oota-llvm.git
6 years agoWhitespace.
NAKAMURA Takumi [Tue, 27 Oct 2015 19:02:52 +0000 (19:02 +0000)]
Whitespace.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251437 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRevert r251291, "Loop Vectorizer - skipping "bitcast" before GEP"
NAKAMURA Takumi [Tue, 27 Oct 2015 19:02:36 +0000 (19:02 +0000)]
Revert r251291, "Loop Vectorizer - skipping "bitcast" before GEP"

It causes miscompilation of llvm/lib/ExecutionEngine/Interpreter/Execution.cpp.
See also PR25324.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251436 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoTidy a comment. NFC.
Diego Novillo [Tue, 27 Oct 2015 18:41:46 +0000 (18:41 +0000)]
Tidy a comment. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251434 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoCreate a new interface addSuccessorWithoutWeight(MBB*) in MBB to add successors when...
Cong Hou [Tue, 27 Oct 2015 17:59:36 +0000 (17:59 +0000)]
Create a new interface addSuccessorWithoutWeight(MBB*) in MBB to add successors when optimization is disabled.

When optimization is disabled, edge weights that are stored in MBB won't be used so that we don't have to store them. Currently, this is done by adding successors with default weight 0, and if all successors have default weights, the weight list will be empty. But that the weight list is empty doesn't mean disabled optimization (as is stated several times in MachineBasicBlock.cpp): it may also mean all successors just have default weights.

We should discourage using default weights when adding successors, because it is very easy for users to forget update the correct edge weights instead of using default ones (one exception is that the MBB only has one successor). In order to detect such usages, it is better to differentiate using default weights from the case when optimizations is disabled.

In this patch, a new interface addSuccessorWithoutWeight(MBB*) is created for when optimization is disabled. In this case, MBB will try to maintain an empty weight list, but it cannot guarantee this as for many uses of addSuccessor() whether optimization is disabled or not is not checked. But it can guarantee that if optimization is enabled, then the weight list always has the same size of the successor list.

Differential revision: http://reviews.llvm.org/D13963

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251429 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SLP] Be more aggressive about reduction width selection.
Charlie Turner [Tue, 27 Oct 2015 17:59:03 +0000 (17:59 +0000)]
[SLP] Be more aggressive about reduction width selection.

Summary:
This change could be way off-piste, I'm looking for any feedback on whether it's an acceptable approach.

It never seems to be a problem to gobble up as many reduction values as can be found, and then to attempt to reduce the resulting tree. Some of the workloads I'm looking at have been aggressively unrolled by hand, and by selecting reduction widths that are not constrained by a vector register size, it becomes possible to profitably vectorize. My test case shows such an unrolling which SLP was not vectorizing (on neither ARM nor X86) before this patch, but with it does vectorize.

I measure no significant compile time impact of this change when combined with D13949 and D14063. There are also no significant performance regressions on ARM/AArch64 in SPEC or LNT.

The more principled approach I thought of was to generate several candidate tree's and use the cost model to pick the cheapest one. That seemed like quite a big design change (the algorithms seem very much one-shot), and would likely be a costly thing for compile time. This seemed to do the job at very little cost, but I'm worried I've misunderstood something!

Reviewers: nadav, jmolloy

Subscribers: mssimpso, llvm-commits, aemerson

Differential Revision: http://reviews.llvm.org/D14116

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251428 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SLP] Try a bit harder to find reduction PHIs
Charlie Turner [Tue, 27 Oct 2015 17:54:16 +0000 (17:54 +0000)]
[SLP] Try a bit harder to find reduction PHIs

Summary:
Currently, when the SLP vectorizer considers whether a phi is part of a reduction, it dismisses phi's whose incoming blocks are not the same as the block containing the phi. For the patterns I'm looking at, extending this rule to allow phis whose incoming block is a containing loop latch allows me to vectorize certain workloads.

There is no significant compile-time impact, and combined with D13949, no performance improvement measured in ARM/AArch64 in any of SPEC2000, SPEC2006 or LNT.

Reviewers: jmolloy, mcrosier, nadav

Subscribers: mssimpso, nadav, aemerson, llvm-commits

Differential Revision: http://reviews.llvm.org/D14063

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251425 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SLP] Treat SelectInsts as reduction values.
Charlie Turner [Tue, 27 Oct 2015 17:49:11 +0000 (17:49 +0000)]
[SLP] Treat SelectInsts as reduction values.

Summary:
Certain workloads, in particular sum-of-absdiff loops, can be vectorized using SLP if it can treat select instructions as reduction values.

The test case is a bit awkward. The AArch64 cost model needs some tuning to not be so pessimistic about selects. I've had to tweak the SLP threshold here.

Reviewers: jmolloy, mzolotukhin, spatel, nadav

Subscribers: nadav, mssimpso, aemerson, llvm-commits

Differential Revision: http://reviews.llvm.org/D13949

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251424 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Orc] Fix indentation.
Lang Hames [Tue, 27 Oct 2015 17:45:48 +0000 (17:45 +0000)]
[Orc] Fix indentation.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251423 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix SamplePGO segfault when debug info is missing.
Diego Novillo [Tue, 27 Oct 2015 17:37:00 +0000 (17:37 +0000)]
Fix SamplePGO segfault when debug info is missing.

When emitting a remark for a conditional branch annotation, the remark
uses the line location information of the conditional branch in the
message.  In some cases, that information is unavailable and the
optimization would segfaul. I'm still not sure whether this is a bug or
WAI, but the optimizer should not die because of this.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251420 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ms-inline-asm] Leave alignment in bytes if the native assembler uses bytes
Reid Kleckner [Tue, 27 Oct 2015 17:32:48 +0000 (17:32 +0000)]
[ms-inline-asm] Leave alignment in bytes if the native assembler uses bytes

The existing behavior was correct on Darwin, which is probably the
platform it was written for.

Before this change, we would rewrite "align 8" to ".align 3" and then
fail to make it through the integrated assembler because 3 is not a
power of 2.

Differential Revision: http://reviews.llvm.org/D14120

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251418 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRename qsort -> multikey_qsort. NFC.
Rui Ueyama [Tue, 27 Oct 2015 16:57:50 +0000 (16:57 +0000)]
Rename qsort -> multikey_qsort. NFC.

`qsort` as a file-scope local function name was confusing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251414 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoPrefer ranlib mode over ar mode.
Ed Schouten [Tue, 27 Oct 2015 16:37:49 +0000 (16:37 +0000)]
Prefer ranlib mode over ar mode.

For CloudABI's toolchain I have a symlink that goes from <target>-ar and
<target>-ranlib to LLVM's ar binary, to mimick GNU Binutils' naming
scheme. The problem is that if we're targetting ARM64, the name of the
ranlib executable is aarch64-unknown-cloudabi-ranlib. This already
contains the string "ar".

Let's move the "ranlib" test above the "ar" test. It's not that likely
that we're going to see operating systems or harwdare architectures that
are called "ranlib".

Reviewed by: rafael
Differential Revision: http://reviews.llvm.org/D14123

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251413 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[CMake] Get rid of LLVM_DYLIB_EXPORT_ALL, and make it the default, add libLLVM-C...
Chris Bieneman [Tue, 27 Oct 2015 16:02:04 +0000 (16:02 +0000)]
[CMake] Get rid of LLVM_DYLIB_EXPORT_ALL, and make it the default, add libLLVM-C on darwin to cover the C API needs.

Summary:
We've had a lot of discussion in the past about the meaningful and useful default behaviors for the llvm-shlib tool. The original implementation was heavily geared toward Apple's use, and I think that was wrong. This patch seeks to correct that.

I've removed the LLVM_DYLIB_EXPORT_ALL variable and made libLLVM export everything by default.

I've also added a new target that is only built on Darwin for libLLVM-C as a library that re-exports the LLVM-C API. This library is not built on Linux because ELF doesn't support re-export libraries in the same way MachO does.

Reviewers: chapuni, resistor, bogner, axw

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D13842

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251411 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][AVX512] [X86][AVX512] add convert float to half
Asaf Badouh [Tue, 27 Oct 2015 15:37:17 +0000 (15:37 +0000)]
[X86][AVX512] [X86][AVX512] add convert float to half

convert float to half with mask/maskz for the reg to reg version and mask for the reg to mem version (there is no maskz version for reg to mem).

Differential Revision: http://reviews.llvm.org/D14113

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251409 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] Expand ROTL and ROTR of vector value types
Charlie Turner [Tue, 27 Oct 2015 10:25:20 +0000 (10:25 +0000)]
[ARM] Expand ROTL and ROTR of vector value types

Summary: After D13851 landed, we saw backend crashes when compiling the reduced test case included in this patch. The right fix seems to be to allow these vector types for expansion in instruction selection.

Reviewers: rengolin, t.p.northover

Subscribers: RKSimon, t.p.northover, aemerson, llvm-commits, rengolin

Differential Revision: http://reviews.llvm.org/D14082

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251401 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoDo not use "else" when both branches return (NFC)
Mehdi Amini [Tue, 27 Oct 2015 08:12:08 +0000 (08:12 +0000)]
Do not use "else" when both branches return (NFC)

From: Mehdi Amini <mehdi.amini@apple.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251398 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ScalarEvolutionExpander] Properly insert no-op casts + EH Pads
David Majnemer [Tue, 27 Oct 2015 07:36:42 +0000 (07:36 +0000)]
[ScalarEvolutionExpander] Properly insert no-op casts + EH Pads

We want to insert no-op casts as close as possible to the def.  This is
tricky when the cast is of a PHI node and the BasicBlocks between the
def and the use cannot hold any instructions.  Iteratively walk EH pads
until we hit a non-EH pad.

This fixes PR25326.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251393 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Make elfiamcu an OS, not an environment.
Michael Kuperstein [Tue, 27 Oct 2015 07:23:59 +0000 (07:23 +0000)]
[X86] Make elfiamcu an OS, not an environment.

GNU tools require elfiamcu to take up the entire OS field, so, e.g.
i?86-*-linux-elfiamcu is not considered a legal triple.
Make us compatible.

Differential Revision: http://reviews.llvm.org/D14081

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251390 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SimplifyLibCalls] Use range-based loop. No functional change.
Davide Italiano [Tue, 27 Oct 2015 04:17:51 +0000 (04:17 +0000)]
[SimplifyLibCalls] Use range-based loop. No functional change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251383 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoConvert cost table lookup functions to return a pointer to the entry or nullptr inste...
Craig Topper [Tue, 27 Oct 2015 04:14:24 +0000 (04:14 +0000)]
Convert cost table lookup functions to return a pointer to the entry or nullptr instead of the index.

This avoid mentioning the table name an extra time and allows the lookup to be done directly in the ifs by relying on the bool conversion of the pointer.

While there make use of ArrayRef and std::find_if.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251382 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[function-attrs] Refactor code to handle shorter code with early exits.
Chandler Carruth [Tue, 27 Oct 2015 01:41:43 +0000 (01:41 +0000)]
[function-attrs] Refactor code to handle shorter code with early exits.

No functionality changed here, but the indentation is substantially
reduced and IMO the code is much easier to read. I've also added some
helpful comments.

This is just a clean-up I wrote while studying the code, and that has
been in my backlog for a while.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251381 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ValueTracking] Don't special case wrapped ConstantRanges; NFCI
Sanjoy Das [Tue, 27 Oct 2015 01:36:06 +0000 (01:36 +0000)]
[ValueTracking] Don't special case wrapped ConstantRanges; NFCI

Use `getUnsignedMax` directly instead of special casing a wrapped
ConstantRange.

The previous code would have been "buggy" (and this would have been a
semantic change) if LLVM allowed !range metadata to denote full
ranges. E.g. in

  %val = load i1, i1* %ptr, !range !{i1 1, i1 1} ;; == full set

ValueTracking would conclude that the high bit (IOW the only bit) in
%val was zero.

Since !range metadata does not allow empty or full ranges, this change
is just a minor stylistic improvement.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251380 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[x86] replace integer logic ops with packed SSE FP logic ops
Sanjay Patel [Tue, 27 Oct 2015 01:28:07 +0000 (01:28 +0000)]
[x86] replace integer logic ops with packed SSE FP logic ops

If we have an operand to a bitwise logic op that's already in
an XMM register and the result is going to be sent to an XMM
register, then use an SSE logic op to avoid moves between the
integer and vector register files.

Related commits:
http://reviews.llvm.org/rL248395
http://reviews.llvm.org/rL248399
http://reviews.llvm.org/rL248404
http://reviews.llvm.org/rL248409
http://reviews.llvm.org/rL248415

This should solve PR22428:
https://llvm.org/bugs/show_bug.cgi?id=22428

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251378 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SCEV] Refactor out ScalarEvolution::getDataLayout; NFC
Sanjoy Das [Tue, 27 Oct 2015 00:52:09 +0000 (00:52 +0000)]
[SCEV] Refactor out ScalarEvolution::getDataLayout; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251375 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix llc crash processing S/UREM for -Oz builds caused by rL250825.
Steve King [Tue, 27 Oct 2015 00:14:06 +0000 (00:14 +0000)]
Fix llc crash processing S/UREM for -Oz builds caused by rL250825.

When taking the remainder of a value divided by a constant, visitREM()
attempts to convert the REM to a longer but faster sequence of instructions.
This conversion calls combine() on a speculative DIV instruction. Commit
rL250825 may cause this combine() to return a DIVREM, corrupting nearby nodes.
Flow eventually hits unreachable().

This patch adds a test case and a check to prevent visitREM() from trying
to convert the REM instruction in cases where a DIVREM is possible.
See http://reviews.llvm.org/D14035

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251373 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoadd FP logic test cases to show current codegen (PR22428)
Sanjay Patel [Mon, 26 Oct 2015 23:52:42 +0000 (23:52 +0000)]
add FP logic test cases to show current codegen (PR22428)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251370 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips][ias] Fold needsExpansion() and expandInstruction() together. NFC.
Daniel Sanders [Mon, 26 Oct 2015 23:50:00 +0000 (23:50 +0000)]
[mips][ias] Fold needsExpansion() and expandInstruction() together. NFC.

Summary:
Previously we maintained two separate switch statements that had to be kept in
sync. This patch merges them into a single switch.

Reviewers: vkalintiris

Subscribers: llvm-commits, dsanders

Differential Revision: http://reviews.llvm.org/D14012

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251369 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoSwitch ownership of miscellaneous ARM target to myself.
Tim Northover [Mon, 26 Oct 2015 23:33:13 +0000 (23:33 +0000)]
Switch ownership of miscellaneous ARM target to myself.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251367 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[x86] Make the vselect-minmax test 2x to 3x faster by deleting all the
Chandler Carruth [Mon, 26 Oct 2015 22:54:53 +0000 (22:54 +0000)]
[x86] Make the vselect-minmax test 2x to 3x faster by deleting all the
instructions that aren't relevant for instruction selection of vector
min and max.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251366 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoUse Twin instead of std::to_string.
Oleksiy Vyalov [Mon, 26 Oct 2015 22:37:36 +0000 (22:37 +0000)]
Use Twin instead of std::to_string.

http://reviews.llvm.org/D14095

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251365 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix indents. It's a follow up to r251353.
Ivan Krasin [Mon, 26 Oct 2015 22:35:40 +0000 (22:35 +0000)]
Fix indents. It's a follow up to r251353.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251364 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LLVMSymbolize] Don't use LLVMSymbolizer::Options in ModuleInfo. NFC.
Alexey Samsonov [Mon, 26 Oct 2015 22:34:56 +0000 (22:34 +0000)]
[LLVMSymbolize] Don't use LLVMSymbolizer::Options in ModuleInfo. NFC.

LLVMSymbolizer::Options is mostly used in LLVMSymbolizer class anyway.
Let's keep their usage restricted to that class, especially given that
it's worth to move ModuleInfo to a different header, independent from
the symbolizer class.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251363 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoreorganize logic; NFCI (retry r251349)
Sanjay Patel [Mon, 26 Oct 2015 21:54:14 +0000 (21:54 +0000)]
reorganize logic; NFCI (retry r251349)

This is a preliminary step before adding another optimization
to PerformBITCASTCombine().

..and I really hope it's NFC this time!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251357 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoMove imported entities into DwarfCompilationUnit to speed up LTO linking.
Ivan Krasin [Mon, 26 Oct 2015 21:36:35 +0000 (21:36 +0000)]
Move imported entities into DwarfCompilationUnit to speed up LTO linking.

Summary:
In particular, this CL speeds up the official Chrome linking with LTO by
1.8x.

See more details in https://crbug.com/542426

Reviewers: dblaikie

Subscribers: jevinskie

Differential Revision: http://reviews.llvm.org/D13918

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251353 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoARM: make sure VFP loads and stores are properly aligned.
Tim Northover [Mon, 26 Oct 2015 21:32:53 +0000 (21:32 +0000)]
ARM: make sure VFP loads and stores are properly aligned.

Both VLDRS and VLDRD fault if the memory is not 4 byte aligned, which wasn't
really being checked before, leading to faults at runtime.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251352 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agorevert r251349; it included code for a functional change
Sanjay Patel [Mon, 26 Oct 2015 21:28:02 +0000 (21:28 +0000)]
revert r251349; it included code for a functional change

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251350 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoreorganize logic; NFCI
Sanjay Patel [Mon, 26 Oct 2015 21:24:09 +0000 (21:24 +0000)]
reorganize logic; NFCI

This is a preliminary step before adding another optimization
to PerformBITCASTCombine().

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251349 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoInitialize BasicAAWrapperPass in it's constructor
Keno Fischer [Mon, 26 Oct 2015 21:22:58 +0000 (21:22 +0000)]
Initialize BasicAAWrapperPass in it's constructor

Summary: This idiom is used elsewhere in LLVM, but was overlooked here.

Reviewers: chandlerc

Subscribers: llvm-commits

Differential Revision: http://reviews.llvm.org/D13628

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251348 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix build failure on GCC 4.7 (old libstdc++ doesn't have std::map::emplace).
Alexey Samsonov [Mon, 26 Oct 2015 21:20:37 +0000 (21:20 +0000)]
Fix build failure on GCC 4.7 (old libstdc++ doesn't have std::map::emplace).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251347 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRemove use of std::map<>::emplace which is not supported on some older versions of...
David Blaikie [Mon, 26 Oct 2015 21:10:36 +0000 (21:10 +0000)]
Remove use of std::map<>::emplace which is not supported on some older versions of libstdc++

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251346 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRemove unused local variable. NFC.
Diego Novillo [Mon, 26 Oct 2015 20:50:26 +0000 (20:50 +0000)]
Remove unused local variable. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251344 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix tests.
Peter Collingbourne [Mon, 26 Oct 2015 20:49:49 +0000 (20:49 +0000)]
Fix tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251343 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoARM/ELF: Restore original (pre-r251322) logic for deciding whether to use GOT.
Peter Collingbourne [Mon, 26 Oct 2015 20:46:44 +0000 (20:46 +0000)]
ARM/ELF: Restore original (pre-r251322) logic for deciding whether to use GOT.

Unbreaks linking with gold, which cannot resolve direct relocations referring
to global symbols.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251342 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LLVMSymbolize] Use symbol table only if function linkage name was requested.
Alexey Samsonov [Mon, 26 Oct 2015 20:12:29 +0000 (20:12 +0000)]
[LLVMSymbolize] Use symbol table only if function linkage name was requested.

Now it's enough to just specify -functions=short without additionally
providing -use-symbol-table=false.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251339 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoFix build error by fully qualifying llvm::make_unique.
Alexey Samsonov [Mon, 26 Oct 2015 20:12:27 +0000 (20:12 +0000)]
Fix build error by fully qualifying llvm::make_unique.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251338 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoOptimize StringTableBuilder.
Rui Ueyama [Mon, 26 Oct 2015 19:58:29 +0000 (19:58 +0000)]
Optimize StringTableBuilder.

This is a patch to improve StringTableBuilder's performance. That class'
finalize function is very hot particularly in LLD because the function
does tail-merge strings in string tables or SHF_MERGE sections.

Generic std::sort-style sorter is not efficient for sorting strings.
The function implemented in this patch seems to be more efficient.

Here's a benchmark of LLD to link Clang with or without this patch.
The numbers are medians of 50 runs.

-O0
real 0m0.455s
real 0m0.430s (5.5% faster)

-O3
real 0m0.487s
real 0m0.452s (7.2% faster)

Since that is a benchmark of the whole linker, the speedup of
StringTableBuilder itself is much more than that.

http://reviews.llvm.org/D14053

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251337 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LLVMSymbolize] Use std::unique_ptr more extensively to clarify ownership.
Alexey Samsonov [Mon, 26 Oct 2015 19:41:23 +0000 (19:41 +0000)]
[LLVMSymbolize] Use std::unique_ptr more extensively to clarify ownership.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251336 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[RS4GC] Strip noalias attribute after statepoint rewrite
Igor Laevsky [Mon, 26 Oct 2015 19:06:01 +0000 (19:06 +0000)]
[RS4GC] Strip noalias attribute after statepoint rewrite

We should remove noalias along with dereference and dereference_or_null attributes
because statepoint could potentially touch the entire heap including noalias objects.

Differential Revision: http://reviews.llvm.org/D14032

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251333 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoSamplePGO - Add optimization reports.
Diego Novillo [Mon, 26 Oct 2015 18:52:53 +0000 (18:52 +0000)]
SamplePGO - Add optimization reports.

This adds a couple of optimization remarks to the SamplePGO
transformation. When it decides to inline a hot function (to mimic the
inline stack and repeat useful inline decisions in the original build).

It will also report branch destinations. For instance, given the code
fragment:

     6      if (i < 1000)
     7        sum -= i;
     8      else
     9        sum += -i * rand();

If the 'else' branch is taken most of the time, building this code with
-Rpass=sample-profile will produce:

a.cc:9:14: remark: most popular destination for conditional branches at small.cc:6:9 [-Rpass=sample-profile]
      sum += -i * rand();
             ^

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251330 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRemove assert(false) in favor of asserting the if conditional it is contained within.
David Blaikie [Mon, 26 Oct 2015 18:41:13 +0000 (18:41 +0000)]
Remove assert(false) in favor of asserting the if conditional it is contained within.

Also adjust the code to avoid 3 redundant map lookups.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251327 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoMove the canonical header to the top of its matching cpp file as per coding convention
David Blaikie [Mon, 26 Oct 2015 18:40:56 +0000 (18:40 +0000)]
Move the canonical header to the top of its matching cpp file as per coding convention

This ensures that the header will be verified to be standalone (and
avoid mistakes like the one fixed in r251178)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251326 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAdd an (optional) identification block in the bitcode
Mehdi Amini [Mon, 26 Oct 2015 18:37:00 +0000 (18:37 +0000)]
Add an (optional) identification block in the bitcode

Processing bitcode from a different LLVM version can lead to
unexpected behavior. The LLVM project guarantees autoupdating
bitcode from a previous minor revision for the same major, but
can't make any promise when reading bitcode generated from a
either a non-released LLVM, a vendor toolchain, or a "future"
LLVM release. This patch aims at being more user-friendly and
allows a bitcode produce to emit an optional block at the
beginning of the bitcode that will contains an opaque string
intended to describe the bitcode producer information. The
bitcode reader will dump this information alongside any error it
reports.

The optional block also includes an "epoch" number, monotonically
increasing when incompatible changes are made to the bitcode. The
reader will reject bitcode whose epoch is different from the one
expected.

Differential Revision: http://reviews.llvm.org/D13666

From: Mehdi Amini <mehdi.amini@apple.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251325 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[safestack] Fast access to the unsafe stack pointer on AArch64/Android.
Evgeniy Stepanov [Mon, 26 Oct 2015 18:28:25 +0000 (18:28 +0000)]
[safestack] Fast access to the unsafe stack pointer on AArch64/Android.

Android libc provides a fixed TLS slot for the unsafe stack pointer,
and this change implements direct access to that slot on AArch64 via
__builtin_thread_pointer() + offset.

This change also moves more code into TargetLowering and its
target-specific subclasses to get rid of target-specific codegen
in SafeStackPass.

This change does not touch the ARM backend because ARM lowers
builting_thread_pointer as aeabi_read_tp, which is not available
on Android.

The previous iteration of this change was reverted in r250461. This
version leaves the generic, compiler-rt based implementation in
SafeStack.cpp instead of moving it to TargetLoweringBase in order to
allow testing without a TargetMachine.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251324 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoBitstreamWriter: Fix integer overflow.
Peter Collingbourne [Mon, 26 Oct 2015 18:23:21 +0000 (18:23 +0000)]
BitstreamWriter: Fix integer overflow.

We were previously overflowing a 32-bit multiply operation when emitting large
(>512MB) bitcode files, resulting in corrupted bitcode. Fix by extending
one of the operands to 64 bits.

There are a few other 32-bit integer types in this code that seem like they
also ought to be extended to 64 bits; this will be done separately.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251323 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoARM/ELF: Better codegen for global variable addresses.
Peter Collingbourne [Mon, 26 Oct 2015 18:23:16 +0000 (18:23 +0000)]
ARM/ELF: Better codegen for global variable addresses.

In PIC mode we were previously computing global variable addresses (or GOT
entry addresses) by adding the PC, the PC-relative GOT displacement and
the GOT-relative symbol/GOT entry displacement. Because the latter two
displacements are fixed, we ended up performing one more addition than
necessary.

This change causes us to compute addresses using a single PC-relative
displacement, resulting in a shorter code sequence. This reduces code size
by about 4% in a recent build of Chromium for Android.

As a result of this change we no longer need to compute the GOT base address
in the ARM backend, which allows us to remove the Global Base Reg pass and
SDAG lowering for the GOT.

We also now no longer use the GOT when addressing a symbol which is known
to be defined in the same linkage unit. Specifically, the symbol must have
either hidden visibility or a strong definition in the current module in
order to not use the the GOT.

This is a change from the previous behaviour where we would use the GOT to
address externally visible symbols defined in the same module. I think the
only cases where this could matter are cases involving symbol interposition,
but we don't really support that well anyway.

Differential Revision: http://reviews.llvm.org/D13650

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251322 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoCleanup test case debug info. NFC.
Diego Novillo [Mon, 26 Oct 2015 18:14:44 +0000 (18:14 +0000)]
Cleanup test case debug info. NFC.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251320 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRefactor: Simplify boolean conditional return statements in lib/Transforms/Instrument...
Alexey Samsonov [Mon, 26 Oct 2015 18:06:40 +0000 (18:06 +0000)]
Refactor: Simplify boolean conditional return statements in lib/Transforms/Instrumentation

Summary: Use clang-tidy to simplify boolean conditional return statements.

Differential Revision: http://reviews.llvm.org/D9996

Patch by Richard (legalize@xmission.com)!

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251318 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoCheck the case that the numerator and denominator are both zeros when getting edge...
Cong Hou [Mon, 26 Oct 2015 18:00:17 +0000 (18:00 +0000)]
Check the case that the numerator and denominator are both zeros when getting edge probabilities in BPI and return 100% in this case.

This issue is triggered in PGO mode when bootstrapping LLVM. It seems that it is not guaranteed that edge weights are always greater than zero which are read from profile data.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251317 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoMove parts of llvm-symbolizer tool into LLVMSymbolize library.
Alexey Samsonov [Mon, 26 Oct 2015 17:56:12 +0000 (17:56 +0000)]
Move parts of llvm-symbolizer tool into LLVMSymbolize library.

Summary: See http://lists.llvm.org/pipermail/llvm-dev/2015-October/091624.html

Reviewers: echristo

Subscribers: llvm-commits, aizatsky

Differential Revision: http://reviews.llvm.org/D13998

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251316 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SystemZ] LTGFR use regclass should be GR32, not GR64.
Jonas Paulsson [Mon, 26 Oct 2015 15:03:49 +0000 (15:03 +0000)]
[SystemZ] LTGFR use regclass should be GR32, not GR64.

Discovered by testing int-cmp-44.ll with -verify-machineinstrs (added to
test run).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251299 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SystemZ] Also clear kill flag for index reg in splitMove().
Jonas Paulsson [Mon, 26 Oct 2015 15:03:41 +0000 (15:03 +0000)]
[SystemZ] Also clear kill flag for index reg in splitMove().

Discovered by running fp-move-05.ll with -verify-machineinstrs (added
to test case run).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251298 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SystemZ] Don't forget the CC def op on LTEBRCompare pseudos
Jonas Paulsson [Mon, 26 Oct 2015 15:03:32 +0000 (15:03 +0000)]
[SystemZ] Don't forget the CC def op on LTEBRCompare pseudos

Discovered by running fp-cmp-02.ll with -verify-machineinstrs (now added
to test run).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251297 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SystemZ] Tie operands in SystemZShorteInst if MI becomes 2-address.
Jonas Paulsson [Mon, 26 Oct 2015 15:03:07 +0000 (15:03 +0000)]
[SystemZ] Tie operands in SystemZShorteInst if MI becomes 2-address.

Discovered by testing fp-add-02.ll with -verify-machineinstrs.

Test case updated to always run with -verify-machineinstrs.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251296 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips] Check for the correct error message in tests for interrupt attributes.
Vasileios Kalintiris [Mon, 26 Oct 2015 14:24:30 +0000 (14:24 +0000)]
[mips] Check for the correct error message in tests for interrupt attributes.

Instead of XFAIL-ing the tests with the wrong usage of the "interrupt"
attribute, we should check that we emit the correct error messages to
the user.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251295 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ValueTracking] Extend r251146 to catch a fairly common case
James Molloy [Mon, 26 Oct 2015 14:10:46 +0000 (14:10 +0000)]
[ValueTracking] Extend r251146 to catch a fairly common case

Even though we may not know the value of the shifter operand, it's possible we know the shifter operand is non-zero. This can allow us to infer more known bits - for example:

  %1 = load %p !range {1, 5}
  %2 = shl %q, %1

We don't know %1, but we do know that it is nonzero so %2[0] is known zero, and importantly %2 is known non-zero.

Calling isKnownNonZero is nontrivially expensive so use an Optional to run it lazily and cache its result.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251294 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SCEV] Fix issues found during the review of r251283. NFC.
Silviu Baranga [Mon, 26 Oct 2015 13:50:06 +0000 (13:50 +0000)]
[SCEV] Fix issues found during the review of r251283. NFC.

Summary:
Replace (const SCEVAddRecExpr *) with cast<SCEVAddRecExpr>.

Rename SCEVApplyRewriter to SCEVLoopAddRecRewriter (which is a more
appropriate name) since the description is "takes a scalar evolution
expression and applies the Map (Loop -> SCEV) to all AddRecExprs."

Subscribers: llvm-commits, sanjoy

Differential Revision: http://reviews.llvm.org/D14065

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251292 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoLoop Vectorizer - skipping "bitcast" before GEP
Elena Demikhovsky [Mon, 26 Oct 2015 13:42:41 +0000 (13:42 +0000)]
Loop Vectorizer - skipping "bitcast" before GEP

Vectorization of memory instruction (Load/Store) is possible when the pointer is coming from GEP. The GEP analysis allows to estimate the profit.
In some cases we have a "bitcast" between GEP and memory instruction.
I added code that skips the "bitcast".

http://reviews.llvm.org/D13886

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251291 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoTests: be slightly more specific to avoid conflict with path.
Tim Northover [Mon, 26 Oct 2015 13:40:03 +0000 (13:40 +0000)]
Tests: be slightly more specific to avoid conflict with path.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251290 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agofix test errors (on windows) for commit r251287
Igor Breger [Mon, 26 Oct 2015 13:31:41 +0000 (13:31 +0000)]
fix test errors (on windows) for commit r251287

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251288 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAVX512: Enabled VPBROADCASTB lowering for v64i8 vectors.
Igor Breger [Mon, 26 Oct 2015 13:01:02 +0000 (13:01 +0000)]
AVX512: Enabled VPBROADCASTB lowering for v64i8 vectors.

Differential Revision: http://reviews.llvm.org/D13896

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251287 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[mips] Interrupt attribute support for mips32r2+.
Vasileios Kalintiris [Mon, 26 Oct 2015 12:38:43 +0000 (12:38 +0000)]
[mips] Interrupt attribute support for mips32r2+.

Summary:
This patch adds support for using the "interrupt" attribute on Mips
for interrupt handling functions. At this time only mips32r2+ with the
o32 ABI with the static relocation model is supported. Unsupported
configurations will be rejected

Patch by Simon Dardis (+ clang-format & some trivial changes to follow the
LLVM coding standards by me).

Reviewers: mpf, dsanders

Subscribers: dsanders, vkalintiris, llvm-commits

Differential Revision: http://reviews.llvm.org/D10768

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251286 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAVX-512: Use correct extract vector length.
Igor Breger [Mon, 26 Oct 2015 12:26:34 +0000 (12:26 +0000)]
AVX-512: Use correct extract vector length.

Bug https://llvm.org/bugs/show_bug.cgi?id=25318

Differential Revision: http://reviews.llvm.org/D14062

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251285 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[SCEV] Factor out common visiting patterns for SCEV rewriters. NFC.
Silviu Baranga [Mon, 26 Oct 2015 11:18:31 +0000 (11:18 +0000)]
[SCEV] Factor out common visiting patterns for SCEV rewriters. NFC.

Summary:
Add a SCEVRewriteVisitor class which contains the common
visiting patterns used when rewriting SCEVs.

SCEVParameterRewriter and SCEVApplyRewriter now inherit
from SCEVRewriteVisitor (and are therefore much simpler).

Reviewers: anemet, mzolotukhin, sanjoy

Subscribers: rengolin, llvm-commits, sanjoy

Differential Revision: http://reviews.llvm.org/D13242

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251283 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[InstCombine] Teach instcombine not to create extra PHI nodes when folding GEPs
Silviu Baranga [Mon, 26 Oct 2015 10:25:05 +0000 (10:25 +0000)]
[InstCombine] Teach instcombine not to create extra PHI nodes when folding GEPs

Summary:
InstCombine tries to transform GEP(PHI(GEP1, GEP2, ..)) into GEP(GEP(PHI(...))
when possible. However, this may leave the old PHI node around. Even if we
do end up folding the GEPs, having an extra PHI node might not be beneficial.

This change makes the transformation more conservative. We now only do this if
the PHI has only one use, and can therefore be removed after the transformation.

Reviewers: jmolloy, majnemer

Subscribers: mcrosier, mssimpso, llvm-commits

Differential Revision: http://reviews.llvm.org/D13887

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251281 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ARM] Handle the inline asm constraint type 'o'
James Molloy [Mon, 26 Oct 2015 10:04:52 +0000 (10:04 +0000)]
[ARM] Handle the inline asm constraint type 'o'

This means "memory with offset" and requires very little plumbing to get working. This fixes PR25317.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251280 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoDrop code after unreachable. No functionality change.
Benjamin Kramer [Mon, 26 Oct 2015 09:55:45 +0000 (09:55 +0000)]
Drop code after unreachable. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251278 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoAVX512: Add AVX-512 not materializable instructions.
Igor Breger [Mon, 26 Oct 2015 08:37:12 +0000 (08:37 +0000)]
AVX512: Add AVX-512 not materializable instructions.
Otherwise value can be reused , despite its value could be changed - produces incorrect assembler.

https://llvm.org/bugs/show_bug.cgi?id=25270

Differential Revision: http://reviews.llvm.org/D14057

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251275 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Orc] Add license header to OrcTargetSupport.
Lang Hames [Mon, 26 Oct 2015 06:40:28 +0000 (06:40 +0000)]
[Orc] Add license header to OrcTargetSupport.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251274 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[Orc] In the CompileOnDemand layer, wrap module ownership in order to preserve
Lang Hames [Mon, 26 Oct 2015 06:37:04 +0000 (06:37 +0000)]
[Orc] In the CompileOnDemand layer, wrap module ownership in order to preserve
the module pointer type passed in by the user.

The previous ownership scheme, where the user pointer was always moved into a
std::shared_ptr, breaks if the user passes in a raw pointer.

Discovered while working on the Orc C API, which should be landing shortly.
I expect to include a test-case with that.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251273 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoUpdate test to take into account for r251271.
David Majnemer [Mon, 26 Oct 2015 03:34:29 +0000 (03:34 +0000)]
Update test to take into account for r251271.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251272 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MC] Add support for GNU as-compatible binary operator precedence
David Majnemer [Mon, 26 Oct 2015 03:15:34 +0000 (03:15 +0000)]
[MC] Add support for GNU as-compatible binary operator precedence

GNU as and Darwin give the various binary operators different
precedence.  LLVM's MC supported the Darwin semantics but not the GNU
semantics.

This fixes PR25311.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251271 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[MC] Don't crash when .word is given bogus values
David Majnemer [Mon, 26 Oct 2015 02:45:50 +0000 (02:45 +0000)]
[MC] Don't crash when .word is given bogus values

We didn't validate that the .word directive was given a sane value,
leading to crashes when we attempt to write out the object file.

Instead, perform some validation and issue a diagnostic pointing at the
start of the diagnostic.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251270 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoRegionInfo: Correctly expand regions
Tobias Grosser [Sun, 25 Oct 2015 22:55:59 +0000 (22:55 +0000)]
RegionInfo: Correctly expand regions

Instead of playing around with dominance to verify if the possible expansion of
a scop region is indeed a single entry single exit region, we now distinguish
two cases. In case we only append a basic block, all edges entering this basic
block need to have come from within the region that is expanded. In case we join
two regions, the source basic blocks of the edges that end at the entry node of
the region that is appended most be part of either the original region or the
region that is appended.

This change will be tested through Polly.

This fixes llvm.org/PR25242

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251267 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoConvert assert(false) into llvm_unreachable where it makes sense.
Benjamin Kramer [Sun, 25 Oct 2015 22:28:27 +0000 (22:28 +0000)]
Convert assert(false) into llvm_unreachable where it makes sense.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251266 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][AVX] Regenerate tests.
Simon Pilgrim [Sun, 25 Oct 2015 21:47:09 +0000 (21:47 +0000)]
[X86][AVX] Regenerate tests.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251263 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agodocs: document `x` mangling in LangRef
Saleem Abdulrasool [Sun, 25 Oct 2015 20:39:35 +0000 (20:39 +0000)]
docs: document `x` mangling in LangRef

Windows has two different mangling specifiers.  `x` is used on x86 for the `_`
UserLabelPrefix.  Others use `w` for the no UserLabelPrefix.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251260 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ScalarEvolution] Throw away dead code.
Davide Italiano [Sun, 25 Oct 2015 20:00:49 +0000 (20:00 +0000)]
[ScalarEvolution] Throw away dead code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251256 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ScalarEvolution] Get rid of NDEBUG in header (correctly this time).
Davide Italiano [Sun, 25 Oct 2015 19:55:24 +0000 (19:55 +0000)]
[ScalarEvolution] Get rid of NDEBUG in header (correctly this time).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251255 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LCSSA] Unbreak build, don't reuse L; NFC
Sanjoy Das [Sun, 25 Oct 2015 19:27:17 +0000 (19:27 +0000)]
[LCSSA] Unbreak build, don't reuse L; NFC

The build broke in r251248.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251251 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[ScalarEvolution] Get rid of NDEBUG in header.
Davide Italiano [Sun, 25 Oct 2015 19:13:36 +0000 (19:13 +0000)]
[ScalarEvolution] Get rid of NDEBUG in header.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251249 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[LCSSA] Use range for loops; NFC
Sanjoy Das [Sun, 25 Oct 2015 19:08:32 +0000 (19:08 +0000)]
[LCSSA] Use range for loops; NFC

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251248 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE4A] Fix for EXTRQI shuffle lowering.
Simon Pilgrim [Sun, 25 Oct 2015 17:40:54 +0000 (17:40 +0000)]
[X86][SSE4A] Fix for EXTRQI shuffle lowering.

Incorrect range test - found during fuzz testing.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251245 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] Refreshed tests (missing AVX512 patterns)
Simon Pilgrim [Sun, 25 Oct 2015 15:39:22 +0000 (15:39 +0000)]
[X86][SSE] Refreshed tests (missing AVX512 patterns)

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251238 91177308-0d34-0410-b5e6-96231b3b80d8

6 years agoScalarizer for masked.gather and masked.scatter intrinsics.
Elena Demikhovsky [Sun, 25 Oct 2015 15:37:55 +0000 (15:37 +0000)]
Scalarizer for masked.gather and masked.scatter intrinsics.
When the target does not support these intrinsics they should be converted to a chain of scalar load or store operations.
If the mask is not constant, the scalarizer will build a chain of conditional basic blocks.
I added isLegalMaskedGather() isLegalMaskedScatter() APIs.

Differential Revision: http://reviews.llvm.org/D13722

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251237 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] Added tests for shuffling through bitcasts.
Simon Pilgrim [Sun, 25 Oct 2015 15:32:04 +0000 (15:32 +0000)]
[X86][SSE] Added tests for shuffling through bitcasts.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251236 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] vector sext/zext tests - remove unnecessary mcpu arguments
Simon Pilgrim [Sun, 25 Oct 2015 12:15:00 +0000 (12:15 +0000)]
[X86][SSE] vector sext/zext tests - remove unnecessary mcpu arguments

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251233 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86][SSE] shift/rotate tests - remove unnecessary mcpu arguments and regenerate...
Simon Pilgrim [Sun, 25 Oct 2015 12:07:45 +0000 (12:07 +0000)]
[X86][SSE] shift/rotate tests - remove unnecessary mcpu arguments and regenerate/cleanup

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251232 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] PMOV*X* tests - remove unnecessary mcpu arguments and regenerate
Simon Pilgrim [Sun, 25 Oct 2015 11:55:10 +0000 (11:55 +0000)]
[X86] PMOV*X* tests - remove unnecessary mcpu arguments and regenerate

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251230 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Stack folding tests - just use mtriple - no need for mcpu in these tests
Simon Pilgrim [Sun, 25 Oct 2015 11:42:46 +0000 (11:42 +0000)]
[X86] Stack folding tests - just use mtriple - no need for mcpu in these tests

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251229 91177308-0d34-0410-b5e6-96231b3b80d8

6 years ago[X86] Use correct calling convention for MCU psABI libcalls
Michael Kuperstein [Sun, 25 Oct 2015 08:14:05 +0000 (08:14 +0000)]
[X86] Use correct calling convention for MCU psABI libcalls

When using the MCU psABI, compiler-generated library calls should pass
some parameters in-register. However, since inreg marking for x86 is currently
done by the front end, it will not be applied to backend-generated calls.

This is a workaround for PR3997, which describes a similar issue for -mregparm.

Differential Revision: http://reviews.llvm.org/D13977

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@251223 91177308-0d34-0410-b5e6-96231b3b80d8