OSDN Git Service

android-x86/external-llvm.git
7 years ago[Reassociate] Add negated value of negative constant to the Duplicates list.
Chad Rosier [Thu, 23 Feb 2017 18:49:03 +0000 (18:49 +0000)]
[Reassociate] Add negated value of negative constant to the Duplicates list.

In OptimizeAdd, we scan the operand list to see if there are any common factors
between operands that can be factored out to reduce the number of multiplies
(e.g., 'A*A+A*B*C+D' -> 'A*(A+B*C)+D'). For each operand of the operand list, we
only consider unique factors (which is tracked by the Duplicate set). Now if we
find a factor that is a negative constant, we add the negated value as a factor
as well, because we can percolate the negate out. However, we mistakenly don't
add this negated constant to the Duplicates set.

Consider the expression A*2*-2 + B. Obviously, nothing to factor.

For the added value A*2*-2 we over count 2 as a factor without this change,
which causes the assert reported in PR30256.  The problem is that this code is
assuming that all the multiply operands of the add are already reassociated.
This change avoids the issue by making OptimizeAdd tolerate multiplies which
haven't been completely optimized; this sort of works, but we're doing wasted
work: we'll end up revisiting the add later anyway.

Another possible approach would be to enforce RPO iteration order more strongly.
If we have RedoInsts, we process them immediately in RPO order, rather than
waiting until we've finished processing the whole function. Intuitively, it
seems like the natural approach: reassociation works on expression trees, so
the optimization only works in one direction. That said, I'm not sure how
practical that is given the current Reassociate; the "optimal" form for an
expression depends on its use list (see all the uses of "user_back()"), so
Reassociate is really an iterative optimization of sorts, so any changes here
would probably get messy.

PR30256

Differential Revision: https://reviews.llvm.org/D30228

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@296003 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse base discriminator in sample pgo profile matching.
Dehao Chen [Thu, 23 Feb 2017 18:27:45 +0000 (18:27 +0000)]
Use base discriminator in sample pgo profile matching.

Summary: The discriminator has been encoded, and only the base discriminator should be used during profile matching.

Reviewers: dblaikie, davidxl

Reviewed By: dblaikie, davidxl

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D30218

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295999 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Hexagon] Avoid IMPLICIT_DEFs as new-value producers
Krzysztof Parzyszek [Thu, 23 Feb 2017 17:47:34 +0000 (17:47 +0000)]
[Hexagon] Avoid IMPLICIT_DEFs as new-value producers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295997 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LazyMachineBFI] Reimplement with getAnalysisIfAvailable
Adam Nemet [Thu, 23 Feb 2017 17:30:01 +0000 (17:30 +0000)]
[LazyMachineBFI] Reimplement with getAnalysisIfAvailable

Since LoopInfo is not available in machine passes as universally as in IR
passes, using the same approach for OptimizationRemarkEmitter as we did for IR
will run LoopInfo and DominatorTree unnecessarily.  (LoopInfo is not used
lazily by ORE.)

To fix this, I am modifying the approach I took in D29836.  LazyMachineBFI now
uses its client passes including MachineBFI itself that are available or
otherwise compute them on the fly.

So for example GreedyRegAlloc, since it's already using MBFI, will reuse that
instance.  On the other hand, AsmPrinter in Justin's patch will generate DT,
LI and finally BFI on the fly.

(I am of course wondering now if the simplicity of this approach is even
preferable in IR.  I will do some experiments.)

Testing is provided by an updated version of D29837 which requires Justin's
patch to bring ORE to the AsmPrinter.

Differential Revision: https://reviews.llvm.org/D30128

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295996 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AddressSanitizer] Add PS4 offset
Filipe Cabecinhas [Thu, 23 Feb 2017 17:10:28 +0000 (17:10 +0000)]
[AddressSanitizer] Add PS4 offset

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295994 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine] use loop instead of recursion to peek through FPExt; NFCI
Sanjay Patel [Thu, 23 Feb 2017 16:39:51 +0000 (16:39 +0000)]
[InstCombine] use loop instead of recursion to peek through FPExt; NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295992 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine] use 'match' to reduce code; NFCI
Sanjay Patel [Thu, 23 Feb 2017 16:26:03 +0000 (16:26 +0000)]
[InstCombine] use 'match' to reduce code; NFCI

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295991 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU/SI: Fix trunc i16 pattern
Jan Vesely [Thu, 23 Feb 2017 16:12:21 +0000 (16:12 +0000)]
AMDGPU/SI: Fix trunc i16 pattern

Hit on ASICs that support 16bit instructions.

Differential Revision: https://reviews.llvm.org/D30281

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295990 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoStrip trailing whitespace.
Simon Pilgrim [Thu, 23 Feb 2017 16:07:04 +0000 (16:07 +0000)]
Strip trailing whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295989 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Hexagon] Patterns for CTPOP, BSWAP and BITREVERSE
Krzysztof Parzyszek [Thu, 23 Feb 2017 15:02:09 +0000 (15:02 +0000)]
[Hexagon] Patterns for CTPOP, BSWAP and BITREVERSE

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295981 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[docs] Add information about how to checkout polly to getting started page
Tobias Grosser [Thu, 23 Feb 2017 14:27:07 +0000 (14:27 +0000)]
[docs] Add information about how to checkout polly to getting started page

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295974 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] GlobalISel: Lower call returns
Diana Picus [Thu, 23 Feb 2017 14:18:41 +0000 (14:18 +0000)]
[ARM] GlobalISel: Lower call returns

Introduce a common ValueHandler for call returns and formal arguments, and
inherit two different versions for handling the differences (at the moment the
only difference is the way physical registers are marked as used).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295973 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong
Alexey Bataev [Thu, 23 Feb 2017 13:37:09 +0000 (13:37 +0000)]
[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong
result

Summary:
If the same value is used several times as an extra value, SLP
vectorizer takes it into account only once instead of actual number of
using.
For example:
```
int val = 1;
for (int y = 0; y < 8; y++) {
  for (int x = 0; x < 8; x++) {
    val = val + input[y * 8 + x] + 3;
  }
}
```
We have 2 extra rguments: `1` - initial value of horizontal reduction
and `3`, which is added 8*8 times to the reduction. Before the patch we
added `1` to the reduction value and added once `3`, though it must be
added 64 times.

Reviewers: mkuper, mzolotukhin

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D30262

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295972 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] GlobalISel: Lower call parameters in regs
Diana Picus [Thu, 23 Feb 2017 13:25:43 +0000 (13:25 +0000)]
[ARM] GlobalISel: Lower call parameters in regs

Add support for lowering calls with parameters than can fit into regs.  Use the
same ValueHandler that we used for function returns, but rename it to match its
new, extended purpose.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295971 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][AVX] Disable VCVTSS2SD & VCVTSD2SS memory folding and fix the register class...
Ayman Musa [Thu, 23 Feb 2017 13:15:44 +0000 (13:15 +0000)]
[X86][AVX] Disable VCVTSS2SD & VCVTSD2SS memory folding and fix the register class of their first input when creating node in fast-isel.

(Quick fix to buildbot failure after rL295940 commit).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295970 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[mips][ias] Further relax operands of certain assembly instructions
Simon Dardis [Thu, 23 Feb 2017 12:40:58 +0000 (12:40 +0000)]
[mips][ias] Further relax operands of certain assembly instructions

This patch adjusts the most relaxed predicate of immediate operands to accept
immediate forms such as ~(0xf0000000|0x000f00000). Previously these forms
would be accepted by GAS and rejected by IAS.

This partially resolves PR/30383.

Thanks to Sean Bruno for reporting the issue!

Reviewers: slthakur, seanbruno

Differential Revision: https://reviews.llvm.org/D29218

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295965 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix assertion failure in ARMConstantIslandPass.
Kristof Beyls [Thu, 23 Feb 2017 12:24:55 +0000 (12:24 +0000)]
Fix assertion failure in ARMConstantIslandPass.

The ARMConstantIslandPass didn't have support for handling accesses to
constant island objects through ARM::t2LDRBpci instructions. This adds
support for that.

This fixes PR31997.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295964 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix signed/unsigned comparison warning on MSVC
Simon Pilgrim [Thu, 23 Feb 2017 12:00:34 +0000 (12:00 +0000)]
Fix signed/unsigned comparison warning on MSVC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295962 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong"
Alexey Bataev [Thu, 23 Feb 2017 11:09:35 +0000 (11:09 +0000)]
Revert "[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong"

This reverts commit 7c5141e577d9efd1c8e3087566a38ce6b3a41a84.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295957 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong
Alexey Bataev [Thu, 23 Feb 2017 10:57:15 +0000 (10:57 +0000)]
[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong
result

Summary:
If the same value is used several times as an extra value, SLP
vectorizer takes it into account only once instead of actual number of
using.
For example:
```
int val = 1;
for (int y = 0; y < 8; y++) {
  for (int x = 0; x < 8; x++) {
    val = val + input[y * 8 + x] + 3;
  }
}
```
We have 2 extra rguments: `1` - initial value of horizontal reduction
and `3`, which is added 8*8 times to the reduction. Before the patch we
added `1` to the reduction value and added once `3`, though it must be
added 64 times.

Reviewers: mkuper, mzolotukhin

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D30262

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295956 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong"
Alexey Bataev [Thu, 23 Feb 2017 09:59:29 +0000 (09:59 +0000)]
Revert "[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong"

This reverts commit d83c81ee6a8dea662808ac22b396d1bb0595c89d.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295951 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong
Alexey Bataev [Thu, 23 Feb 2017 09:40:38 +0000 (09:40 +0000)]
[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong
result

Summary:
If the same value is used several times as an extra value, SLP
vectorizer takes it into account only once instead of actual number of
using.
For example:
```
int val = 1;
for (int y = 0; y < 8; y++) {
  for (int x = 0; x < 8; x++) {
    val = val + input[y * 8 + x] + 3;
  }
}
```
We have 2 extra rguments: `1` - initial value of horizontal reduction
and `3`, which is added 8*8 times to the reduction. Before the patch we
added `1` to the reduction value and added once `3`, though it must be
added 64 times.

Reviewers: mkuper, mzolotukhin

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D30262

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295949 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][AVX512] Remove VCVTSS2SDZ & VCVTSD2SSZ from memory folding tables as they intro...
Ayman Musa [Thu, 23 Feb 2017 08:13:36 +0000 (08:13 +0000)]
[X86][AVX512] Remove VCVTSS2SDZ & VCVTSD2SSZ from memory folding tables as they introduce new read dependency when folding.

(Quick fix to buildbot fail).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295946 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][AVX512] Change VCVTSS2SD and VCVTSD2SS node types to keep consistency between...
Ayman Musa [Thu, 23 Feb 2017 07:24:21 +0000 (07:24 +0000)]
[X86][AVX512] Change VCVTSS2SD and VCVTSD2SS node types to keep consistency between VEX/EVEX versions.

AVX versions of the converts work on f32/f64 types, while AVX512 version work on vectors.

Differential Revision: https://reviews.llvm.org/D29988

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295940 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoLoadStoreVectorizer: Split even sized illegal chains properly
Matt Arsenault [Thu, 23 Feb 2017 03:58:53 +0000 (03:58 +0000)]
LoadStoreVectorizer: Split even sized illegal chains properly

Implement isLegalToVectorizeLoadChain for AMDGPU to avoid
producing private address spaces accesses that will need to be
split up later. This was doing the wrong thing in the case
where the queried chain was an even number of elements.

A possible <4 x i32> store was being split into
store <2 x i32>
store i32
store i32

rather than
store <2 x i32>
store <2 x i32>

when legal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295933 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][IR] In AutoUpgrade, check explicitly for xop.vpcmov and xop.vpcmov.256 instead...
Craig Topper [Thu, 23 Feb 2017 03:22:14 +0000 (03:22 +0000)]
[X86][IR] In AutoUpgrade, check explicitly for xop.vpcmov and xop.vpcmov.256 instead of anything starting with xop.vpcmov

There were some older intrinsics that only existed for less than a month in 2012 that still exist in some out of tree test files that start with this string, but aren't able to be handled by the current upgrade code and fire an assert. Now we'll go back to treating them as not intrinsics at all and just passing them through to output.

Fixes PR32041, sort of.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295930 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoTargetOptions: Fix not accounting for NoSignedZerosFPMath in ==
Matt Arsenault [Thu, 23 Feb 2017 03:16:44 +0000 (03:16 +0000)]
TargetOptions: Fix not accounting for NoSignedZerosFPMath in ==

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295928 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoTest if we can use raw strings on all platforms compiling LLVM.
Matthias Braun [Thu, 23 Feb 2017 01:09:01 +0000 (01:09 +0000)]
Test if we can use raw strings on all platforms compiling LLVM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295917 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoExplicitly state the behavior of inbounds with a null pointer.
Eli Friedman [Thu, 23 Feb 2017 00:48:18 +0000 (00:48 +0000)]
Explicitly state the behavior of inbounds with a null pointer.

See https://llvm.org/bugs/show_bug.cgi?id=31439; this reflects LLVM's
behavior in practice, and should be compatible with C/C++ rules.

Differential Revision: https://reviews.llvm.org/D28026

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295916 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Replace disabled exp inputs with undef
Matt Arsenault [Thu, 23 Feb 2017 00:44:03 +0000 (00:44 +0000)]
AMDGPU: Replace disabled exp inputs with undef

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295914 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Add another BFE pattern
Matt Arsenault [Thu, 23 Feb 2017 00:23:43 +0000 (00:23 +0000)]
AMDGPU: Add another BFE pattern

This is the pattern that falls out of the instruction's
definition if offset == 0.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295912 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Use clamp with f64
Matt Arsenault [Wed, 22 Feb 2017 23:53:37 +0000 (23:53 +0000)]
AMDGPU: Use clamp with f64

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295908 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert r295868 because it breaks a different SLP lit test.
Michael Kuperstein [Wed, 22 Feb 2017 23:35:13 +0000 (23:35 +0000)]
Revert r295868 because it breaks a different SLP lit test.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295906 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Fold FP clamp as modifier bit
Matt Arsenault [Wed, 22 Feb 2017 23:27:53 +0000 (23:27 +0000)]
AMDGPU: Fold FP clamp as modifier bit

The manual is unclear on the details of this. It's not
clear to me if denormals are not allowed with clamp,
or if that is only omod. Not allowing denorms for
fp16 or fp64 isn't useful so I also question if that
is really a restriction. Same with whether this is valid
without IEEE mode enabled.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295905 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU : Update TrapCode based on Trap Handler ABI.
Wei Ding [Wed, 22 Feb 2017 23:22:19 +0000 (23:22 +0000)]
AMDGPU : Update TrapCode based on Trap Handler ABI.

Differential Revision: http://reviews.llvm.org/D30232

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295904 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[libFuzzer] Update traces hooks test after r293741
Justin Bogner [Wed, 22 Feb 2017 23:12:36 +0000 (23:12 +0000)]
[libFuzzer] Update traces hooks test after r293741

This test now passes on darwin.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295902 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[libFuzzer] Mark a test that infinite loops as unsupported
Justin Bogner [Wed, 22 Feb 2017 23:05:17 +0000 (23:05 +0000)]
[libFuzzer] Mark a test that infinite loops as unsupported

We need to investigate this, but for now it just causes too much
headache when trying to run these tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295900 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Add replacement bfe intrinsics
Matt Arsenault [Wed, 22 Feb 2017 23:04:58 +0000 (23:04 +0000)]
AMDGPU: Add replacement bfe intrinsics

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295899 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[InstCombine] don't try SimplifyDemandedInstructionBits from add/sub because it's...
Sanjay Patel [Wed, 22 Feb 2017 23:01:12 +0000 (23:01 +0000)]
[InstCombine] don't try SimplifyDemandedInstructionBits from add/sub because it's slow and unlikely to succeed

Notably, no regression tests change when we remove these calls, and these are expensive calls.

The motivation comes from the general acknowledgement that the compiler is getting slower:
http://lists.llvm.org/pipermail/llvm-dev/2017-January/109188.html
http://lists.llvm.org/pipermail/llvm-dev/2016-December/108279.html

And specifically the test case attached to PR32037:
https://bugs.llvm.org//show_bug.cgi?id=32037

Profiling the middle-end (opt) part of the compile:
$ ./opt -O2 row_common.bc -o /dev/null

...visitAdd and visitSub are near the top of the instcombine list, and the calls to SimplifyDemandedInstructionBits()
are high within each of those. Those calls account for 1%+ of the opt time in either debug or release profiles. And
that's the rough win I see from this patch when testing opt built release from r295864 on an iMac with Haswell 4GHz
(model 4790K).

It seems unlikely that we'd be able to eliminate add/sub or change their operands given that add/sub normally affect
all bits, and the PR32037 example shows no IR difference after this change using -O2.

Also worth noting - the code comment in visitAdd:
// This handles stuff like (X & 254)+1 -> (X&254)|1
...isn't true. That transform is handled later with a call to haveNoCommonBitsSet().

Differential Revision: https://reviews.llvm.org/D30270

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295898 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVR] Disable integrated assembler for a few tests
Dylan McKay [Wed, 22 Feb 2017 22:41:13 +0000 (22:41 +0000)]
[AVR] Disable integrated assembler for a few tests

Fixes the build.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295895 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[CodeGen] Fix some Clang-tidy modernize and Include What You Use warnings; other...
Eugene Zelenko [Wed, 22 Feb 2017 22:32:51 +0000 (22:32 +0000)]
[CodeGen] Fix some Clang-tidy modernize and Include What You Use warnings; other minor fixes (NFC).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295893 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Hexagon] Implement @llvm.readcyclecounter()
Krzysztof Parzyszek [Wed, 22 Feb 2017 22:28:47 +0000 (22:28 +0000)]
[Hexagon] Implement @llvm.readcyclecounter()

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295892 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Don't add emergency stack slot if all spills are SGPR->VGPR
Matt Arsenault [Wed, 22 Feb 2017 22:23:32 +0000 (22:23 +0000)]
AMDGPU: Don't add emergency stack slot if all spills are SGPR->VGPR

This should avoid reporting any stack needs to be allocated in the
case where no stack is truly used. An unused stack slot is still
left around in other cases where there are real stack objects
but no spilling occurs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295891 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoPredicateInfo: Support switch statements
Daniel Berlin [Wed, 22 Feb 2017 22:20:58 +0000 (22:20 +0000)]
PredicateInfo: Support switch statements

Summary:
Depends on D29606 and D29682

Makes us pass GVN's edge.ll (we also will pass a few other testcases
they just need cleaning up).

Thoughts on the Predicate* hiearchy of classes especially welcome :)
(it's not clear to me how best to organize it, and currently, the getBlock* seems ... uglier than maybe wasting a field somewhere or something).

Reviewers: davide

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D29747

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295889 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAdd pair conversion functions to BasicBlockEdge.
Daniel Berlin [Wed, 22 Feb 2017 22:20:53 +0000 (22:20 +0000)]
Add pair conversion functions to BasicBlockEdge.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295888 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMove updating functions to MemorySSAUpdater.
Daniel Berlin [Wed, 22 Feb 2017 22:19:55 +0000 (22:19 +0000)]
Move updating functions to MemorySSAUpdater.
Add updater to passes that now need it.
Move around code in MemorySSA to expose needed functions.

Summary: Mostly cleanup

Reviewers: george.burgess.iv

Subscribers: llvm-commits, Prazek

Differential Revision: https://reviews.llvm.org/D30221

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295887 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LV] Update floating-point induction test checks (NFC)
Matthew Simpson [Wed, 22 Feb 2017 21:56:02 +0000 (21:56 +0000)]
[LV] Update floating-point induction test checks (NFC)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295885 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LSR] Canonicalize formula and put recursive Reg related with current loop in ScaledReg.
Wei Mi [Wed, 22 Feb 2017 21:47:08 +0000 (21:47 +0000)]
[LSR] Canonicalize formula and put recursive Reg related with current loop in ScaledReg.

After rL294814, LSR formula can have multiple SCEVAddRecExprs inside of its BaseRegs.
Previous canonicalization will swap the first SCEVAddRecExpr in BaseRegs with ScaledReg.
But now we want to swap the SCEVAddRecExpr Reg related with current loop with ScaledReg.
Otherwise, we may generate code like this: RegA + lsr.iv + RegB, where loop invariant
parts RegA and RegB are not grouped together and cannot be promoted outside of loop.
With this patch, it will ensure lsr.iv to be generated later in the expr:
RegA + RegB + lsr.iv, so that RegA + RegB can be promoted outside of loop.

Differential Revision: https://reviews.llvm.org/D26781

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295884 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[RDF] Support for partial structural aliases in RegisterAggr
Krzysztof Parzyszek [Wed, 22 Feb 2017 21:42:15 +0000 (21:42 +0000)]
[RDF] Support for partial structural aliases in RegisterAggr

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295883 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Support] Re-add the special OSX flags on mmap.
Zachary Turner [Wed, 22 Feb 2017 21:24:06 +0000 (21:24 +0000)]
[Support] Re-add the special OSX flags on mmap.

The problem appears to be that these flags can only be used
when mapping a file for read-only, not for readwrite.  So
we do that here.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295880 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Hexagon] Add intrinsics for masked vector stores
Krzysztof Parzyszek [Wed, 22 Feb 2017 21:23:09 +0000 (21:23 +0000)]
[Hexagon] Add intrinsics for masked vector stores

Patch by Harsha Jagasia.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295879 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Don't look at chain users when adjusting writemask
Matt Arsenault [Wed, 22 Feb 2017 21:16:41 +0000 (21:16 +0000)]
AMDGPU: Don't look at chain users when adjusting writemask

Fixes not adjusting using new intrinsics with chains.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295878 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Always allocate emergency stack slot at offset 0
Matt Arsenault [Wed, 22 Feb 2017 21:05:25 +0000 (21:05 +0000)]
AMDGPU: Always allocate emergency stack slot at offset 0

This allows us to ensure that 0 is never a valid pointer
to a user object, and ensures that the offset is always legal
without needing a register to access it. This comes at the cost
of usable offsets and wasted stack space.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295877 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[WebAssembly] Update llvm-readobj tests for switch to version 0x1
Derek Schuff [Wed, 22 Feb 2017 21:01:17 +0000 (21:01 +0000)]
[WebAssembly] Update llvm-readobj tests for switch to version 0x1

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295875 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Change exp with compr bit printing
Matt Arsenault [Wed, 22 Feb 2017 20:37:12 +0000 (20:37 +0000)]
AMDGPU: Change exp with compr bit printing

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295873 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoRevert "AMDGPU : Update TrapCode based on Trap Handler ABI."
Wei Ding [Wed, 22 Feb 2017 20:29:22 +0000 (20:29 +0000)]
Revert "AMDGPU : Update TrapCode based on Trap Handler ABI."

This reverts commit r295867.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295871 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[WebAssembly] Update llvm-objdump tests for the new wasm version number.
Dan Gohman [Wed, 22 Feb 2017 20:24:16 +0000 (20:24 +0000)]
[WebAssembly] Update llvm-objdump tests for the new wasm version number.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295869 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong result
Alexey Bataev [Wed, 22 Feb 2017 20:06:40 +0000 (20:06 +0000)]
[SLP] Fix for PR32036: Vectorized horizontal reduction returning wrong result

Summary:
If the same value is used several times as an extra value, SLP
vectorizer takes it into account only once instead of actual number of
using.
For example:
```
int val = 1;
for (int y = 0; y < 8; y++) {
  for (int x = 0; x < 8; x++) {
    val = val + input[y * 8 + x] + 3;
  }
}
```
We have 2 extra rguments: `1` - initial value of horizontal reduction
and `3`, which is added 8*8 times to the reduction. Before the patch we
added `1` to the reduction value and added once `3`, though it must be
added 64 times.

Reviewers: mkuper, mzolotukhin

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D30262

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295868 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU : Update TrapCode based on Trap Handler ABI.
Wei Ding [Wed, 22 Feb 2017 20:05:06 +0000 (20:05 +0000)]
AMDGPU : Update TrapCode based on Trap Handler ABI.

Differential Revision: http://reviews.llvm.org/D30232

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295867 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMove llvm_unreachable out of switch.
Rafael Espindola [Wed, 22 Feb 2017 19:42:14 +0000 (19:42 +0000)]
Move llvm_unreachable out of switch.

This should make gcc happy and still produce a clang warning if we add
another value to the enum.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295865 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoBring back 2>&1 redirection for this test
Matthias Braun [Wed, 22 Feb 2017 19:16:33 +0000 (19:16 +0000)]
Bring back 2>&1 redirection for this test

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295864 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AArch64] Extend AArch64RedundantCopyElimination to do simple copy propagation.
Geoff Berry [Wed, 22 Feb 2017 19:10:45 +0000 (19:10 +0000)]
[AArch64] Extend AArch64RedundantCopyElimination to do simple copy propagation.

Summary:
Extend AArch64RedundantCopyElimination to catch cases where the register
that is known to be zero is COPY'd in the predecessor block.  Before
this change, this pass would catch cases like:

      CBZW %W0, <BB#1>
  BB#1:
      %W0 = COPY %WZR // removed

After this change, cases like the one below are also caught:

      %W0 = COPY %W1
      CBZW %W1, <BB#1>
  BB#1:
      %W0 = COPY %WZR // removed

This change results in a 4% increase in static copies removed by this
pass when compiling the llvm test-suite.  It also fixes regressions
caused by doing post-RA copy propagation (a separate change to be put up
for review shortly).

Reviewers: junbuml, mcrosier, t.p.northover, qcolombet, MatzeB

Subscribers: aemerson, rengolin, llvm-commits

Differential Revision: https://reviews.llvm.org/D30113

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295863 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LV] Add scalar floating-point induction test (NFC)
Matthew Simpson [Wed, 22 Feb 2017 19:09:38 +0000 (19:09 +0000)]
[LV] Add scalar floating-point induction test (NFC)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295862 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ModuleSummaryAnalysis] Don't crash when referencing unnamed globals.
Davide Italiano [Wed, 22 Feb 2017 18:53:38 +0000 (18:53 +0000)]
[ModuleSummaryAnalysis] Don't crash when referencing unnamed globals.

Instead, just be conservative as these are unfrequent enough. Thanks
to Peter Collingbourne for the discussion about this on IRC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295861 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[WebAssembly] Implement the wasm binary container header.
Dan Gohman [Wed, 22 Feb 2017 18:50:20 +0000 (18:50 +0000)]
[WebAssembly] Implement the wasm binary container header.

Also, update the version number to 0x1, which is what engines are now
expecting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295860 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoMIRTests: Remove unnecessary 2>&1 redirection
Matthias Braun [Wed, 22 Feb 2017 18:47:41 +0000 (18:47 +0000)]
MIRTests: Remove unnecessary 2>&1 redirection

llc mir output goes to stdout nowadays, so the 2>&1 is not necessary
anymore for most tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295859 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LoopVectorize] Added address space check when analysing interleaved accesses
Karl-Johan Karlsson [Wed, 22 Feb 2017 18:37:36 +0000 (18:37 +0000)]
[LoopVectorize] Added address space check when analysing interleaved accesses

Prevent memory objects of different address spaces to be part of
the same load/store groups when analysing interleaved accesses.

This is fixing pr31900.

Reviewers: HaoLiu, mssimpso, mkuper

Reviewed By: mssimpso, mkuper

Subscribers: llvm-commits, efriedma, mzolotukhin

Differential Revision: https://reviews.llvm.org/D29717

This reverts r295042 (re-applies r295038) with an additional fix for the
buildbot problem.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295858 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[WebAssembly] Define a table of function signatures for runtime library calls.
Dan Gohman [Wed, 22 Feb 2017 18:34:16 +0000 (18:34 +0000)]
[WebAssembly] Define a table of function signatures for runtime library calls.

LLVM CodeGen emits references to external symbols that are never declared in
LLVM IR level, so they have no declared signature. However, WebAssembly requires
all functions be declared with signatures. This patch adds a table for providing
signatures for known runtime libcalls that will be used in subsequent patches to
emit declarations for such functions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295857 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[RDF] Skip undef uses when calculating kill flags
Krzysztof Parzyszek [Wed, 22 Feb 2017 18:29:16 +0000 (18:29 +0000)]
[RDF] Skip undef uses when calculating kill flags

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295856 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[RDF] Only access block live-ins when tracking liveness
Krzysztof Parzyszek [Wed, 22 Feb 2017 18:27:36 +0000 (18:27 +0000)]
[RDF] Only access block live-ins when tracking liveness

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295855 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Support] Provide linux/magic.h fallback for older kernels
Michal Gorny [Wed, 22 Feb 2017 18:09:15 +0000 (18:09 +0000)]
[Support] Provide linux/magic.h fallback for older kernels

The function for distinguishing local and remote files added in r295768
unconditionally uses linux/magic.h header to provide necessary
filesystem magic numbers. However, in kernel headers predating 2.6.18
the magic numbers are spread throughout multiple include files.
Furthermore, LLVM did not require kernel headers being installed so far.

To increase the portability across different versions of Linux kernel
and different Linux systems, add CMake header checks for linux/magic.h
and -- if it is missing -- the linux/nfs_fs.h and linux/smb.h headers
which contained the numbers previously.

Furthermore, since the numbers are static and the feature does not seem
critical enough to make LLVM require kernel headers at all, add fallback
constants for the case when none of the necessary headers is available.

Differential Revision: https://reviews.llvm.org/D30261

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295854 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SLP] Test with extra argument used several times.
Alexey Bataev [Wed, 22 Feb 2017 17:47:28 +0000 (17:47 +0000)]
[SLP] Test with extra argument used several times.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295853 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix an obvious bug in SampleProfileReaderGCC.
Dehao Chen [Wed, 22 Feb 2017 17:27:21 +0000 (17:27 +0000)]
Fix an obvious bug in SampleProfileReaderGCC.

Summary: The CallTargetProfile should be added to FProfile to be consistent with other profile readers.

Reviewers: dnovillo, davidxl

Reviewed By: davidxl

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D30233

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295852 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[WebAssembly] Configure codegen to legalize f16 values.
Dan Gohman [Wed, 22 Feb 2017 16:28:00 +0000 (16:28 +0000)]
[WebAssembly] Configure codegen to legalize f16 values.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295850 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[DAGCombiner] revert r295336
Bill Seurer [Wed, 22 Feb 2017 16:27:33 +0000 (16:27 +0000)]
[DAGCombiner] revert r295336

r295336 causes a bootstrapped clang to fail for many compilations on
powerpc BE.  See
http://lab.llvm.org:8011/builders/clang-ppc64be-linux-multistage/builds/2315
for example.

Reverting as per the developer's request.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295849 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][SSE] getTargetConstantBitsFromNode - insert constant bits directly into masks.
Simon Pilgrim [Wed, 22 Feb 2017 15:38:13 +0000 (15:38 +0000)]
[X86][SSE] getTargetConstantBitsFromNode - insert constant bits directly into masks.

Minor optimization, don't create temporary mask APInts that are just going to be OR'd into the accumulate masks - insert directly instead.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295848 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][SSE] Use APInt::getBitsSet() instead of APInt::getLowBitsSet().shl() separately...
Simon Pilgrim [Wed, 22 Feb 2017 15:04:55 +0000 (15:04 +0000)]
[X86][SSE] Use APInt::getBitsSet() instead of APInt::getLowBitsSet().shl() separately. NFCI.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295845 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[Support] XFAIL is_local for mips
Simon Dardis [Wed, 22 Feb 2017 14:34:45 +0000 (14:34 +0000)]
[Support] XFAIL is_local for mips

is_local can't pass on some our buildbots as some of our buildbots use network
shares for building and testing LLVM.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295840 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago* [AMDGPU][mc][tests] Updated coverage/smoke tests for gfx7 and gfx8; minor test...
Dmitry Preobrazhensky [Wed, 22 Feb 2017 13:59:39 +0000 (13:59 +0000)]
* [AMDGPU][mc][tests] Updated coverage/smoke tests for gfx7 and gfx8; minor test corrections.

NB: several old tests have been corrected because they violated constant bus limitations

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295834 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix -Wunused-but-set-variable warning by removing unused 'aggregateIsPacked' checking
Simon Pilgrim [Wed, 22 Feb 2017 13:37:31 +0000 (13:37 +0000)]
Fix -Wunused-but-set-variable warning by removing unused 'aggregateIsPacked' checking

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295830 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[GlobalISel] Fix compiler warnings and make assert assert something.
Benjamin Kramer [Wed, 22 Feb 2017 12:59:47 +0000 (12:59 +0000)]
[GlobalISel] Fix compiler warnings and make assert assert something.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295827 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[SLP] Remove unused initial value from the variable, NFC.
Alexey Bataev [Wed, 22 Feb 2017 12:57:58 +0000 (12:57 +0000)]
[SLP] Remove unused initial value from the variable, NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295826 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86][GlobalISel] Initial implementation , select G_ADD gpr, gpr
Igor Breger [Wed, 22 Feb 2017 12:25:09 +0000 (12:25 +0000)]
[X86][GlobalISel] Initial implementation , select G_ADD gpr, gpr

Summary: Initial implementation for X86InstructionSelector. Handle selection COPY and G_ADD/G_SUB gpr, gpr .

Reviewers: qcolombet, rovka, zvi, ab

Reviewed By: rovka

Subscribers: mgorny, dberris, kristof.beyls, llvm-commits

Differential Revision: https://reviews.llvm.org/D29816

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295824 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86] Regenerate CSE test with codegen instead of just the instruction count
Simon Pilgrim [Wed, 22 Feb 2017 10:12:46 +0000 (10:12 +0000)]
[X86] Regenerate CSE test with codegen instead of just the instruction count

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295819 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] Fix constant islands pass.
Roger Ferrer Ibanez [Wed, 22 Feb 2017 09:06:21 +0000 (09:06 +0000)]
[ARM] Fix constant islands pass.

The pass tries to fix a spill of LR that turns out to be unnecessary.
So it removes the tPOP but forgets to remove tPUSH.
This causes the stack be misaligned upon returning the function.

Thus, remove the tPUSH as well in this case.

Differential Revision: https://reviews.llvm.org/D30207

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295816 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoWrite to a temporary file in test instead of random file in the test directory.
Benjamin Kramer [Wed, 22 Feb 2017 09:02:27 +0000 (09:02 +0000)]
Write to a temporary file in test instead of random file in the test directory.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295815 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[X86] Fix memory operands definition for some instructions.
Ayman Musa [Wed, 22 Feb 2017 08:06:29 +0000 (08:06 +0000)]
[X86] Fix memory operands definition for some instructions.

Change integer memory operands to FP memory operands to some FP instructions.

Differential Revision: https://reviews.llvm.org/D30201

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295813 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoOptDiag: Add const to some interfaces that don't modify anything. NFC
Justin Bogner [Wed, 22 Feb 2017 07:38:17 +0000 (07:38 +0000)]
OptDiag: Add const to some interfaces that don't modify anything. NFC

This needed a const_cast for the dominator tree recalculation in
OptimizationRemarkEmitter, but we do that all over the place already
and it's safe.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295812 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ARM] Classification Improvements to ARM Sched-Models. NFCI.
Javed Absar [Wed, 22 Feb 2017 07:22:57 +0000 (07:22 +0000)]
[ARM] Classification Improvements to ARM Sched-Models. NFCI.

This patch adds missing sched classes for Thumb2 instructions.
This has been missing so far, and as a consequence, machine
scheduler models for individual sub-targets have tended to
be larger than they needed to be. These patches should help
write schedulers better and faster in the future
for ARM sub-targets.

Reviewer: Diana Picus
Differential Revision: https://reviews.llvm.org/D29953

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295811 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[AVX-512] Allow legacy scalar min/max intrinsics to select EVEX instructions when...
Craig Topper [Wed, 22 Feb 2017 06:54:18 +0000 (06:54 +0000)]
[AVX-512] Allow legacy scalar min/max intrinsics to select EVEX instructions when available

This patch introduces new X86ISD::FMAXS and X86ISD::FMINS opcodes. The legacy intrinsics now lower to this node. As do the AVX-512 masked intrinsics when the rounding mode is CUR_DIRECTION.

I've merged a copy of the tablegen multiclass avx512_fp_scalar into avx512_fp_scalar_sae. avx512_fp_scalar still needs to support CUR_DIRECTION appearing as a rounding mode for X86ISD::FADD_ROUND and others.

Differential revision: https://reviews.llvm.org/D30186

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295810 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[ValueTracking] Make poison propagation more aggressive
Sanjoy Das [Wed, 22 Feb 2017 06:52:32 +0000 (06:52 +0000)]
[ValueTracking] Make poison propagation more aggressive

Summary:
Motivation: fix PR31181 without regression (the actual fix is still in
progress).  However, the actual content of PR31181 is not relevant
here.

This change makes poison propagation more aggressive in the following
cases:

 1. poision * Val == poison, for any Val.  In particular, this changes
    existing intentional and documented behavior in these two cases:
     a. Val is 0
     b. Val is 2^k * N
 2. poison << Val == poison, for any Val
 3. getelementptr is poison if any input is poison

I think all of these are justified (and are axiomatically true in the
new poison / undef model):

1a: we need poison * 0 to be poison to allow transforms like these:

  A * (B + C) ==> A * B + A * C

If poison * 0 were 0 then the above transform could not be allowed
since e.g. we could have A = poison, B = 1, C = -1, making the LHS

  poison * (1 + -1) = poison * 0 = 0

and the RHS

  poison * 1 + poison * -1 = poison + poison = poison

1b: we need e.g. poison * 4 to be poison since we want to allow

  A * 4 ==> A + A + A + A

If poison * 4 were a value with all of their bits poison except the
last four; then we'd not be able to do this transform since then if A
were poison the LHS would only be "partially" poison while the RHS
would be "full" poison.

2: Same reasoning as (1b), we'd like have the following kinds
transforms be legal:

  A << 1 ==> A + A

Reviewers: majnemer, efriedma

Subscribers: mcrosier, llvm-commits

Differential Revision: https://reviews.llvm.org/D30185

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295809 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoUse const-ref in range-loop for to avoid copying pairs of std::string
Sean Silva [Wed, 22 Feb 2017 06:34:04 +0000 (06:34 +0000)]
Use const-ref in range-loop for to avoid copying pairs of std::string

No reason to create temporaries.

Differential Revision: https://reviews.llvm.org/D29871

Patch by sergio.martins!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295807 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[WebAssembly] Add skeleton MC support for the Wasm container format
Dan Gohman [Wed, 22 Feb 2017 01:23:18 +0000 (01:23 +0000)]
[WebAssembly] Add skeleton MC support for the Wasm container format

This just adds the basic skeleton for supporting a new object file format.
All of the actual encoding will be implemented in followup patches.

Differential Revision: https://reviews.llvm.org/D26722

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295803 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoFix -Wcovered-switch-default.
Rui Ueyama [Wed, 22 Feb 2017 01:01:45 +0000 (01:01 +0000)]
Fix -Wcovered-switch-default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295799 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Add cvt.pkrtz intrinsic
Matt Arsenault [Wed, 22 Feb 2017 00:27:34 +0000 (00:27 +0000)]
AMDGPU: Add cvt.pkrtz intrinsic

Convert llvm.SI.packf16 test uses

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295797 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[LoopUnroll] Enable PGO-based loop peeling by default.
Michael Kuperstein [Wed, 22 Feb 2017 00:27:34 +0000 (00:27 +0000)]
[LoopUnroll] Enable PGO-based loop peeling by default.

This enables peeling of loops with low dynamic iteration count by default,
when profile information is available.

Differential Revision: https://reviews.llvm.org/D27734

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295796 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Remove some uses of llvm.SI.export in tests
Matt Arsenault [Wed, 22 Feb 2017 00:02:21 +0000 (00:02 +0000)]
AMDGPU: Remove some uses of llvm.SI.export in tests

Merge some of the old, smaller tests into more complete versions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295792 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Remove llvm.AMDGPU.clamp intrinsic
Matt Arsenault [Tue, 21 Feb 2017 23:46:04 +0000 (23:46 +0000)]
AMDGPU: Remove llvm.AMDGPU.clamp intrinsic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295789 91177308-0d34-0410-b5e6-96231b3b80d8

7 years agoAMDGPU: Redefine clamp node as clamp 0.0-1.0
Matt Arsenault [Tue, 21 Feb 2017 23:35:48 +0000 (23:35 +0000)]
AMDGPU: Redefine clamp node as clamp 0.0-1.0

Change implementation to use max instead of add.
min/max/med3 do not flush denormals regardless of the mode,
so it is OK to use it whether or not they are enabled.

Also allow using clamp with f16, and use knowledge
of dx10_clamp.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295788 91177308-0d34-0410-b5e6-96231b3b80d8

7 years ago[NVPTX] Unify vectorization of load/stores of aggregate arguments and return values.
Artem Belevich [Tue, 21 Feb 2017 22:56:05 +0000 (22:56 +0000)]
[NVPTX] Unify vectorization of load/stores of aggregate arguments and return values.

Original code only used vector loads/stores for explicit vector arguments.
It could also do more loads/stores than necessary (e.g v5f32 would
touch 8 f32 values). Aggregate types were loaded one element at a time,
even the vectors contained within.

This change attempts to generalize (and simplify) parameter space
loads/stores so that vector loads/stores can be used more broadly.
Functionality of the patch has been verified by compiling thrust
test suite and manually checking the differences between PTX
generated by llvm with and without the patch.

General algorithm:
* ComputePTXValueVTs() flattens input/output argument into a flat list
  of scalars to load/store and returns their types and offsets.
* VectorizePTXValueVTs() uses that data to create vectorization plan
  which returns an array of flags marking boundaries of vectorized
  load/stores. Scalars are represented as 1-element vectors.
* Code that generates loads/stores implements a simple state machine
  that constructs a vector according to the plan.

Differential Revision: https://reviews.llvm.org/D30011

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@295784 91177308-0d34-0410-b5e6-96231b3b80d8