platform/upstream/llvm.git
5 years ago[Codegen][SelectionDAG] X u% C == 0 fold: non-splat vector improvements
Roman Lebedev [Sat, 20 Jul 2019 16:33:15 +0000 (16:33 +0000)]
[Codegen][SelectionDAG] X u% C == 0 fold: non-splat vector improvements

Summary:
Four things here:
1. Generalize the fold to handle non-splat divisors. Reasonably trivial.
2. Unban power-of-two divisors. I don't see any reason why they should
   be illegal.
   * There is no ban in Hacker's Delight
   * I think the ban came from the same bug that caused the miscompile
      in the base patch - in `floor((2^W - 1) / D)` we were dividing by
      `D0` instead of `D`, and we **were** ensuring that `D0` is not `1`,
      which made sense.
3. Unban `1` divisors. I no longer believe Hacker's Delight actually says
   that the fold is invalid for `D = 0`. Further considerations:
   * We know that
     * `(X u% 1) == 0`  can be constant-folded to `1`,
     * `(X u% 1) != 0`  can be constant-folded to `0`,
   *  Also, we know that
     * `X u<= -1` can be constant-folded to `1`,
     * `X u>  -1` can be constant-folded to `0`,
   * https://godbolt.org/z/7jnZJX https://rise4fun.com/Alive/oF6p
   * We know will end up with the following:
       `(setule/setugt (rotr (mul N, P), K), Q)`
   * Therefore, for given new DAG nodes and comparison predicates
     (`ule`/`ugt`), we will still produce the correct answer if:
     `Q` is a all-ones constant; and both `P` and `K` are *anything*
     other than `undef`.
   * The fold will indeed produce `Q = all-ones`.
4. Try to re-splat the `P` and `K` vectors - we don't care about
   their values for the lanes where divisor was `1`.

Reviewers: RKSimon, hermord, craig.topper, spatel, xbolva00

Reviewed By: RKSimon

Subscribers: hiraditya, javed.absar, dexonsmith, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D63963

llvm-svn: 366637

5 years ago[X86][SSE] Use PSADBW to improve vXi8 sum reduction (PR42674)
Simon Pilgrim [Sat, 20 Jul 2019 15:20:11 +0000 (15:20 +0000)]
[X86][SSE] Use PSADBW to improve vXi8 sum reduction (PR42674)

As detailed on PR42674, we can reduce a vXi8 down until we have the final <8 x i8>, and then use PSADBW with zero, to sum those values. We then extract the bottom i8, discarding any overflow from the upper bits of the i16 result.

llvm-svn: 366636

5 years agoFix cppcheck reduce scope variable warnings. NFCI
Simon Pilgrim [Sat, 20 Jul 2019 13:01:16 +0000 (13:01 +0000)]
Fix cppcheck reduce scope variable warnings. NFCI

Move a couple of variables inside the block where they are actually needed.

llvm-svn: 366635

5 years agoRevert "Fix asan infinite loop on undefined symbol"
Serge Guelton [Sat, 20 Jul 2019 13:00:12 +0000 (13:00 +0000)]
Revert "Fix asan infinite loop on undefined symbol"

This reverts commit cbd28cd05bb11e9d76d71c6cc2d38d89dbb1fe1a.

Buildbot fail: http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux/builds/22434/steps/64-bit%20check-asan-dynamic/

llvm-svn: 366634

5 years ago[Local] Zap blockaddress without users in ConstantFoldTerminator.
Florian Hahn [Sat, 20 Jul 2019 12:25:47 +0000 (12:25 +0000)]
[Local] Zap blockaddress without users in ConstantFoldTerminator.

If the blockaddress is not destoryed, the destination block will still
be marked as having its address taken, limiting further transformations.

I think there are other places where the dead blockaddress constants are kept
around, I'll look into that as follow up.

Reviewers: craig.topper, brzycki, davide

Reviewed By: brzycki, davide

Differential Revision: https://reviews.llvm.org/D64936

llvm-svn: 366633

5 years agoFix asan infinite loop on undefined symbol
Serge Guelton [Sat, 20 Jul 2019 12:01:18 +0000 (12:01 +0000)]
Fix asan infinite loop on undefined symbol

Fix llvm#39641

Recommit of r366413

Differential Revision: https://reviews.llvm.org/D63877

llvm-svn: 366632

5 years ago[CMake] Align debugserver with lldb-server on Darwin
Stefan Granitz [Sat, 20 Jul 2019 11:18:31 +0000 (11:18 +0000)]
[CMake] Align debugserver with lldb-server on Darwin

Summary: Make debugserver a tool like lldb-server, so it can be included/excluded via `LLDB_TOOL_DEBUGSERVER_BUILD`. This replaces the old `LLDB_NO_DEBUGSERVER` flag. Doing the same for darwin-debug while I am here.

Reviewers: xiaobai, JDevlieghere, davide

Reviewed By: xiaobai, JDevlieghere

Subscribers: mgorny, lldb-commits, #lldb

Tags: #lldb

Differential Revision: https://reviews.llvm.org/D64994

llvm-svn: 366631

5 years ago[c++20] P1161R3: a[b,c] is deprecated.
Richard Smith [Sat, 20 Jul 2019 09:32:27 +0000 (09:32 +0000)]
[c++20] P1161R3: a[b,c] is deprecated.

llvm-svn: 366630

5 years agoMark P1301R4 in C++2a as being SVN instead.
Aaron Ballman [Sat, 20 Jul 2019 08:57:08 +0000 (08:57 +0000)]
Mark P1301R4 in C++2a as being SVN instead.

llvm-svn: 366629

5 years agoWe support P1301R4 in C++2a as of r366626.
Aaron Ballman [Sat, 20 Jul 2019 08:24:56 +0000 (08:24 +0000)]
We support P1301R4 in C++2a as of r366626.

llvm-svn: 366628

5 years ago[cxx_status] Update status page for WG21 Cologne meeting motions.
Richard Smith [Sat, 20 Jul 2019 08:20:54 +0000 (08:20 +0000)]
[cxx_status] Update status page for WG21 Cologne meeting motions.

Note that many of the paper links will be dead until the post-meeting
mailing is released.

llvm-svn: 366627

5 years agoImplement P1301R4, which allows specifying an optional message on the [[nodiscard...
Aaron Ballman [Sat, 20 Jul 2019 07:56:34 +0000 (07:56 +0000)]
Implement P1301R4, which allows specifying an optional message on the [[nodiscard]] attribute.

This also bumps the attribute feature test value and introduces the notion of a C++2a extension warning.

llvm-svn: 366626

5 years ago[GlobalISel][AArch64] Contract trivial same-size cross-bank copies into G_STOREs
Jessica Paquette [Sat, 20 Jul 2019 01:55:35 +0000 (01:55 +0000)]
[GlobalISel][AArch64] Contract trivial same-size cross-bank copies into G_STOREs

Sometimes, you can end up with cross-bank copies between same-sized GPRs and
FPRs, which feed into G_STOREs. When these copies feed only into stores, they
aren't necessary; we can just store using the original register bank.

This provides some minor code size savings for some floating point SPEC
benchmarks. (Around 0.2% for 453.povray and 450.soplex)

This issue doesn't seem to show up due to regbankselect or anything similar. So,
this patch introduces an early select function, `contractCrossBankCopyIntoStore`
which performs the contraction when possible. The selector then continues
normally and selects the correct store opcode, eliminating needless copies
along the way.

Differential Revision: https://reviews.llvm.org/D65024

llvm-svn: 366625

5 years ago[WebAssembly] Compute and export TLS block alignment
Guanzhong Chen [Fri, 19 Jul 2019 23:34:16 +0000 (23:34 +0000)]
[WebAssembly] Compute and export TLS block alignment

Summary:
Add immutable WASM global `__tls_align` which stores the alignment
requirements of the TLS segment.

Add `__builtin_wasm_tls_align()` intrinsic to get this alignment in Clang.

The expected usage has now changed to:

    __wasm_init_tls(memalign(__builtin_wasm_tls_align(),
                             __builtin_wasm_tls_size()));

Reviewers: tlively, aheejin, sbc100, sunfish, alexcrichton

Reviewed By: tlively

Subscribers: dschuff, jgravelle-google, hiraditya, cfe-commits, llvm-commits

Tags: #clang, #llvm

Differential Revision: https://reviews.llvm.org/D65028

llvm-svn: 366624

5 years ago[LTO] Always mark regular LTO units with EnableSplitLTOUnit=1
Teresa Johnson [Fri, 19 Jul 2019 23:02:58 +0000 (23:02 +0000)]
[LTO] Always mark regular LTO units with EnableSplitLTOUnit=1

Summary:
Regular LTO modules do not need LTO Unit splitting, only ThinLTO does
(they must be consistently split into regular and Thin units for
optimizations such as whole program devirtualization and lower type
tests). In order to avoid spurious errors from LTO when combining with
split ThinLTO modules, always set this flag for regular LTO modules.

Reviewers: pcc

Subscribers: mehdi_amini, Prazek, inglorion, steven_wu, dexonsmith, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D65009

llvm-svn: 366623

5 years agoRe-commit: r366610 and r366612: Expand pseudo-components before embedding in llvm...
Daniel Sanders [Fri, 19 Jul 2019 22:46:47 +0000 (22:46 +0000)]
Re-commit: r366610 and r366612: Expand pseudo-components before embedding in llvm-config

There were two main problems:
* The 'nativecodegen' pseudo-component was unconditionally adding
  ${native_tgt}CodeGen even though it conditionally added ${native_tgt}Info and
  ${native_tgt}Desc. This has been fixed by making ${native_tgt}CodeGen
  conditional too
* The 'all' pseudo-component was causing library names like LLVMLLVMDemangle as
  the expansion was to a library name and not a component. There doesn't seem to
  be a list of available components anywhere so this has been fixed by moving the
  expansion of 'all' back where it was before. This manifested in different ways
  on different builders but it was the same root cause

llvm-svn: 366622

5 years agoAMDGPU/GlobalISel: Legalize GEP for other 32-bit address spaces
Matt Arsenault [Fri, 19 Jul 2019 22:28:44 +0000 (22:28 +0000)]
AMDGPU/GlobalISel: Legalize GEP for other 32-bit address spaces

llvm-svn: 366621

5 years ago[NFC] Remove unused variable
Denis Bakhvalov [Fri, 19 Jul 2019 21:59:42 +0000 (21:59 +0000)]
[NFC] Remove unused variable

Change-Id: I5aee24dcdf6eebfbf788e52be22463387f23d927
llvm-svn: 366620

5 years ago[AMDGPU] Autogenerate register sequences in tuples
Stanislav Mekhanoshin [Fri, 19 Jul 2019 21:43:42 +0000 (21:43 +0000)]
[AMDGPU] Autogenerate register sequences in tuples

Differential Revision: https://reviews.llvm.org/D65007

llvm-svn: 366619

5 years agoRevert "Fix asan infinite loop on undefined symbol"
Matthew Voss [Fri, 19 Jul 2019 21:41:07 +0000 (21:41 +0000)]
Revert "Fix asan infinite loop on undefined symbol"

This reverts commit 8e46275488ca453a24964cb26be0736172ad1db8.

This was failing on sanitizer-x86_64-linux and our internal CI.

llvm-svn: 366618

5 years agoDisallow most calling convention attributes on PS4
Sunil Srivastava [Fri, 19 Jul 2019 21:38:34 +0000 (21:38 +0000)]
Disallow most calling convention attributes on PS4

PS4 now only allows "cdecl", and its equivalent on PS4, "sysv_abi".

Differential Revision: https://reviews.llvm.org/D64780

llvm-svn: 366617

5 years ago[AMDGPU] Fixed occupancy calculation for gfx10
Stanislav Mekhanoshin [Fri, 19 Jul 2019 21:29:51 +0000 (21:29 +0000)]
[AMDGPU] Fixed occupancy calculation for gfx10

Differential Revision: https://reviews.llvm.org/D65010

llvm-svn: 366616

5 years agoRevert r366610 and r366612: Expand pseudo-components before embedding in llvm-config
Daniel Sanders [Fri, 19 Jul 2019 21:11:05 +0000 (21:11 +0000)]
Revert r366610 and r366612: Expand pseudo-components before embedding in llvm-config

Some targets are missing LLVMDemangle, one is adding the LLVM prefix twice, and two
are hitting the very error this patch fixes for my target. Reverting while I work
through the reports.

llvm-svn: 366615

5 years ago[InstCombine] Fix copy/paste mistake in the test cases I added for PR42691. NFC
Craig Topper [Fri, 19 Jul 2019 21:09:21 +0000 (21:09 +0000)]
[InstCombine] Fix copy/paste mistake in the test cases I added for PR42691. NFC

llvm-svn: 366614

5 years agoAMDGPU: Avoid custom predicates for stores with glue
Matt Arsenault [Fri, 19 Jul 2019 21:01:30 +0000 (21:01 +0000)]
AMDGPU: Avoid custom predicates for stores with glue

llvm-svn: 366613

5 years agoFix a latent bug discovered by r366610: nativecodegen includes X86CodeGen when X86...
Daniel Sanders [Fri, 19 Jul 2019 20:58:11 +0000 (20:58 +0000)]
Fix a latent bug discovered by r366610: nativecodegen includes X86CodeGen when X86 is not compiled

I believe this to have been a latent bug as the same expansion checks for the
existence of ${native_tgt}Info and ${native_tgt}Desc and only adds them if
they were compiled but unconditionally adds ${native_tgt}CodeGen.

This should fix llvm-clang-x86_64-win-fast which builds ARM only on an X86 host and similar builders.

llvm-svn: 366612

5 years ago[InstCombine] Add test cases for PR42691. NFC
Craig Topper [Fri, 19 Jul 2019 20:48:52 +0000 (20:48 +0000)]
[InstCombine] Add test cases for PR42691. NFC

llvm-svn: 366611

5 years agoExpand pseudo-components before embedding in llvm-config
Daniel Sanders [Fri, 19 Jul 2019 20:38:05 +0000 (20:38 +0000)]
Expand pseudo-components before embedding in llvm-config

Summary:
If you use pseudo-targets like AllTargetsCodeGens in LLVM_DYLIB_COMPONENTS
then a test will fail because `./bin/llvm-config --shared-mode` can't
handle these targets. We can fix this by expanding them before embedding
the string into llvm-config

Reviewers: bogner

Reviewed By: bogner

Subscribers: mgorny, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D65011

llvm-svn: 366610

5 years agoAMDGPU: Redefine setcc condition PatLeafs
Matt Arsenault [Fri, 19 Jul 2019 20:24:40 +0000 (20:24 +0000)]
AMDGPU: Redefine setcc condition PatLeafs

Avoid using custom code predicates.

llvm-svn: 366609

5 years agoAMDGPU: Don't rely on m0 being -1 for GWS offsets
Matt Arsenault [Fri, 19 Jul 2019 20:01:24 +0000 (20:01 +0000)]
AMDGPU: Don't rely on m0 being -1 for GWS offsets

This only works if the high bits of m0 are also 0, so m0 would have to
be set to 0xffff.

llvm-svn: 366608

5 years agoAMDGPU: Force s_waitcnt after GWS instructions
Matt Arsenault [Fri, 19 Jul 2019 19:47:30 +0000 (19:47 +0000)]
AMDGPU: Force s_waitcnt after GWS instructions

This is apparently required to be the immediately following
instruction, so force it into a bundle with a waitcnt.

llvm-svn: 366607

5 years ago[libc++] Mark libcpp_deallocate.sh.cpp as UNSUPPORTED instead of XFAIL on AppleClang 9
Louis Dionne [Fri, 19 Jul 2019 19:41:40 +0000 (19:41 +0000)]
[libc++] Mark libcpp_deallocate.sh.cpp as UNSUPPORTED instead of XFAIL on AppleClang 9

Some minor versions of AppleClang 9 appear not to fail the test. It's
such a mess that the only sane thing to do is to mark the test as
UNSUPPORTED.

llvm-svn: 366606

5 years agoLiveIntervals: Fix handleMove asserting on BUNDLE
Matt Arsenault [Fri, 19 Jul 2019 19:32:00 +0000 (19:32 +0000)]
LiveIntervals: Fix handleMove asserting on BUNDLE

The top-level BUNDLE instruction should behave as an ordinary
instruction. It is supposed to have all relevant registers as implicit
operands. Moving it should work as any other instruction. I believe
the assert intended to avoid moving instructions inside bundles.

llvm-svn: 366605

5 years agoRevert "[libc++] Integrate the PSTL into libc++"
Louis Dionne [Fri, 19 Jul 2019 18:52:46 +0000 (18:52 +0000)]
Revert "[libc++] Integrate the PSTL into libc++"

This reverts r366593, which caused unforeseen breakage on the build bots.
I'm reverting until the problems have been figured out and fixed.

llvm-svn: 366603

5 years ago[AMDGPU] Add test case on crashing of `si-lower-sgpr-spills` pass
Michael Liao [Fri, 19 Jul 2019 18:50:53 +0000 (18:50 +0000)]
[AMDGPU] Add test case on crashing of `si-lower-sgpr-spills` pass

Reviewers: arsenm

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

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64273

llvm-svn: 366602

5 years ago[libc++] Allow passing additional CMake arguments in macOS trunk CI script
Louis Dionne [Fri, 19 Jul 2019 18:47:00 +0000 (18:47 +0000)]
[libc++] Allow passing additional CMake arguments in macOS trunk CI script

llvm-svn: 366601

5 years agoRevert "Use the MachineBasicBlock symbol for a callbr target"
Nick Desaulniers [Fri, 19 Jul 2019 18:18:02 +0000 (18:18 +0000)]
Revert "Use the MachineBasicBlock symbol for a callbr target"

This reverts commit r366523/ccbffefccaff42b0d094c9ef0f49fc3e8c8456ea.

Two regressions were immediately reported:
- https://github.com/ClangBuiltLinux/linux/issues/614
- https://github.com/ClangBuiltLinux/linux/issues/615

Reported-by: nathanchance
llvm-svn: 366600

5 years ago[RISCV] Disable tests failing on buildbots.
Matt Morehouse [Fri, 19 Jul 2019 18:05:12 +0000 (18:05 +0000)]
[RISCV] Disable tests failing on buildbots.

r366399 enabled a couple tests that are failing on a few buildbots.

llvm-svn: 366599

5 years ago[AMDGPU] Allow register tuples to set asm names
Stanislav Mekhanoshin [Fri, 19 Jul 2019 18:05:01 +0000 (18:05 +0000)]
[AMDGPU] Allow register tuples to set asm names

This change reverts most of the previous register name generation.
The real problem is that RegisterTuple does not generate asm names.
Added optional operand to RegisterTuple. This way we can simplify
register name access and dramatically reduce the size of static
tables for the backend.

Differential Revision: https://reviews.llvm.org/D64967

llvm-svn: 366598

5 years agoAMDGPU/GlobalISel: Fix MMO flags for kernel argument loads
Matt Arsenault [Fri, 19 Jul 2019 17:52:56 +0000 (17:52 +0000)]
AMDGPU/GlobalISel: Fix MMO flags for kernel argument loads

The DAG lowering sets dereferencable and invariant, not nontemporal.

llvm-svn: 366597

5 years agoGlobalISel: Add GINodeEquiv for fcopysign
Matt Arsenault [Fri, 19 Jul 2019 17:32:19 +0000 (17:32 +0000)]
GlobalISel: Add GINodeEquiv for fcopysign

I don't need this at the moment, but it should be here.

llvm-svn: 366596

5 years ago[llvm-lipo] Remove trailing whitespace. NFC
Shoaib Meenai [Fri, 19 Jul 2019 17:19:57 +0000 (17:19 +0000)]
[llvm-lipo] Remove trailing whitespace. NFC

llvm-svn: 366595

5 years ago[libc++] Use _EnableIf instead of std::enable_if in deduction guides for map and set
Louis Dionne [Fri, 19 Jul 2019 17:13:39 +0000 (17:13 +0000)]
[libc++] Use _EnableIf instead of std::enable_if in deduction guides for map and set

llvm-svn: 366594

5 years ago[libc++] Integrate the PSTL into libc++
Louis Dionne [Fri, 19 Jul 2019 17:02:42 +0000 (17:02 +0000)]
[libc++] Integrate the PSTL into libc++

Summary:
This commit allows specifying LIBCXX_ENABLE_PARALLEL_ALGORITHMS when
configuring libc++ in CMake. When that option is enabled, libc++ will
assume that the PSTL can be found somewhere on the CMake module path,
and it will provide the C++17 parallel algorithms based on the PSTL
(that is assumed to be available).

The commit also adds support for running the PSTL tests as part of
the libc++ test suite.

Reviewers: rodgert, EricWF

Subscribers: mgorny, christof, jkorous, dexonsmith, libcxx-commits, mclow.lists, EricWF

Tags: #libc

Differential Revision: https://reviews.llvm.org/D60480

llvm-svn: 366593

5 years ago[Format/ObjC] Avoid breaking between unary operators and operands
Ben Hamilton [Fri, 19 Jul 2019 16:50:24 +0000 (16:50 +0000)]
[Format/ObjC] Avoid breaking between unary operators and operands

Summary:
Test Plan:
  New tests added. Ran tests with:
  % ninja FormatTests && ./tools/clang/unittests/Format/FormatTests
  Confirmed tests failed before change and passed after change.

Reviewers: krasimir, djasper, sammccall, klimek

Reviewed By: sammccall

Subscribers: klimek, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D64775

llvm-svn: 366592

5 years agoAMDGPU: Add some function return test cases
Matt Arsenault [Fri, 19 Jul 2019 16:45:48 +0000 (16:45 +0000)]
AMDGPU: Add some function return test cases

llvm-svn: 366591

5 years ago[lldb][NFC] Cleanup mentions and code related to lldb-mi
Raphael Isemann [Fri, 19 Jul 2019 15:55:23 +0000 (15:55 +0000)]
[lldb][NFC] Cleanup mentions and code related to lldb-mi

Summary: lldb-mi has been removed, but there are still a bunch of references in the code base. This patch removes all of them.

Reviewers: JDevlieghere, jfb

Reviewed By: JDevlieghere

Subscribers: dexonsmith, ki.stfu, mgorny, abidh, jfb, lldb-commits

Tags: #lldb

Differential Revision: https://reviews.llvm.org/D64992

llvm-svn: 366590

5 years ago[AMDGPU] Regenerate test file for upcoming patch. NFCI.
Simon Pilgrim [Fri, 19 Jul 2019 15:43:56 +0000 (15:43 +0000)]
[AMDGPU] Regenerate test file for upcoming patch. NFCI.

llvm-svn: 366589

5 years agoFix asan infinite loop on undefined symbol
Serge Guelton [Fri, 19 Jul 2019 15:20:36 +0000 (15:20 +0000)]
Fix asan infinite loop on undefined symbol

Fix llvm#39641

Recommit of r366413

Differential Revision: https://reviews.llvm.org/D63877

llvm-svn: 366588

5 years ago[libunwind][ARM] Fix loading FP registers on big-endian targets
Mikhail Maltsev [Fri, 19 Jul 2019 15:20:32 +0000 (15:20 +0000)]
[libunwind][ARM] Fix loading FP registers on big-endian targets

Summary:
The function Unwind-EHABI.cpp:_Unwind_VRS_Pop loads the saved values of
64-bit FP registers as two 32-bit words because they might not be
8-byte aligned. Combining these words into a 64-bit value has to be
done differently on big-endian platforms.

Reviewers: ostannard, john.brawn, dmgreen

Reviewed By: ostannard

Subscribers: kristof.beyls, christof, libcxx-commits

Tags: #libc

Differential Revision: https://reviews.llvm.org/D64996

llvm-svn: 366587

5 years agoAMDGPU: Attempt to fix bot error
Matt Arsenault [Fri, 19 Jul 2019 14:56:24 +0000 (14:56 +0000)]
AMDGPU: Attempt to fix bot error

Manually remove file name from check line, since it somehow ends
up being different on an msvc bot.

llvm-svn: 366586

5 years agoAMDGPU/GlobalISel: Selection for fminnum/fmaxnum
Matt Arsenault [Fri, 19 Jul 2019 14:42:40 +0000 (14:42 +0000)]
AMDGPU/GlobalISel: Selection for fminnum/fmaxnum

v2f16 case doesn't work yet because the VOP3P complex patterns haven't
been ported yet.

llvm-svn: 366585

5 years agoAMDGPU/GlobalISel: Support arguments with multiple registers
Matt Arsenault [Fri, 19 Jul 2019 14:29:30 +0000 (14:29 +0000)]
AMDGPU/GlobalISel: Support arguments with multiple registers

Handles structs used directly in argument lists.

llvm-svn: 366584

5 years agoAMDGPU/GlobalISel: Rewrite lowerFormalArguments
Matt Arsenault [Fri, 19 Jul 2019 14:15:18 +0000 (14:15 +0000)]
AMDGPU/GlobalISel: Rewrite lowerFormalArguments

This should now handle everything except structs passed as multiple
registers.

I think most of the packing logic should be handled by
handleAssignments, but I'm unclear on what the contract is for
multiple registers. This is copying how x86 handles this.

This does change the behavior of the test_sgpr_alignment0 amdgpu_vs
test. I don't think shader arguments should try to follow the
alignment, and registers need to be repacked. I also don't think it
matters, since I think the pointers are packed to the beginning of the
argument list anyway.

llvm-svn: 366582

5 years agoSupport Linux signal return trampolines in frame initialization
Joseph Tremoulet [Fri, 19 Jul 2019 14:05:55 +0000 (14:05 +0000)]
Support Linux signal return trampolines in frame initialization

Summary:
Add __kernel_rt_sigreturn to the list of trap handlers for Linux (it's
used as such on aarch64 at least), and __restore_rt as well (used on
x86_64).

Skip decrement-and-recompute for trap handlers in
InitializeNonZerothFrame, as signal dispatch may point the child frame's
return address to the start of the return trampoline.

Parse the 'S' flag for signal handlers from eh_frame augmentation, and
propagate it to the unwind plan.

Reviewers: labath, jankratochvil, compnerd, jfb, jasonmolenda

Reviewed By: jasonmolenda

Subscribers: clayborg, MaskRay, wuzish, nemanjai, kbarton, jrtc27, atanasyan, jsji, javed.absar, kristof.beyls, lldb-commits

Tags: #lldb

Differential Revision: https://reviews.llvm.org/D63667

llvm-svn: 366580

5 years ago[libc++] Add missing %link_flags to .sh.cpp test
Louis Dionne [Fri, 19 Jul 2019 14:01:48 +0000 (14:01 +0000)]
[libc++] Add missing %link_flags to .sh.cpp test

Without the link flags, the test always fails on Linux. For some reason,
however, it works on Darwin -- which is why it wasn't caught at first.

llvm-svn: 366579

5 years agoAMDGPU: Decompose all values to 32-bit pieces for calling conventions
Matt Arsenault [Fri, 19 Jul 2019 13:57:44 +0000 (13:57 +0000)]
AMDGPU: Decompose all values to 32-bit pieces for calling conventions

This is the more natural lowering, and presents more opportunities to
reduce 64-bit ops to 32-bit.

This should also help avoid issues graphics shaders have had with
64-bit values, and simplify argument lowering in globalisel.

llvm-svn: 366578

5 years ago[clangd] Provide a way to publish highlightings in non-racy manner
Ilya Biryukov [Fri, 19 Jul 2019 13:51:01 +0000 (13:51 +0000)]
[clangd] Provide a way to publish highlightings in non-racy manner

Summary:
By exposing a callback that can guard code publishing results of
'onMainAST' callback in the same manner we guard diagnostics.

Reviewers: sammccall

Reviewed By: sammccall

Subscribers: javed.absar, MaskRay, jkorous, arphaman, kadircet, hokein, jvikstrom, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D64985

llvm-svn: 366577

5 years agogn build: Set +x on symlink_or_copy.py
Nico Weber [Fri, 19 Jul 2019 13:40:54 +0000 (13:40 +0000)]
gn build: Set +x on symlink_or_copy.py

llvm-svn: 366576

5 years ago[clangd] Disable background-index on lit-tests by default
Kadir Cetinkaya [Fri, 19 Jul 2019 13:40:30 +0000 (13:40 +0000)]
[clangd] Disable background-index on lit-tests by default

Summary:
Since background-index can perform disk writes, we don't want to turn
it on tests that won't clear it.

Reviewers: sammccall

Subscribers: ilya-biryukov, MaskRay, jkorous, arphaman, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D64990

llvm-svn: 366575

5 years agoDAG: Handle dbg_value for arguments split into multiple subregs
Matt Arsenault [Fri, 19 Jul 2019 13:36:46 +0000 (13:36 +0000)]
DAG: Handle dbg_value for arguments split into multiple subregs

This was handled previously for arguments split due to not fitting in
an MVT. This was dropping the register for argument registers split
due to TLI::getRegisterTypeForCallingConv.

llvm-svn: 366574

5 years agolld-link: Demangle symbols from archives in diagnostics
Nico Weber [Fri, 19 Jul 2019 13:29:10 +0000 (13:29 +0000)]
lld-link: Demangle symbols from archives in diagnostics

Also add test coverage for thin archives (which are the only way I could
come up with to test at least some of the diagnostic changes).

Differential Revision: https://reviews.llvm.org/D64927

llvm-svn: 366573

5 years ago[NFC] include cstdint/string prior to using uint8_t/string
Than McIntosh [Fri, 19 Jul 2019 13:13:54 +0000 (13:13 +0000)]
[NFC] include cstdint/string prior to using uint8_t/string

Summary: include proper header prior to use of uint8_t typedef
and std::string.

Subscribers: llvm-commits

Reviewers: cherry

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64937

llvm-svn: 366572

5 years ago[AMDGPU][MC] Corrected parsing of branch offsets
Dmitry Preobrazhensky [Fri, 19 Jul 2019 13:12:47 +0000 (13:12 +0000)]
[AMDGPU][MC] Corrected parsing of branch offsets

See bug 40820: https://bugs.llvm.org/show_bug.cgi?id=40820

Reviewers: artem.tamazov, arsenm

Differential Revision: https://reviews.llvm.org/D64629

llvm-svn: 366571

5 years ago[MachineCSE][MachinePRE] Avoid hoisting code from code regions into hot BBs.
Kai Luo [Fri, 19 Jul 2019 12:58:16 +0000 (12:58 +0000)]
[MachineCSE][MachinePRE] Avoid hoisting code from code regions into hot BBs.

Summary:
Current PRE hoists common computations into
CMBB = DT->findNearestCommonDominator(MBB, MBB1).
However, if CMBB is in a hot loop body, we might get performance
degradation.

Differential Revision: https://reviews.llvm.org/D64394

llvm-svn: 366570

5 years ago[X86] for split stack, not save/restore nested arg if unused
Than McIntosh [Fri, 19 Jul 2019 12:54:44 +0000 (12:54 +0000)]
[X86] for split stack, not save/restore nested arg if unused

Summary:
For split-stack, if the nested argument (i.e. R10) is not used, no need to save/restore it in the prologue.

Reviewers: thanm

Reviewed By: thanm

Subscribers: mstorsjo, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64673

llvm-svn: 366569

5 years ago[Clangd] Fixed ExtractVariable test
Shaurya Gupta [Fri, 19 Jul 2019 12:11:04 +0000 (12:11 +0000)]
[Clangd] Fixed ExtractVariable test

llvm-svn: 366568

5 years ago[libc++] Fix link error with _LIBCPP_HIDE_FROM_ABI_PER_TU and std::string
Louis Dionne [Fri, 19 Jul 2019 11:52:55 +0000 (11:52 +0000)]
[libc++] Fix link error with _LIBCPP_HIDE_FROM_ABI_PER_TU and std::string

Summary:
This is effectively a revert of r344616, which was a partial fix for
PR38964 (compilation of <string> with GCC in C++03 mode). However, that
configuration is explicitly not supported anymore and that partial fix
breaks compilation with Clang when per-TU insulation is provided.

PR42676
rdar://52899715

Reviewers: mclow.lists, EricWF

Subscribers: christof, jkorous, dexonsmith, libcxx-commits

Tags: #libc

Differential Revision: https://reviews.llvm.org/D64941

llvm-svn: 366567

5 years ago[Clangd] Fixed SelectionTree bug for macros
Shaurya Gupta [Fri, 19 Jul 2019 11:41:02 +0000 (11:41 +0000)]
[Clangd] Fixed SelectionTree bug for macros

Summary:
Fixed SelectionTree bug for macros
- Fixed SelectionTree claimRange for macros and template instantiations
- Fixed SelectionTree unit tests
- Changed a breaking test in TweakTests

Reviewers: sammccall, kadircet

Subscribers: ilya-biryukov, MaskRay, jkorous, arphaman, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D64329

llvm-svn: 366566

5 years ago[NFC][InstCombine] Tests for 'rem' formation from sub-of-mul-by-'div' (PR42673)
Roman Lebedev [Fri, 19 Jul 2019 11:29:18 +0000 (11:29 +0000)]
[NFC][InstCombine] Tests for 'rem' formation from sub-of-mul-by-'div' (PR42673)

https://rise4fun.com/Alive/8Rp
https://bugs.llvm.org/show_bug.cgi?id=42673

llvm-svn: 366565

5 years ago[NFC][InstCombine] Redundant masking before left-shift: tests with assume
Roman Lebedev [Fri, 19 Jul 2019 11:29:04 +0000 (11:29 +0000)]
[NFC][InstCombine] Redundant masking before left-shift: tests with assume

If the legality check is `(shiftNbits-maskNbits) s>= 0`,
then we can simplify it to `shiftNbits u>= maskNbits`,
which is easier to check for.

However, currently switching the `dropRedundantMaskingOfLeftShiftInput()`
to `SimplifyICmpInst()` does not catch these cases and regresses
currently-handled cases, so i'll leave it as is for now.

https://rise4fun.com/Alive/25P

llvm-svn: 366564

5 years agoFix MSVC "result of 32-bit shift implicitly converted to 64 bits" warning. NFCI.
Simon Pilgrim [Fri, 19 Jul 2019 11:18:46 +0000 (11:18 +0000)]
Fix MSVC "result of 32-bit shift implicitly converted to 64 bits" warning. NFCI.

llvm-svn: 366563

5 years agoDon't update NoTrappingFPMath and FPDenormalMode in resetTargetOptions
Oliver Stannard [Fri, 19 Jul 2019 10:37:37 +0000 (10:37 +0000)]
Don't update NoTrappingFPMath and FPDenormalMode in resetTargetOptions

We'd like to remove this whole function, because these are properties of
functions, not the target as a whole. These two are easy to remove
because they are only used for emitting ARM build attributes, which
expects them to represent the defaults for the whole module, not just
the last function generated.

This is needed to get correct build attributes when using IPRA on ARM,
because IPRA causes resetTargetOptions to get called before
ARMAsmPrinter::emitAttributes.

Differential revision: https://reviews.llvm.org/D64929

llvm-svn: 366562

5 years ago[lldb][NFC] Tablegenify target
Raphael Isemann [Fri, 19 Jul 2019 10:23:22 +0000 (10:23 +0000)]
[lldb][NFC] Tablegenify target

llvm-svn: 366561

5 years ago[NFC] Remove indent after r366433
Stefan Granitz [Fri, 19 Jul 2019 10:20:35 +0000 (10:20 +0000)]
[NFC] Remove indent after r366433

llvm-svn: 366560

5 years agoRevert "Revert r366458, r366467 and r366468"
Kadir Cetinkaya [Fri, 19 Jul 2019 10:18:52 +0000 (10:18 +0000)]
Revert "Revert r366458, r366467 and r366468"

This reverts commit 9c377105da0be7c2c9a3c70035ce674c71b846af.

[clangd][BackgroundIndexLoader] Directly store DependentTU while loading shard

Summary:
We were deferring the population of DependentTU field in LoadedShard
until BackgroundIndexLoader was consumed. This actually triggers a use after
free since the shards FileToTU was pointing at could've been moved while
consuming the Loader.

Reviewers: sammccall

Subscribers: ilya-biryukov, MaskRay, jkorous, arphaman, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D64980

llvm-svn: 366559

5 years ago[llvm-readelf] - A fix for: "--hash-symbols asserts for 64-bit ELFs"
George Rimar [Fri, 19 Jul 2019 10:15:03 +0000 (10:15 +0000)]
[llvm-readelf] - A fix for: "--hash-symbols asserts for 64-bit ELFs"

Fixes https://bugs.llvm.org/show_bug.cgi?id=42622.
(--hash-symbols switch is currently broken for 64-bit ELF files, due to r352630.)

Differential revision: https://reviews.llvm.org/D64788

llvm-svn: 366558

5 years ago[IPRA] Don't rely on non-exact function definitions
Oliver Stannard [Fri, 19 Jul 2019 09:59:26 +0000 (09:59 +0000)]
[IPRA] Don't rely on non-exact function definitions

If a function definition is not exact, then the linker could select a
differently-compiled version of it, which could use different registers.

https://reviews.llvm.org/D64909

llvm-svn: 366557

5 years ago[ARM] Add <saturate> operand to SQRSHRL and UQRSHLL
Mikhail Maltsev [Fri, 19 Jul 2019 09:46:28 +0000 (09:46 +0000)]
[ARM] Add <saturate> operand to SQRSHRL and UQRSHLL

Summary:
According to the new Armv8-M specification
https://static.docs.arm.com/ddi0553/bh/DDI0553B_h_armv8m_arm.pdf the
instructions SQRSHRL and UQRSHLL now have an additional immediate
operand <saturate>. The new assembly syntax is:

SQRSHRL<c> RdaLo, RdaHi, #<saturate>, Rm
UQRSHLL<c> RdaLo, RdaHi, #<saturate>, Rm

where <saturate> can be either 64 (the existing behavior) or 48, in
that case the result is saturated to 48 bits.

The new operand is encoded as follows:
  #64 Encoded as sat = 0
  #48 Encoded as sat = 1
sat is bit 7 of the instruction bit pattern.

This patch adds a new assembler operand class MveSaturateOperand which
implements parsing and encoding. Decoding is implemented in
DecodeMVEOverlappingLongShift.

Reviewers: ostannard, simon_tatham, t.p.northover, samparker, dmgreen, SjoerdMeijer

Reviewed By: simon_tatham

Subscribers: javed.absar, kristof.beyls, hiraditya, pbarrio, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64810

llvm-svn: 366555

5 years agoRevert r366458, r366467 and r366468
Azharuddin Mohammed [Fri, 19 Jul 2019 09:26:33 +0000 (09:26 +0000)]
Revert r366458, r366467 and r366468

r366458 is causing test failures. r366467 and r366468 had to be reverted as
they were casuing conflict while reverting r366458.

r366468 [clangd] Remove dead code from BackgroundIndex
r366467 [clangd] BackgroundIndex stores shards to the closest project
r366458 [clangd] Refactor background-index shard loading

llvm-svn: 366551

5 years ago[OpenCL] Define CLK_NULL_EVENT without cast
Sven van Haastregt [Fri, 19 Jul 2019 09:11:48 +0000 (09:11 +0000)]
[OpenCL] Define CLK_NULL_EVENT without cast

Defining CLK_NULL_EVENT with a `(void*)` cast has the (unintended?)
side-effect that the address space will be fixed (as generic in OpenCL
2.0 mode).  The consequence is that any target specific address space
for the clk_event_t type will not be applied.

It is not clear why the void pointer cast was needed in the first
place, and it seems we can do without it.

Differential Revision: https://reviews.llvm.org/D63876

llvm-svn: 366546

5 years ago[clangd] Handle windows line endings in QueryDriver
Kadir Cetinkaya [Fri, 19 Jul 2019 09:08:22 +0000 (09:08 +0000)]
[clangd] Handle windows line endings in QueryDriver

Summary:
The previous patch did not fix the end mark. D64789
fixes second case of https://github.com/clangd/clangd/issues/93

Patch by @lh123 !

Reviewers: sammccall, kadircet

Reviewed By: kadircet

Subscribers: MaskRay, ilya-biryukov, jkorous, arphaman, cfe-commits

Tags: #clang-tools-extra, #clang

Differential Revision: https://reviews.llvm.org/D64970

llvm-svn: 366545

5 years ago[sanitizers] Use covering ObjectFormatType switches
Hubert Tong [Fri, 19 Jul 2019 08:46:18 +0000 (08:46 +0000)]
[sanitizers] Use covering ObjectFormatType switches

Summary:
This patch removes the `default` case from some switches on
`llvm::Triple::ObjectFormatType`, and cases for the missing enumerators
(`UnknownObjectFormat`, `Wasm`, and `XCOFF`) are then added.

For `UnknownObjectFormat`, the effect of the action for the `default`
case is maintained; otherwise, where `llvm_unreachable` is called,
`report_fatal_error` is used instead.

Where the `default` case returns a default value, `report_fatal_error`
is used for XCOFF as a placeholder. For `Wasm`, the effect of the action
for the `default` case in maintained.

The code is structured to avoid strongly implying that the `Wasm` case
is present for any reason other than to make the switch cover all
`ObjectFormatType` enumerator values.

Reviewers: sfertile, jasonliu, daltenty

Reviewed By: sfertile

Subscribers: hiraditya, aheejin, sunfish, llvm-commits, cfe-commits

Tags: #clang, #llvm

Differential Revision: https://reviews.llvm.org/D64222

llvm-svn: 366544

5 years ago[AMDGPU] Simplify the exclusive scan used for optimized atomics
Jay Foad [Fri, 19 Jul 2019 08:40:37 +0000 (08:40 +0000)]
[AMDGPU] Simplify the exclusive scan used for optimized atomics

Summary:
Change the scan algorithm to use only power-of-two shifts (1, 2, 4, 8,
16, 32) instead of starting off shifting by 1, 2 and 3 and then doing
a 3-way ADD, because:

1. It simplifies the compiler a little.
2. It minimizes vgpr pressure because each instruction is now of the
   form vn = vn + vn << c.
3. It is more friendly to the DPP combiner, which currently can't
   combine into an ADD3 instruction.

Because of #2 and #3 the end result is improved from this:

  v_add_u32_dpp v4, v3, v3  row_shr:1 row_mask:0xf bank_mask:0xf bound_ctrl:0
  v_mov_b32_dpp v5, v3  row_shr:2 row_mask:0xf bank_mask:0xf
  v_mov_b32_dpp v1, v3  row_shr:3 row_mask:0xf bank_mask:0xf
  v_add3_u32 v1, v4, v5, v1
  s_nop 1
  v_add_u32_dpp v1, v1, v1  row_shr:4 row_mask:0xf bank_mask:0xe
  s_nop 1
  v_add_u32_dpp v1, v1, v1  row_shr:8 row_mask:0xf bank_mask:0xc
  s_nop 1
  v_add_u32_dpp v1, v1, v1  row_bcast:15 row_mask:0xa bank_mask:0xf
  s_nop 1
  v_add_u32_dpp v1, v1, v1  row_bcast:31 row_mask:0xc bank_mask:0xf

To this:

  v_add_u32_dpp v1, v1, v1  row_shr:1 row_mask:0xf bank_mask:0xf bound_ctrl:0
  s_nop 1
  v_add_u32_dpp v1, v1, v1  row_shr:2 row_mask:0xf bank_mask:0xf bound_ctrl:0
  s_nop 1
  v_add_u32_dpp v1, v1, v1  row_shr:4 row_mask:0xf bank_mask:0xe
  s_nop 1
  v_add_u32_dpp v1, v1, v1  row_shr:8 row_mask:0xf bank_mask:0xc
  s_nop 1
  v_add_u32_dpp v1, v1, v1  row_bcast:15 row_mask:0xa bank_mask:0xf
  s_nop 1
  v_add_u32_dpp v1, v1, v1  row_bcast:31 row_mask:0xc bank_mask:0xf

I.e. two fewer computational instructions, one extra nop where we could
schedule something else.

Reviewers: arsenm, sheredom, critson, rampitec, vpykhtin

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

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64411

llvm-svn: 366543

5 years ago[Loop Peeling] Enable peeling of multiple exits by default.
Serguei Katkov [Fri, 19 Jul 2019 08:35:45 +0000 (08:35 +0000)]
[Loop Peeling] Enable peeling of multiple exits by default.

Enable loop peeling with multiple exits where all non-latch exits
ends up with deopt by default.

Reviewers: reames, fhahn
Reviewed By: reames
Subscribers: xbolva00, hiraditya, zzheng, llvm-commits
Differential Revision: https://reviews.llvm.org/D64619

llvm-svn: 366542

5 years ago[clangd] cleanup: unify the implemenation of checking a location is inside main file.
Haojian Wu [Fri, 19 Jul 2019 08:33:39 +0000 (08:33 +0000)]
[clangd] cleanup: unify the implemenation of checking a location is inside main file.

Summary: We have variant implementations in the codebase, this patch unifies them.

Reviewers: ilya-biryukov, kadircet

Subscribers: MaskRay, jkorous, arphaman, cfe-commits

Tags: #clang

Differential Revision: https://reviews.llvm.org/D64915

llvm-svn: 366541

5 years ago[InstCombine] Dropping redundant masking before left-shift [5/5] (PR42563)
Roman Lebedev [Fri, 19 Jul 2019 08:26:58 +0000 (08:26 +0000)]
[InstCombine] Dropping redundant masking before left-shift [5/5] (PR42563)

Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.

There are many variants to this pattern:
f. `((x << MaskShAmt) a>> MaskShAmt) << ShiftShAmt`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
f. `(ShiftShAmt-MaskShAmt) s>= 0` (i.e. `ShiftShAmt u>= MaskShAmt`)

Normally, the inner pattern is sign-extend,
but for our purposes it's no different to other patterns:

alive proofs:
f: https://rise4fun.com/Alive/7U3

For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.

https://bugs.llvm.org/show_bug.cgi?id=42563

Differential Revision: https://reviews.llvm.org/D64524

llvm-svn: 366540

5 years ago[InstCombine] Dropping redundant masking before left-shift [4/5] (PR42563)
Roman Lebedev [Fri, 19 Jul 2019 08:26:47 +0000 (08:26 +0000)]
[InstCombine] Dropping redundant masking before left-shift [4/5] (PR42563)

Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.

There are many variants to this pattern:
e. `((x << MaskShAmt) l>> MaskShAmt) << ShiftShAmt`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
e. `(ShiftShAmt-MaskShAmt) s>= 0` (i.e. `ShiftShAmt u>= MaskShAmt`)

alive proofs:
e: https://rise4fun.com/Alive/0FT

For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.

https://bugs.llvm.org/show_bug.cgi?id=42563

Differential Revision: https://reviews.llvm.org/D64521

llvm-svn: 366539

5 years ago[InstCombine] Dropping redundant masking before left-shift [3/5] (PR42563)
Roman Lebedev [Fri, 19 Jul 2019 08:26:37 +0000 (08:26 +0000)]
[InstCombine] Dropping redundant masking before left-shift [3/5] (PR42563)

Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.

There are many variants to this pattern:
d. `(x & ((-1 << MaskShAmt) >> MaskShAmt)) << ShiftShAmt`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
d. `(ShiftShAmt-MaskShAmt) s>= 0` (i.e. `ShiftShAmt u>= MaskShAmt`)

alive proofs:
d: https://rise4fun.com/Alive/I5Y

For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.

https://bugs.llvm.org/show_bug.cgi?id=42563

Differential Revision: https://reviews.llvm.org/D64519

llvm-svn: 366538

5 years ago[InstCombine] Dropping redundant masking before left-shift [2/5] (PR42563)
Roman Lebedev [Fri, 19 Jul 2019 08:26:25 +0000 (08:26 +0000)]
[InstCombine] Dropping redundant masking before left-shift [2/5] (PR42563)

Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.

There are many variants to this pattern:
c. `(x & (-1 >> MaskShAmt)) << ShiftShAmt`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
c. `(ShiftShAmt-MaskShAmt) s>= 0` (i.e. `ShiftShAmt u>= MaskShAmt`)

alive proofs:
c: https://rise4fun.com/Alive/RgJh

For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.

https://bugs.llvm.org/show_bug.cgi?id=42563

Differential Revision: https://reviews.llvm.org/D64517

llvm-svn: 366537

5 years ago[InstCombine] Dropping redundant masking before left-shift [1/5] (PR42563)
Roman Lebedev [Fri, 19 Jul 2019 08:26:13 +0000 (08:26 +0000)]
[InstCombine] Dropping redundant masking before left-shift [1/5] (PR42563)

Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.

There are many variants to this pattern:
b. `(x & (~(-1 << maskNbits))) << shiftNbits`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
b. `(MaskShAmt+ShiftShAmt) u>= bitwidth(x)`

alive proof:
b: https://rise4fun.com/Alive/y8M

For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.

https://bugs.llvm.org/show_bug.cgi?id=42563

Differential Revision: https://reviews.llvm.org/D64514

llvm-svn: 366536

5 years ago[InstCombine] Dropping redundant masking before left-shift [0/5] (PR42563)
Roman Lebedev [Fri, 19 Jul 2019 08:25:43 +0000 (08:25 +0000)]
[InstCombine] Dropping redundant masking before left-shift [0/5] (PR42563)

Summary:
If we have some pattern that leaves only some low bits set, and then performs
left-shift of those bits, if none of the bits that are left after the final
shift are modified by the mask, we can omit the mask.

There are many variants to this pattern:
a. `(x & ((1 << MaskShAmt) - 1)) << ShiftShAmt`
All these patterns can be simplified to just:
`x << ShiftShAmt`
iff:
a. `(MaskShAmt+ShiftShAmt) u>= bitwidth(x)`

alive proof:
a: https://rise4fun.com/Alive/wi9

Indeed, not all of these patterns are canonical.
But since this fold will only produce a single instruction
i'm really interested in handling even uncanonical patterns,
since i have this general kind of pattern in hotpaths,
and it is not totally outlandish for bit-twiddling code.

For now let's start with patterns where both shift amounts are variable,
with trivial constant "offset" between them, since i believe this is
both simplest to handle and i think this is most common.
But again, there are likely other variants where we could use
ValueTracking/ConstantRange to handle more cases.

https://bugs.llvm.org/show_bug.cgi?id=42563

Reviewers: spatel, nikic, huihuiz, xbolva00

Reviewed By: xbolva00

Subscribers: efriedma, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64512

llvm-svn: 366535

5 years ago[ELF][test] Fix aarch64-condb-reloc.s
Fangrui Song [Fri, 19 Jul 2019 08:00:22 +0000 (08:00 +0000)]
[ELF][test] Fix aarch64-condb-reloc.s

llvm-svn: 366534

5 years ago[NFC] Fix an indentation issue in llvm/Support/TargetRegistry.h
Hubert Tong [Fri, 19 Jul 2019 07:21:59 +0000 (07:21 +0000)]
[NFC] Fix an indentation issue in llvm/Support/TargetRegistry.h

llvm-svn: 366533

5 years ago[ELF][AArch64] Improve some aarch64-*.s tests
Fangrui Song [Fri, 19 Jul 2019 06:33:36 +0000 (06:33 +0000)]
[ELF][AArch64] Improve some aarch64-*.s tests

* Delete aarch64-tls-static.s: it is covered by aarch64-tlsdesc.c
* Add --no-show-raw-insn to llvm-objdump -d tests
* When linking an executable with %t.so, the path %t.so will be recorded in the DT_NEEDED entry if %t.so doesn't have DT_SONAME. The DT_NEEDED has varying lengths on different systems.
  Add -soname to make tests more robust. This issue will become outstanding if we allow overlapping PT_LOAD (D64930).

llvm-svn: 366532

5 years ago[DebugInfo] Some fields do not need relocations even relax is enabled.
Hsiangkai Wang [Fri, 19 Jul 2019 06:10:36 +0000 (06:10 +0000)]
[DebugInfo] Some fields do not need relocations even relax is enabled.

In debug frame information, some fields, e.g., Length in CIE/FDE and
Offset in FDE are attributes to describe the structure of CIE/FDE. They
are not related to the relaxed code. However, these attributes are
symbol differences. So, in current design, these attributes will be
filled as zero and LLVM generates relocations for them.

We only need to generate relocations for symbols in executable sections.
So, if the symbols are not located in executable sections, we still
evaluate their values under relaxation.

Differential Revision: https://reviews.llvm.org/D61584

llvm-svn: 366531

5 years agounbreak links
Chris Lattner [Fri, 19 Jul 2019 05:49:11 +0000 (05:49 +0000)]
unbreak links

llvm-svn: 366530

5 years agoreplace the old kaleidoscope tutorial files with orphaned pages that forward to the...
Chris Lattner [Fri, 19 Jul 2019 05:23:17 +0000 (05:23 +0000)]
replace the old kaleidoscope tutorial files with orphaned pages that forward to the new copy.

llvm-svn: 366529

5 years agoPoint to the dusted off version of the kaleidoscope tutorial.
Chris Lattner [Fri, 19 Jul 2019 05:15:57 +0000 (05:15 +0000)]
Point to the dusted off version of the kaleidoscope tutorial.

llvm-svn: 366528

5 years ago[test] [llvm-objcopy] Fix broken test case
Alex Brachet [Fri, 19 Jul 2019 02:31:21 +0000 (02:31 +0000)]
[test] [llvm-objcopy] Fix broken test case

Summary: The test case added in D62718 did not work unless the user was root because write bits were not set for the output file. This change uses only permissions with user write (0200) to ensure tests pass regardless of the users permissions.

Reviewers: jhenderson, rupprecht, MaskRay, espindola, alexshap

Reviewed By: MaskRay

Subscribers: emaste, arichardson, jakehehrlich, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D64302

llvm-svn: 366527