
https://github.com/llvm/llvm-project/blob/master/llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp#L620-L628 from triggering.

This patch fixes the issue by adding manually removing the old phis.

Differential Revision: https://reviews.llvm.org/D71164

4 years ago[mlir][Linalg] Delete unused LinalgLibraryOps.td
Nicolas Vasilache [Mon, 30 Dec 2019 20:41:12 +0000 (15:41 -0500)]
[mlir][Linalg] Delete unused LinalgLibraryOps.td

Summary: This has been previously renamed to LinalgStructuredOps.td

Reviewers: ftynse

Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, llvm-commits, ftynse

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D72013

4 years ago[llvm-exegesis] Check counters before running
Miloš Stojanović [Tue, 31 Dec 2019 13:14:41 +0000 (14:14 +0100)]
[llvm-exegesis] Check counters before running

Check if the appropriate counters for the specified mode are defined on
the target. This is checked before any other work is done.

Differential Revision: https://reviews.llvm.org/D71927

4 years ago[ARM][TypePromotion] Re-enable by default
Sam Parker [Tue, 31 Dec 2019 11:31:06 +0000 (11:31 +0000)]
[ARM][TypePromotion] Re-enable by default

Re-enable the pass after it was reverted and the bug fixed.

4 years agoFix external-names.c test when separator is \\
Michael Platings [Mon, 30 Dec 2019 08:21:42 +0000 (08:21 +0000)]
Fix external-names.c test when separator is \\

This fixes the following failure:

C:\[...]\llvm\tools\clang\test\VFS\external-names.c:34:26: error: CHECK-DEBUG-EXTERNAL: expected string not found in input
// CHECK-DEBUG-EXTERNAL: ![[Num]] = !DIFile(filename: "{{[^"]*}}Inputs{{.}}external-names.h"
                         ^
[...]
<stdin>:42:54: note: possible intended match here
!10 = !DIFile(filename: "C:/[...]\\llvm\\tools\\clang\\test\\VFS\\Inputs\\external-names.h", directory: "")

Differential Revision: https://reviews.llvm.org/D71991

4 years ago[InstCombine] Don't rewrite phi-of-bitcast when the phi has other users
Connor Abbott [Tue, 31 Dec 2019 11:11:35 +0000 (12:11 +0100)]
[InstCombine] Don't rewrite phi-of-bitcast when the phi has other users

Judging by the existing comments, this was the intention, but the
transform never actually checked if the existing phi's would be removed.
See https://bugs.llvm.org/show_bug.cgi?id=44242 for an example where
this causes much worse code generation on AMDGPU.

Differential Revision: https://reviews.llvm.org/D71209

4 years ago[InstCombine] Add tests for PR44242
Connor Abbott [Tue, 31 Dec 2019 11:10:06 +0000 (12:10 +0100)]
[InstCombine] Add tests for PR44242

Differential Revision: https://reviews.llvm.org/D71260

4 years agoclang-tidy doc: modernize-make-unique has an autofix
Sylvestre Ledru [Tue, 31 Dec 2019 10:56:17 +0000 (11:56 +0100)]
clang-tidy doc: modernize-make-unique has an autofix

4 years ago[Attributor] Suppress unused warnings when assertions are disabled. NFC
Ilya Biryukov [Tue, 31 Dec 2019 09:21:52 +0000 (10:21 +0100)]
[Attributor] Suppress unused warnings when assertions are disabled. NFC

4 years ago[Utils] Deal with occasionally deleted functions
Johannes Doerfert [Fri, 11 Oct 2019 01:23:17 +0000 (20:23 -0500)]
[Utils] Deal with occasionally deleted functions

When functions exist for some but not all run lines we need to be
careful when selecting the prefix. So far, a common prefix was
potentially chosen as there was never a "conflict" that would have
caused otherwise. With this patch we avoid common prefixes if they
are used by run lines that do not emit the function.

Reviewed By: lebedev.ri

Differential Revision: https://reviews.llvm.org/D68850

4 years ago[Attributor] Function signature rewrite infrastructure
Johannes Doerfert [Thu, 10 Oct 2019 06:39:16 +0000 (01:39 -0500)]
[Attributor] Function signature rewrite infrastructure

As part of the Attributor manifest we want to change the signature of
functions. This patch introduces a fairly generic interface to do so.
As a first, very simple, use case, we remove unused arguments. A second
use case, pointer privatization, will be committed with this patch as
well.

A lot of the code and ideas are taken from argument promotion and we
run all argument promotion tests through this framework as well.

Reviewed By: uenoku

Differential Revision: https://reviews.llvm.org/D68765

4 years ago[X86] Slightly improve our attempted error recovery for 64-bit -mno-sse2 in LowerCall...
Craig Topper [Tue, 31 Dec 2019 07:41:35 +0000 (23:41 -0800)]
[X86] Slightly improve our attempted error recovery for 64-bit -mno-sse2 in LowerCallResult to use FP1 if there are two return values.

If the return value is a struct of 2 doubles we need two return
registers.

If SSE2 is disabled we can't return in XMM registers like the ABI says.
After logging an error we attempt to recover by using FP0 instead
of an XMM register. But if the return needs two registers, we may have
already used FP0. So if the register we were supposed to copy to is
XMM1, copy to FP1 in the recovery instead.

This seems to fix the assertion/crash in PR44413.

4 years ago[Utils][Fix] Minor test result change
Johannes Doerfert [Tue, 31 Dec 2019 08:11:30 +0000 (02:11 -0600)]
[Utils][Fix] Minor test result change

4 years ago [NFC] Style cleanup
Shengchen Kan [Tue, 31 Dec 2019 08:08:05 +0000 (16:08 +0800)]
 [NFC] Style cleanup

1. make function Is16BitMemOperand static
2. Use Doxygen features in comment
3. Rename functions to make them start with a lower case letter

4 years ago[Utils] Reuse argument variable names in the body
Johannes Doerfert [Fri, 1 Nov 2019 17:51:26 +0000 (12:51 -0500)]
[Utils] Reuse argument variable names in the body

If we have `int foo(int a) { return a; }` and we run with --function-signature
enabled, we want a single variable declaration for `a` which is reused
later.

Reviewed By: lebedev.ri

Differential Revision: https://reviews.llvm.org/D69722

4 years ago[Utils] Allow update_test_checks to scrub attribute annotations
Johannes Doerfert [Fri, 11 Oct 2019 01:32:04 +0000 (20:32 -0500)]
[Utils] Allow update_test_checks to scrub attribute annotations

Attribute annotations on calls, e.g., #0, are not useful on their own.
This patch adds a flag to update_test_checks.py to scrub them.

Reviewed By: lebedev.ri

Differential Revision: https://reviews.llvm.org/D68851

4 years ago[Attributor] Propagate known align from arguments to call sites arguments
Johannes Doerfert [Tue, 31 Dec 2019 07:27:50 +0000 (01:27 -0600)]
[Attributor] Propagate known align from arguments to call sites arguments

Since the information is known we can simply use it at the call site.
This is especially useful for callbacks but also helps regular calls.

The test changes are mechanical.

4 years ago[Attributor] Use abstract call sites to determine associated arguments
Johannes Doerfert [Thu, 10 Oct 2019 06:19:57 +0000 (01:19 -0500)]
[Attributor] Use abstract call sites to determine associated arguments

This is the second step after D67871 to make use of abstract call sites.
In this patch the argument we associate with a abstract call site
argument can be the one in the callback callee instead of the one in the
callback broker.

Caveat: We cannot allow no-alias arguments for problematic callbacks:
As described in [1], adding no-alias (or restrict) to arguments could
break synchronization as the synchronization effect, e.g., a barrier,
does not "alias" with the pointer anymore. This disables no-alias
annotation for potentially problematic arguments until we implement the
fix described in [1].

Reviewed By: uenoku

Differential Revision: https://reviews.llvm.org/D68008

[1] Compiler Optimizations for OpenMP, J. Doerfert and H. Finkel,
    International Workshop on OpenMP 2018,
    http://compilers.cs.uni-saarland.de/people/doerfert/par_opt18.pdf

4 years ago[Attributor] Annotate the memory behavior of call site arguments
Johannes Doerfert [Tue, 31 Dec 2019 06:57:00 +0000 (00:57 -0600)]
[Attributor] Annotate the memory behavior of call site arguments

Especially for callbacks, annotating the call site arguments is
important. Doing so exposed a too strong dependence of AAMemoryBehavior
on AANoCapture since we handle the case of potentially captured pointers
explicitly.

The changes to the tests are all mechanical.

4 years ago[NFC] Make X86MCCodeEmitter::isPCRel32Branch static
Shengchen Kan [Tue, 31 Dec 2019 07:10:08 +0000 (15:10 +0800)]
[NFC] Make X86MCCodeEmitter::isPCRel32Branch static

4 years agoRevert "DebugInfo: Fix rangesBaseAddress DICompileUnit bitcode serialization/deserial...
David Blaikie [Tue, 31 Dec 2019 06:32:08 +0000 (22:32 -0800)]
Revert "DebugInfo: Fix rangesBaseAddress DICompileUnit bitcode serialization/deserialization"

Seeing some curious CFI failures internally - which makes little sense
to me, as I don't think anyone is using this flag (even us,
internally)... so sounds like a bug in my code somewhere (possibly a
latent one that propagating this flag exposed, not sure). Reverting
while I investigate.

This reverts commit c51b45e32ef7f35c11891f60871aa9c2c04cd991.

4 years ago[NFC] Style cleanup
Shengchen Kan [Tue, 31 Dec 2019 06:23:07 +0000 (14:23 +0800)]
[NFC] Style cleanup

1. Remove function is64BitMode() and use STI.hasFeature(X86::Mode16Bit) directly
2. Use Doxygen features in comment
3. Rename functions to make them start with a lower case letter
4. Format the code with clang-format

4 years ago[mlir] Refactor operation results to use a single use list for all results of the...
River Riddle [Tue, 31 Dec 2019 04:49:47 +0000 (20:49 -0800)]
[mlir] Refactor operation results to use a single use list for all results of the operation.

Summary: A new class is added, IRMultiObjectWithUseList, that allows for representing an IR use list that holds multiple sub values(used in this case for OpResults). This class provides all of the same functionality as the base IRObjectWithUseList, but for specific sub-values. This saves a word per operation result and is a necessary step in optimizing the layout of operation results. For now the use list is placed on the operation itself, so zero-result operations grow by a word. When the work for optimizing layout is finished, this can be moved back to being a trailing object based on memory/runtime benchmarking.

Reviewed By: jpienaar

Differential Revision: https://reviews.llvm.org/D71955

4 years ago[TargetLowering][AMDGPU] Make scalarizeVectorLoad return a pair of SDValues instead...
Craig Topper [Tue, 31 Dec 2019 03:07:36 +0000 (19:07 -0800)]
[TargetLowering][AMDGPU] Make scalarizeVectorLoad return a pair of SDValues instead of creating a MERGE_VALUES node. NFCI

This allows us to clean up some places that were peeking through
the MERGE_VALUES node after the call. By returning the SDValues
directly, we can clean that up.

Unfortunately, there are several call sites in AMDGPU that wanted
the MERGE_VALUES and now need to create their own.

4 years ago[SelectionDAG] Fix copy/paste mistake in comment. NFC
Craig Topper [Tue, 31 Dec 2019 02:01:59 +0000 (18:01 -0800)]
[SelectionDAG] Fix copy/paste mistake in comment. NFC

I think this was copied from scalarizeVectorLoad where that is
what happens.

4 years ago[NFC] Add comments in unit test aix-xcoff-toc.ll to clarify the intent
jasonliu [Tue, 31 Dec 2019 03:29:50 +0000 (03:29 +0000)]
[NFC] Add comments in unit test aix-xcoff-toc.ll to clarify the intent

Address David's post review comment in https://reviews.llvm.org/D71667.
Add comments to clarify what we are testing in that file.

4 years ago[X86] Add test case for PR44412. NFC
Craig Topper [Mon, 30 Dec 2019 22:40:56 +0000 (14:40 -0800)]
[X86] Add test case for PR44412. NFC

4 years ago[CodeGen] Use IRBuilder::CreateFNeg for __builtin_conj
Craig Topper [Mon, 30 Dec 2019 21:25:23 +0000 (13:25 -0800)]
[CodeGen] Use IRBuilder::CreateFNeg for __builtin_conj

This replaces the fsub -0.0 idiom with an fneg instruction. We didn't see to have a test that showed the current codegen. Just some tests for constant folding and a test that was only checking the declare lines for libcalls. The latter just checked that we did not have a declare for @conj when using __builtin_conj.

Differential Revision: https://reviews.llvm.org/D72012

4 years ago[CodeGen] Use CreateFNeg in buildFMulAdd
Craig Topper [Mon, 30 Dec 2019 21:24:08 +0000 (13:24 -0800)]
[CodeGen] Use CreateFNeg in buildFMulAdd

We have an fneg instruction now and should use it instead of the fsub -0.0 idiom. Looks like we had no test that showed that we handled the negation cases here so I've added new tests.

Differential Revision: https://reviews.llvm.org/D72010

4 years agoRemove a redundant `default:` on an exhaustive switch(enum).
Eric Astor [Mon, 30 Dec 2019 21:11:28 +0000 (16:11 -0500)]
Remove a redundant `default:` on an exhaustive switch(enum).

4 years ago[OpenMP][FIX] Generalize a test check line
Johannes Doerfert [Mon, 30 Dec 2019 20:58:28 +0000 (14:58 -0600)]
[OpenMP][FIX] Generalize a test check line

The new check line is compatible with the clang code generation check
line as it allows a 64 and 32 bit value.

I hope this makes the llvm-clang-win-x-armv7l buildbot happy.

4 years ago[libomptarget][nfc] Change unintentional target_impl prefix to kmpc_impl
Jon Chesterfield [Mon, 30 Dec 2019 20:49:56 +0000 (20:49 +0000)]
[libomptarget][nfc] Change unintentional target_impl prefix to kmpc_impl

4 years ago[PowerPC][docs] Update Embedded PowerPC docs in Compiler Writers Info page
Jinsong Ji [Mon, 30 Dec 2019 20:21:46 +0000 (20:21 +0000)]
[PowerPC][docs] Update Embedded PowerPC docs in Compiler Writers Info page

Summary:
Embedded PowerPC are still actively supported, especially SPE...
So update some important references here:

* adding EREF
* adding SPE/VLE ref

Delete deprecated ones into "Other documents..".

Reviewers: #powerpc, jhibbits, hfinkel

Reviewed By: #powerpc, jhibbits

Subscribers: wuzish, merge_guards_bot, nemanjai, shchenz, steven.zhang, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D72008

4 years ago[OpenMP] Use the OpenMPIRBuilder for `omp parallel`
Johannes Doerfert [Thu, 26 Dec 2019 17:23:38 +0000 (11:23 -0600)]
[OpenMP] Use the OpenMPIRBuilder for `omp parallel`

This allows to use the OpenMPIRBuilder for parallel regions. Code was
extracted from D61953 and adapted to work with the new version (D70109).

All but one feature should be supported. An update of this patch will
provide test coverage and privatization other than shared.

Reviewed By: fghanim

Differential Revision: https://reviews.llvm.org/D70290

4 years ago[OpenMP] Use the OpenMPIRBuilder for `omp cancel`
Johannes Doerfert [Fri, 27 Dec 2019 21:53:37 +0000 (15:53 -0600)]
[OpenMP] Use the OpenMPIRBuilder for `omp cancel`

An `omp cancel parallel` needs to be emitted by the OpenMPIRBuilder if
the `parallel` was emitted by the OpenMPIRBuilder. This patch makes
this possible. The cancel logic is shared with the cancel barriers.
Testing is done via unit tests and the clang cancel_codegen.cpp file
once D70290 lands.

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D71948

4 years ago[X86][AsmParser] re-introduce 'offset' operator
Eric Astor [Mon, 30 Dec 2019 19:33:56 +0000 (14:33 -0500)]
[X86][AsmParser] re-introduce 'offset' operator

Summary:
Amend MS offset operator implementation, to more closely fit with its MS counterpart:

    1. InlineAsm: evaluate non-local source entities to their (address) location
    2. Provide a mean with which one may acquire the address of an assembly label via MS syntax, rather than yielding a memory reference (i.e. "offset asm_label" and "$asm_label" should be synonymous
    3. address PR32530

Based on http://llvm.org/D37461

Fix broken test where the break appears unrelated.

- Set up appropriate memory-input rewrites for variable references.

- Intel-dialect assembly printing now correctly handles addresses by adding "offset".

- Pass offsets as immediate operands (using "r" constraint for offsets of locals).

Reviewed By: rnk

Differential Revision: https://reviews.llvm.org/D71436

4 years agoAMDGPU/GlobalISel: Select mul24 intrinsics
Matt Arsenault [Sun, 8 Sep 2019 22:11:51 +0000 (18:11 -0400)]
AMDGPU/GlobalISel: Select mul24 intrinsics

4 years agoTableGen: Fix assert on PatFrags with predicate code
Matt Arsenault [Mon, 30 Dec 2019 17:05:25 +0000 (12:05 -0500)]
TableGen: Fix assert on PatFrags with predicate code

This assumed a single pattern if there was a predicate. Relax this a
bit, and allow multiple patterns as long as they have the same class.

This was only broken for the DAG path. GlobalISel seems to have
handled this correctly already.

4 years ago[X86] Add X86ISD::PCMPGT to SimplifyMultipleUseDemandedBitsForTargetNode.
Craig Topper [Mon, 30 Dec 2019 18:50:04 +0000 (10:50 -0800)]
[X86] Add X86ISD::PCMPGT to SimplifyMultipleUseDemandedBitsForTargetNode.

If only the sign bit is demanded, and the LHS is all zeroes, then
we can bypass the PCMPGT.

4 years ago[test] do not parse ls output for file size; NFCI
Bryan Chan [Fri, 27 Dec 2019 22:26:24 +0000 (17:26 -0500)]
[test] do not parse ls output for file size; NFCI

Parsing `ls -l` output to obtain the size of a file is unreliable; the
exact output format is not specified, and some user or group names may
contain multiple words, causing `cut -f5 -d' '` to extract an incorrect
value. `wc -c`, on the other hand, is portable, and there are precendents
of its use in test cases.

4 years agoAMDGPU/GlobalISel: Re-use MRI available in selector
Matt Arsenault [Sun, 29 Dec 2019 14:05:56 +0000 (09:05 -0500)]
AMDGPU/GlobalISel: Re-use MRI available in selector

4 years agoIgnore "no-frame-pointer-elim" and "no-frame-pointer-elim-non-leaf" in favor of ...
Fangrui Song [Wed, 25 Dec 2019 02:12:15 +0000 (18:12 -0800)]
Ignore "no-frame-pointer-elim" and "no-frame-pointer-elim-non-leaf" in favor of "frame-pointer"

D56351 (included in LLVM 8.0.0) introduced "frame-pointer".  All tests
which use "no-frame-pointer-elim" or "no-frame-pointer-elim-non-leaf"
have been migrated to use "frame-pointer".

Implement UpgradeFramePointerAttributes to upgrade the two obsoleted
function attributes for bitcode. Their semantics are ignored.

Differential Revision: https://reviews.llvm.org/D71863

4 years ago[InstCombine] remove stale comment on test; NFC
Sanjay Patel [Mon, 30 Dec 2019 17:38:49 +0000 (12:38 -0500)]
[InstCombine] remove stale comment on test; NFC

4 years ago[MIPS GlobalISel] Select bitreverse. Recommit
Petar Avramovic [Mon, 30 Dec 2019 17:06:29 +0000 (18:06 +0100)]
[MIPS GlobalISel] Select bitreverse. Recommit

G_BITREVERSE is generated from llvm.bitreverse.<type> intrinsics,
clang genrates these intrinsics from __builtin_bitreverse32 and
__builtin_bitreverse64.
Add lower and narrowscalar for G_BITREVERSE.
Lower G_BITREVERSE on MIPS32.

Recommit notes:
Introduce temporary variables in order to make sure
instructions get inserted into MachineFunction in same order
regardless of compiler used to build llvm.

Differential Revision: https://reviews.llvm.org/D71363

4 years agoAMDGPU/GlobalISel: Select llvm.amdgcn.fmad.ftz
Matt Arsenault [Sun, 8 Sep 2019 21:44:09 +0000 (17:44 -0400)]
AMDGPU/GlobalISel: Select llvm.amdgcn.fmad.ftz

4 years ago[InstCombine] propagate sign argument through nested copysigns
Sanjay Patel [Mon, 30 Dec 2019 16:04:00 +0000 (11:04 -0500)]
[InstCombine] propagate sign argument through nested copysigns

This is another optimization suggested in PR44153:
https://bugs.llvm.org/show_bug.cgi?id=44153

4 years ago[ARM][Thumb][FIX] Add unwinding information to t4
Diogo Sampaio [Mon, 30 Dec 2019 15:43:32 +0000 (15:43 +0000)]
[ARM][Thumb][FIX] Add unwinding information to t4

Summary:
Add missing part of patch D71361. Now that the stack-frame
can be operated using a addw/subw instruction, they should
appear in the unwinding list.

Reviewers: dmgreen, efriedma

Reviewed By: dmgreen

Subscribers: kristof.beyls, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D72000

4 years agoAMDGPU/GlobalISel: Add select test for fexp2
Matt Arsenault [Sun, 8 Sep 2019 19:39:00 +0000 (15:39 -0400)]
AMDGPU/GlobalISel: Add select test for fexp2

4 years agoGlobalISel: moreElementsVector for FP min/max
Matt Arsenault [Sat, 27 Jul 2019 21:47:08 +0000 (17:47 -0400)]
GlobalISel: moreElementsVector for FP min/max

4 years agoAMDGPU: Improve llvm.round.f64 lowering for CI+
Matt Arsenault [Tue, 24 Dec 2019 16:07:45 +0000 (11:07 -0500)]
AMDGPU: Improve llvm.round.f64 lowering for CI+

The path already used for f16/f32 works a lot better when v_trunc_f64
is available.

4 years agoAMDGPU: Generate check lines
Matt Arsenault [Tue, 24 Dec 2019 16:24:30 +0000 (11:24 -0500)]
AMDGPU: Generate check lines

4 years agoAMDGPU/GlobalISel: Account for G_PHI result bank
Matt Arsenault [Sat, 21 Dec 2019 20:34:30 +0000 (15:34 -0500)]
AMDGPU/GlobalISel: Account for G_PHI result bank

Sometimes the result bank of the phi is already assigned to something,
and should not be ignored. This is in preparation for additional
boolean phi handling changes.

Also refine the logic to fix some cases that were incorrectly deciding
to use SGPRs.

4 years ago[PowerPC] Legalize rounding nodes
Nemanja Ivanovic [Mon, 30 Dec 2019 13:38:27 +0000 (07:38 -0600)]
[PowerPC] Legalize rounding nodes

VSX provides a full complement of rounding instructions yet we somehow ended up
with some of them legal and others not. This just legalizes all of the FP
rounding nodes and the FP -> int rounding nodes with unsafe math.

Differential revision: https://reviews.llvm.org/D69949

4 years agoRevert "[MIPS GlobalISel] Select bitreverse"
Dmitri Gribenko [Mon, 30 Dec 2019 13:28:56 +0000 (14:28 +0100)]
Revert "[MIPS GlobalISel] Select bitreverse"

This reverts commit dbc136e0fe7e14c64dcb78e72321bb41af60afa4.
It broke buildbots:
http://lab.llvm.org:8011/builders/clang-x86_64-debian-fast/builds/21066

4 years ago[ARM] Sink splat to ICmp
David Green [Mon, 30 Dec 2019 09:39:14 +0000 (09:39 +0000)]
[ARM] Sink splat to ICmp

This adds ICmp to the list of instructions that we sink a splat to in a
loop, allowing the register forms of instructions to be selected more
often. It does not add FCmp yet as the results look a little odd, trying
to keep the register in an float reg and having to move it back to a GPR.

Differential Revision: https://reviews.llvm.org/D70997

4 years ago[ARM] MVE sink ICmp test. NFC
David Green [Mon, 30 Dec 2019 09:28:10 +0000 (09:28 +0000)]
[ARM] MVE sink ICmp test. NFC

4 years ago[LV][NFC] Keep dominator tree up to date during vectorization.
Evgeniy Brevnov [Mon, 9 Dec 2019 08:11:30 +0000 (15:11 +0700)]
[LV][NFC] Keep dominator tree up to date during vectorization.

4 years ago[LV][NFC] Some refactoring and renaming to facilitate next change.
Evgeniy Brevnov [Thu, 5 Dec 2019 10:57:27 +0000 (17:57 +0700)]
[LV][NFC] Some refactoring and renaming to facilitate next change.

4 years ago[ARM][THUMB2] Allow emitting T3 types of add and sub
Diogo Sampaio [Mon, 30 Dec 2019 10:59:45 +0000 (10:59 +0000)]
[ARM][THUMB2] Allow emitting T3 types of add and sub

Summary:
This patch allows to emit thumb2 add and sub
instructions with 12 bit immediates in the
emitT2RegPlusImmediate function.
- Splitting parts of the D70680

Reviewers: eli.friedman, olista01, efriedma

Reviewed By: efriedma

Subscribers: efriedma, kristof.beyls, hiraditya, dmgreen, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D71361

4 years ago[OpenCL] Add mipmap builtin functions
Sven van Haastregt [Mon, 30 Dec 2019 10:47:58 +0000 (10:47 +0000)]
[OpenCL] Add mipmap builtin functions

Add the mipmap builtin functions from the OpenCL extension
specification.

Patch by Pierre Gondois and Sven van Haastregt.

4 years ago[MIPS GlobalISel] Select bitreverse
Petar Avramovic [Mon, 30 Dec 2019 10:26:45 +0000 (11:26 +0100)]
[MIPS GlobalISel] Select bitreverse

G_BITREVERSE is generated from llvm.bitreverse.<type> intrinsics,
clang genrates these intrinsics from __builtin_bitreverse32 and
__builtin_bitreverse64.
Add lower and narrowscalar for G_BITREVERSE.
Lower G_BITREVERSE on MIPS32.

Differential Revision: https://reviews.llvm.org/D71363

4 years ago[MIPS GlobalISel] Select bswap
Petar Avramovic [Mon, 30 Dec 2019 10:13:22 +0000 (11:13 +0100)]
[MIPS GlobalISel] Select bswap

G_BSWAP is generated from llvm.bswap.<type> intrinsics, clang genrates
these intrinsics from __builtin_bswap32 and __builtin_bswap64.
Add lower and narrowscalar for G_BSWAP.
Lower G_BSWAP on MIPS32, select G_BSWAP on MIPS32 revision 2 and later.

Differential Revision: https://reviews.llvm.org/D71362

4 years ago[MCP] Add stats for backward copy propagation. NFC.
Kai Luo [Mon, 30 Dec 2019 08:31:41 +0000 (16:31 +0800)]
[MCP] Add stats for backward copy propagation. NFC.

4 years ago[opt] Fix run-twice crash and detection problem
Peter Kokai [Mon, 30 Dec 2019 08:22:55 +0000 (00:22 -0800)]
[opt] Fix run-twice crash and detection problem

1. Execute `opt -run-twice a.ll` with in a terminal will crash.
   https://bugs.llvm.org/show_bug.cgi?id=44382
2. `-run-twice` saves output into two buffers and compares them.
   When outputing the result is disabled, that produces two empty string thus
   they are going to be equal all the time resulting false-positive results.

The proposed solution is to generate the results even if the output will not be
emitted, as that is required for the comparision.

Differential Revision: https://reviews.llvm.org/D71967

4 years ago[Diagnostic] Add ftabstop to -Wmisleading-indentation
Tyker [Sun, 29 Dec 2019 23:14:20 +0000 (00:14 +0100)]
[Diagnostic] Add ftabstop to -Wmisleading-indentation

Summary:
this allow much better support of codebases like the linux kernel that mix tabs and spaces.

-ftabstop=//Width// allow specifying how large tabs are considered to be.

Reviewers: xbolva00, aaron.ballman, rsmith

Reviewed By: aaron.ballman

Subscribers: jyknight, riccibruno, rsmith, nathanchance

Differential Revision: https://reviews.llvm.org/D71037

4 years ago[NFC] Add test for load-insert-store pattern
Qiu Chaofan [Mon, 30 Dec 2019 08:09:07 +0000 (16:09 +0800)]
[NFC] Add test for load-insert-store pattern

This patch adds necessary test cases for load-update-store pattern
which only updates single element of vector.

Differential Revision: https://reviews.llvm.org/D71886

4 years ago[Attributor] Use `changeUseAfterManifest` in AAValueSimplify manifest
Hideto Ueno [Mon, 30 Dec 2019 08:08:48 +0000 (17:08 +0900)]
[Attributor] Use `changeUseAfterManifest` in AAValueSimplify manifest

Summary: This patch makes `AAValueSimplify` use `changeUsesAfterManifest` in `manifest`. This will invoke simple folding after the manifest.

Reviewers: jdoerfert, sstefan1

Reviewed By: jdoerfert

Subscribers: hiraditya, arphaman, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D71972

4 years ago[ELF][PPC64] Improve "call lacks nop" diagnostic and make it compatible with GCC...
Fangrui Song [Wed, 18 Dec 2019 00:45:04 +0000 (16:45 -0800)]
[ELF][PPC64] Improve "call lacks nop" diagnostic and make it compatible with GCC<5.5 and GCC<6.4

GCC before r245813 (https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79439)
did not emit nop after b/bl. This can happen with recursive calls.
r245813 was back ported to GCC 5.5 and GCC 6.4.

This is common, for example, libstdc++.a(locale.o) shipped with GCC 4.9
and many objects in netlib lapack can cause lld to error.  gold allows
such calls to the same section. Our __plt_foo symbol's `section` field
is used for ThunkSection, so we can't implement a similar loosen rule
easily. But we can make use of its `file` field which is currently NULL.

Differential Revision: https://reviews.llvm.org/D71639

4 years ago[ELF][PPC32] Implement IPLT code sequence for non-preemptible IFUNC
Fangrui Song [Tue, 17 Dec 2019 19:23:37 +0000 (11:23 -0800)]
[ELF][PPC32] Implement IPLT code sequence for non-preemptible IFUNC

Similar to D71509 (EM_PPC64), on EM_PPC, the IPLT code sequence should
be similar to a PLT call stub. Unlike EM_PPC64, EM_PPC -msecure-plt has
small/large PIC model differences.

* -fpic/-fpie: R_PPC_PLTREL24 r_addend=0.  The call stub loads an address relative to `_GLOBAL_OFFSET_TABLE_`.
* -fPIC/-fPIE: R_PPC_PLTREL24 r_addend=0x8000. (A partial linked object
  file may have an addend larger than 0x8000.) The call stub loads an address relative to .got2+0x8000.

Just assume large PIC model for now. This patch makes:

  // clang -fuse-ld=lld -msecure-plt -fno-pie -no-pie a.c
  // clang -fuse-ld=lld -msecure-plt -fPIE -pie a.c
  #include <stdio.h>
  static void impl(void) { puts("meow"); }
  void thefunc(void) __attribute__((ifunc("resolver")));
  void *resolver(void) { return &impl; }
  int main(void) {
    thefunc();
    void (*theptr)(void) = &thefunc;
    theptr();
  }

work on Linux glibc. -fpie will crash because the compiler and the
linker do not agree on the value which r30 stores (_GLOBAL_OFFSET_TABLE_
vs .got2+0x8000).

Differential Revision: https://reviews.llvm.org/D71621

4 years ago[ELF][PPC64] Implement IPLT code sequence for non-preemptible IFUNC
Fangrui Song [Sat, 14 Dec 2019 02:30:21 +0000 (18:30 -0800)]
[ELF][PPC64] Implement IPLT code sequence for non-preemptible IFUNC

Non-preemptible IFUNC are placed in in.iplt (.glink on EM_PPC64).  If
there is a non-GOT non-PLT relocation, for pointer equality, we change
the type of the symbol from STT_IFUNC and STT_FUNC and bind it to the
.glink entry.

On EM_386, EM_X86_64, EM_ARM, and EM_AARCH64, the PLT code sequence
loads the address from its associated .got.plt slot. An IPLT also has an
associated .got.plt slot and can use the same code sequence.

On EM_PPC64, the PLT code sequence is actually a bl instruction in
.glink .  It jumps to `__glink_PLTresolve` (the PLT header). and
`__glink_PLTresolve` computes the .plt slot (relocated by
R_PPC64_JUMP_SLOT).

An IPLT does not have an associated R_PPC64_JUMP_SLOT, so we cannot use
`bl` in .iplt . Instead, create a call stub which has a similar code
sequence as PPC64PltCallStub. We don't save the TOC pointer, so such
scenarios will not work: a function pointer to a non-preemptible ifunc,
which resolves to a function defined in another DSO. This is the
restriction described by https://sourceware.org/glibc/wiki/GNU_IFUNC
(though on many architectures it works in practice):

  Requirement (a): Resolver must be defined in the same translation unit as the implementations.

If an ifunc is taken address but not called, technically we don't need
an entry for it, but we currently do that.

This patch makes

  // clang -fuse-ld=lld -fno-pie -no-pie a.c
  // clang -fuse-ld=lld -fPIE -pie a.c
  #include <stdio.h>
  static void impl(void) { puts("meow"); }
  void thefunc(void) __attribute__((ifunc("resolver")));
  void *resolver(void) { return &impl; }
  int main(void) {
    thefunc();
    void (*theptr)(void) = &thefunc;
    theptr();
  }

work on Linux glibc and FreeBSD. Calling a function pointer pointing to
a Non-preemptible IFUNC never worked before.

Differential Revision: https://reviews.llvm.org/D71509

4 years ago[SelectionDAT] Simplify SelectionDAGBuilder::visitInlineAsm
Fangrui Song [Mon, 30 Dec 2019 04:17:31 +0000 (20:17 -0800)]
[SelectionDAT] Simplify SelectionDAGBuilder::visitInlineAsm

Indirect C_Immediate or C_Other constraints have been excluded.

Also simplify an unneeded change to indirect 'X' by D60942.

4 years ago[CMake] Added remote test execution support into CrossWinToARMLinux CMake cache file.
Vladimir Vereschaka [Tue, 17 Dec 2019 20:44:26 +0000 (12:44 -0800)]
[CMake] Added remote test execution support into CrossWinToARMLinux CMake cache file.

Added two confguration argument to provide a host name and SSH user name
to run the tests on the remote target host.

* REMOTE_TEST_HOST  - remote host name or address.
* REMOTE_TEST_USER  - passwordless SSH account name.

Differential Revision: https://reviews.llvm.org/D71625

4 years ago[PowerPC] Exploit the rlwinm instructions for "and" with constant
QingShan Zhang [Mon, 30 Dec 2019 03:18:31 +0000 (03:18 +0000)]
[PowerPC] Exploit the rlwinm instructions for "and" with constant

For now, PowerPC will using several instructions to get the constant and "and" it with the following case:

define i32 @test1(i32 %a) {
  %and = and i32 %a, -2
  ret i32 %and
}

However, we could exploit it with the rotate mask instructions.
               MB  ME
+----------------------+
|xxxxxxxxxxx00011111000|
+----------------------+
 0         32         64
Notice that, we can only do it if the MB is larger than 32 and MB <= ME as
RLWINM will replace the content of [0 - 32) with [32 - 64) even we didn't rotate it.

Differential Revision: https://reviews.llvm.org/D71829

4 years ago[X86] Use APInt::isOneValue and ConstantSDNode::isOne. NFC
Craig Topper [Mon, 30 Dec 2019 00:41:28 +0000 (16:41 -0800)]
[X86] Use APInt::isOneValue and ConstantSDNode::isOne. NFC

These are implemented slightly more efficiently than comparing
to 1 in the case that the value is more than 64 bits.

4 years ago[X86] Use isOneConstant to simplify some code. NFC
Craig Topper [Mon, 30 Dec 2019 00:22:10 +0000 (16:22 -0800)]
[X86] Use isOneConstant to simplify some code. NFC

4 years ago[X86] Remove dyn_casts to ConstantSDNode for operand 1 of X86ISD::VSRLI/VSRAI/VSRLI...
Craig Topper [Mon, 30 Dec 2019 00:19:32 +0000 (16:19 -0800)]
[X86] Remove dyn_casts to ConstantSDNode for operand 1 of X86ISD::VSRLI/VSRAI/VSRLI. Use getConstantOperandVal and APInt operations.

These nodes should only ever be formed with an i8 TargetConstant
so we don't need to check for it to be a constant. It's also
always 8-bits so we don't need to use APInt compare functions.

4 years ago[SelectionDAG] Disallow indirect "i" constraint
Fangrui Song [Sun, 29 Dec 2019 23:53:46 +0000 (15:53 -0800)]
[SelectionDAG] Disallow indirect "i" constraint

This allows us to delete InlineAsm::Constraint_i workarounds in
SelectionDAGISel::SelectInlineAsmMemoryOperand overrides and
TargetLowering::getInlineAsmMemConstraint overrides.

They were introduced to X86 in r237517 to prevent crashes for
constraints like "=*imr". They were later copied to other targets.

4 years ago[lldb][NFC] Simplify ClangASTContext::GetTypeForDecl
Raphael Isemann [Sun, 29 Dec 2019 22:01:53 +0000 (23:01 +0100)]
[lldb][NFC] Simplify ClangASTContext::GetTypeForDecl

Also removes the GetASTContext call from this code.

4 years ago[lldb][NFC] Make integer types functions in ClangASTContext not static
Raphael Isemann [Sun, 29 Dec 2019 20:28:31 +0000 (21:28 +0100)]
[lldb][NFC] Make integer types functions in ClangASTContext not static

These functions need a ClangASTContext instance that we would otherwise
recalculate by calling GetASTContext (which is no longer necessary with
this patch).