platform/upstream/llvm.git
4 years ago[mlir] check for failures when packing function sigunatures in std->llvm conversion
Alex Zinenko [Mon, 14 Sep 2020 11:47:27 +0000 (13:47 +0200)]
[mlir] check for failures when packing function sigunatures in std->llvm conversion

When packing function results into a structure during the standard-to-llvm
dialect conversion, do not assume the conversion was successful and propagate
nullptr as error state.

Fixes PR45184.

Reviewed By: nicolasvasilache

Differential Revision: https://reviews.llvm.org/D87605

4 years ago[mlir][integration_test] Linalg Conv folder renamed to CPU
Jakub Lichman [Tue, 15 Sep 2020 10:07:29 +0000 (10:07 +0000)]
[mlir][integration_test] Linalg Conv folder renamed to CPU

Changing directory name to reflect naming convention discussed here:
https://llvm.discourse.group/t/vectorops-rfc-add-suite-of-integration-tests-for-vector-dialect-operations/1213

Differential Revision: https://reviews.llvm.org/D87678

4 years agoRevert "[InstCombine] Simplify select operand based on equality condition"
Benjamin Kramer [Tue, 15 Sep 2020 10:22:47 +0000 (12:22 +0200)]
Revert "[InstCombine] Simplify select operand based on equality condition"

This reverts commit cfff88c03cf9e9b72906a41fd11e06721d54f293. Sends
instcombine into an infinite loop.

```
define i1 @foo(i32 %arg, i32 %arg1) {
bb:
  %tmp = udiv i32 %arg, %arg1
  %tmp2 = mul nsw i32 %tmp, %arg1
  %tmp3 = icmp eq i32 %tmp2, %arg
  %tmp4 = select i1 %tmp3, i32 %tmp, i32 undef
  %tmp5 = icmp sgt i32 %tmp4, 255
  ret i1 %tmp5
}
```

4 years ago[X86] Update SSE/AVX integer MINMAX intrinsics to emit llvm.smax.* etc. (PR46851)
Simon Pilgrim [Tue, 15 Sep 2020 10:18:44 +0000 (11:18 +0100)]
[X86] Update SSE/AVX integer MINMAX intrinsics to emit llvm.smax.* etc. (PR46851)

We're now getting close to having the necessary analysis/combines etc. for the new generic llvm smax/smin/umax/umin intrinsics.

This patch updates the SSE/AVX integer MINMAX intrinsics to emit the generic equivalents instead of the icmp+select code pattern.

Differential Revision: https://reviews.llvm.org/D87603

4 years ago[ARM] Corrected condition in isSaturatingConditional
Meera Nakrani [Tue, 15 Sep 2020 10:14:30 +0000 (10:14 +0000)]
[ARM] Corrected condition in isSaturatingConditional

Fixed a small error in an if condition to prevent usat/ssat being generated if (upper constant + 1) is not a
power of 2.

4 years ago[SelectionDAG] Remove unused FP constant in getNegatedExpression
Qiu Chaofan [Tue, 15 Sep 2020 09:59:10 +0000 (17:59 +0800)]
[SelectionDAG] Remove unused FP constant in getNegatedExpression

960cbc53 immediately removes nodes that won't be used to avoid
compilation time explosion. This patch adds the removal to constants to
fix PR47517.

Reviewed By: RKSimon, steven.zhang

Differential Revision: https://reviews.llvm.org/D87614

4 years ago[Transforms][Coroutines] Add missing header path to CMakeLists.txt
Simon Pilgrim [Tue, 15 Sep 2020 09:37:25 +0000 (10:37 +0100)]
[Transforms][Coroutines] Add missing header path to CMakeLists.txt

Helps Visual Studio check include dependencies.

4 years ago[MVE] fix typo in llvm debug message. NFC.
Sjoerd Meijer [Tue, 15 Sep 2020 09:04:02 +0000 (10:04 +0100)]
[MVE] fix typo in llvm debug message. NFC.

4 years ago[X86] detectAVGPattern - accept non-pow2 vectors by padding.
Simon Pilgrim [Tue, 15 Sep 2020 09:06:35 +0000 (10:06 +0100)]
[X86] detectAVGPattern - accept non-pow2 vectors by padding.

Drop the pow2 vector limitation for AVG generation by padding the vector to the next pow2, creating the PAVG nodes and then extracting the final subvector.

Fixes some poor codegen that has been annoying me for years.....

4 years ago[llvm-readobj/elf] - Simplify and refine the implementation which dumps .stack_sizes
Georgii Rymar [Sat, 5 Sep 2020 16:13:50 +0000 (19:13 +0300)]
[llvm-readobj/elf] - Simplify and refine the implementation which dumps .stack_sizes

Our implementation of stack sizes section dumping heavily uses `ELFObjectFile<ELFT>`,
while the rest of the code uses `ELFFile<ELFT>`.

That APIs are very different. `ELFObjectFile<ELFT>` is very generic
and has `SectionRef`, `RelocationRef`, `SymbolRef` and other generic concepts.
The `ELFFile<ELFT>` class works directly with `Elf_Shdr`, `Elf_Rel[a]`, `Elf_Sym` etc,
what is probably much cleaner for ELF dumper.

Also, `ELFObjectFile<ELFT>` API does not always provide a way to check
for possible errors. E.g. the implementation of `symbol_end()` does not verify the `sh_size`:

```
template <class ELFT>
basic_symbol_iterator ELFObjectFile<ELFT>::symbol_end() const {
  const Elf_Shdr *SymTab = DotSymtabSec;
  if (!SymTab)
    return symbol_begin();
  DataRefImpl Sym = toDRI(SymTab, SymTab->sh_size / sizeof(Elf_Sym));
  return basic_symbol_iterator(SymbolRef(Sym, this));
}
```
There are many other examples which makes me thing we might win from
switching to `ELFFile<ELFT>` API, where we heavily validate an input data already.

This patch is the first step in this direction. I've converted the large portion of the code
to use `ELFFile<ELFT>`.

Differential revision: https://reviews.llvm.org/D87362

4 years ago[NFC][DebugInfo] Use consistent regex group spelling
Petr Hosek [Tue, 15 Sep 2020 08:46:58 +0000 (01:46 -0700)]
[NFC][DebugInfo] Use consistent regex group spelling

This is a follow up to c1f2fb5184ca.

4 years ago[lib/Object] - Refine interface of ELFFile<ELFT>. NFCI.
Georgii Rymar [Wed, 9 Sep 2020 14:03:53 +0000 (17:03 +0300)]
[lib/Object] - Refine interface of ELFFile<ELFT>. NFCI.

`ELFFile<ELFT>` has many methods that take pointers,
though they assume that arguments are never null and
hence could take references instead.

This patch performs such clean-up.

Differential revision: https://reviews.llvm.org/D87385

4 years agoGlobalISel/IRTranslator resetTargetOptions based on function attributes
Petar Avramovic [Tue, 15 Sep 2020 08:25:38 +0000 (10:25 +0200)]
GlobalISel/IRTranslator resetTargetOptions based on function attributes

Update TargetMachine.Options with function attributes before we start
to generate MIR instructions. This allows access to correct function
attributes via TargetMachine.Options (it used to access attributes of
the function that was translated first).
This affects some existing tests with "no-nans-fp-math" attribute.
Follow-up on D87456.

Differential Revision: https://reviews.llvm.org/D87511

4 years ago[MVE] Rename of tests making them consistent with tail-predication tests. NFC.
Sjoerd Meijer [Tue, 15 Sep 2020 08:09:59 +0000 (09:09 +0100)]
[MVE] Rename of tests making them consistent with tail-predication tests. NFC.

4 years ago[Support] Make building with snmalloc work
Hans Wennborg [Thu, 10 Sep 2020 17:32:45 +0000 (19:32 +0200)]
[Support] Make building with snmalloc work

Differential revision: https://reviews.llvm.org/D87471

4 years ago[DebugInfo] Support both forward and backward slashes in tests
Petr Hosek [Tue, 15 Sep 2020 07:48:12 +0000 (00:48 -0700)]
[DebugInfo] Support both forward and backward slashes in tests

This addresses test failure revealed by 042c23506869.

4 years ago[doc] Fix broken link
Han Seoul-Oh [Tue, 15 Sep 2020 04:15:16 +0000 (21:15 -0700)]
[doc] Fix broken link

4 years ago[SVE] Fix isLoadInvariantInLoop for scalable vectors
David Sherwood [Mon, 7 Sep 2020 08:17:10 +0000 (09:17 +0100)]
[SVE] Fix isLoadInvariantInLoop for scalable vectors

I've amended the isLoadInvariantInLoop function to bail out for
scalable vectors for now since the invariant.start intrinsic is only
ever generated by the clang frontend for thread locals or struct
and class constructors, neither of which support sizeless types.
In addition, the intrinsic itself does not currently support the
concept of a scaled size, which makes it impossible to compare
the sizes of different scalable objects, e.g. <vscale x 32 x i8>
and <vscale x 16 x i8>.

Added new tests here:

  Transforms/LICM/AArch64/sve-load-hoist.ll
  Transforms/LICM/hoisting.ll

Differential Revision: https://reviews.llvm.org/D87227

4 years ago[NFC][Asan] Remove chunk pointer from metadata
Vitaly Buka [Tue, 15 Sep 2020 07:22:10 +0000 (00:22 -0700)]
[NFC][Asan] Remove chunk pointer from metadata

kAllocBegMagic should be enough.
kAllocBegMagic is already set for the Secondary allocations.
kAllocBegMagic is good enough for the Primary, but it's even safer for
the Secondary allocator as all allocated block are from mmap.

Depends on D87646.

Reviewed By: morehouse

Differential Revision: https://reviews.llvm.org/D87647

4 years ago[NFC] Test commit
Chris Hamilton [Tue, 15 Sep 2020 07:19:02 +0000 (02:19 -0500)]
[NFC] Test commit

4 years ago[NFC][Asan] Return uptr as before D87646
Vitaly Buka [Tue, 15 Sep 2020 07:16:55 +0000 (00:16 -0700)]
[NFC][Asan] Return uptr as before D87646

4 years ago[Asan] Cleanup kAllocBegMagic setup
Vitaly Buka [Tue, 15 Sep 2020 07:12:02 +0000 (00:12 -0700)]
[Asan] Cleanup kAllocBegMagic setup

Make it atomic.
Wrap it into class.
Set it late after chunk is initialized.
Reset it soon when the chunk is still valid.

Depends on D87645.

Reviewed By: morehouse

Differential Revision: https://reviews.llvm.org/D87646

4 years ago[NFC][Asan] Reorder bitfields
Vitaly Buka [Tue, 15 Sep 2020 07:07:56 +0000 (00:07 -0700)]
[NFC][Asan] Reorder bitfields

Depends on D87644.

Reviewed By: morehouse

Differential Revision: https://reviews.llvm.org/D87645

4 years agoRevert "[Sema] Address-space sensitive index check for unbounded arrays"
Chris Hamilton [Tue, 15 Sep 2020 06:54:41 +0000 (01:54 -0500)]
Revert "[Sema] Address-space sensitive index check for unbounded arrays"

This reverts commit da55e9ba1273284f1af61bceeaeb25e487838034.

Build bots uncovered coverage gap in testing.  Change not ready.

4 years ago[NFC][Asan] Remove from_memalign and rz_log
Vitaly Buka [Tue, 15 Sep 2020 06:54:48 +0000 (23:54 -0700)]
[NFC][Asan] Remove from_memalign and rz_log

Before D87643 they where used to optimize UsedSize(). Which was
called frequently from leak scanner.
It was also used for calls from QuarantineCallback
but we have heavy get_allocator().Deallocate call there anyway.

Depends on D87643.

Reviewed By: morehouse

Differential Revision: https://reviews.llvm.org/D87644

4 years ago[NFC][Asan] Don't use MetaData for size
Vitaly Buka [Tue, 15 Sep 2020 06:53:58 +0000 (23:53 -0700)]
[NFC][Asan] Don't use MetaData for size

Now we have enough space in the ChunkHeader.
45 bit is enough for kMaxAllowedMallocSize.

Depends on D87642.

Reviewed By: morehouse

Differential Revision: https://reviews.llvm.org/D87643

4 years ago[X86] Pre-commit test cases for D87593
Craig Topper [Tue, 15 Sep 2020 05:41:39 +0000 (22:41 -0700)]
[X86] Pre-commit test cases for D87593

The memory operand for these is incorrect.

4 years ago[NFC][lsan][fuzzer] Relax fuzzer-leak.test
Vitaly Buka [Tue, 15 Sep 2020 02:44:27 +0000 (19:44 -0700)]
[NFC][lsan][fuzzer] Relax fuzzer-leak.test

With lsan we can't guarantee to catch leak on the same iteration.

4 years ago[Windows][Polly] Disable LLVMPolly module for all compilers on Windows
Mateusz Mikuła [Tue, 15 Sep 2020 06:12:02 +0000 (09:12 +0300)]
[Windows][Polly] Disable LLVMPolly module for all compilers on Windows

Before this patch, the cmake disabled loadable modules when compiling
with Visual Studio. However, the reason for this is a limitation of the
Windows DLLs, thus this restriction should apply to any compiler for the
Windows platform, such as MinGW, Cygwin, icc, etc.

Differential Revision: https://reviews.llvm.org/D87524

4 years ago[LLD] Allow configuring default ld.lld backend
Mateusz Mikuła [Tue, 15 Sep 2020 05:39:15 +0000 (08:39 +0300)]
[LLD] Allow configuring default ld.lld backend

The motivation for this is ld.lld --help targeting MinGW which
currently prints help for the ELF backend unless -m i386pe{,p} is
added. This confuses build systems that grep through linker help to
find supported flags.

This matches LD from Binutils which always prints help for MinGW
when configured to target it.

After this change, the backend can still be overridden to any
supported ELF/MinGW target by using correct -m <arch>.

Differential Revision: https://reviews.llvm.org/D87418

4 years ago[llvm-readobj] [ARMWinEH] Print ARM64 packed unwind info
Martin Storsjö [Fri, 4 Sep 2020 20:42:22 +0000 (23:42 +0300)]
[llvm-readobj] [ARMWinEH] Print ARM64 packed unwind info

In addition to printing the individual fields, synthesize and
print the corresponding prolog for the unwind info (in reverse
order, to match how it's printed for non-packed unwind info).

Differential Revision: https://reviews.llvm.org/D87370

4 years ago[BinaryFormat/MachO] Add a missing constant.
Davide Italiano [Tue, 15 Sep 2020 05:29:53 +0000 (22:29 -0700)]
[BinaryFormat/MachO] Add a missing constant.

Reference:
https://opensource.apple.com/source/cctools/cctools-949.0.1/include/mach-o/loader.h.auto.html

4 years ago[DebugInfo] Make offsets of dwarf units 64-bit (19/19).
Igor Kudrin [Tue, 15 Sep 2020 04:32:08 +0000 (11:32 +0700)]
[DebugInfo] Make offsets of dwarf units 64-bit (19/19).

In the case of LTO, several DWARF units can be emitted in one section.
For an extremely large application, they may exceed the limit of 4GiB
for 32-bit offsets. As it is now possible to emit 64-bit debugging info,
the patch enables storing the larger offsets.

Differential Revision: https://reviews.llvm.org/D87026

4 years ago[DebugInfo] Make the offset of string pool entries 64-bit (18/19).
Igor Kudrin [Tue, 15 Sep 2020 04:32:01 +0000 (11:32 +0700)]
[DebugInfo] Make the offset of string pool entries 64-bit (18/19).

The string pool is shared among several units in the case of LTO,
and it potentially can exceed the limit of 4GiB for an extremely
large application. As it is now possible to emit 64-bit debugging
info, the limitation can be removed.

Differential Revision: https://reviews.llvm.org/D87025

4 years ago[DebugInfo] Fix emitting DWARF64 .debug_macro[.dwo] sections (17/19).
Igor Kudrin [Tue, 15 Sep 2020 04:31:55 +0000 (11:31 +0700)]
[DebugInfo] Fix emitting DWARF64 .debug_macro[.dwo] sections (17/19).

The patch fixes emitting flags and the debug_line_offset field in
the header, as well as the reference to the macro string for
a pre-standard GNU .debug_macro extension.

Differential Revision: https://reviews.llvm.org/D87024

4 years ago[DebugInfo] Fix emitting DWARF64 .debug_names sections (16/19).
Igor Kudrin [Tue, 15 Sep 2020 04:31:49 +0000 (11:31 +0700)]
[DebugInfo] Fix emitting DWARF64 .debug_names sections (16/19).

The patch fixes emitting the unit length field in the header of
the table and offsets to the entry pool. Note that while the patch
changes the common method to emit offsets, in fact, nothing is changed
for Apple accelerator tables, because we do not yet support DWARF64 for
those targets.

Differential Revision: https://reviews.llvm.org/D87023

4 years ago[DebugInfo] Fix emitting DWARF64 .debug_addr sections (15/19).
Igor Kudrin [Tue, 15 Sep 2020 04:31:41 +0000 (11:31 +0700)]
[DebugInfo] Fix emitting DWARF64 .debug_addr sections (15/19).

The patch fixes emitting the header of the table. The content is
independent of the DWARF format.

Differential Revision: https://reviews.llvm.org/D87022

4 years ago[DebugInfo] Fix emitting DWARF64 .debug_loclists sections (14/19).
Igor Kudrin [Tue, 15 Sep 2020 04:31:34 +0000 (11:31 +0700)]
[DebugInfo] Fix emitting DWARF64 .debug_loclists sections (14/19).

The size of the offsets in the table depends on the DWARF format.

Differential Revision: https://reviews.llvm.org/D87020

4 years ago[DebugInfo] Fix emitting DWARF64 .debug_rnglists sections (13/19).
Igor Kudrin [Tue, 15 Sep 2020 04:31:28 +0000 (11:31 +0700)]
[DebugInfo] Fix emitting DWARF64 .debug_rnglists sections (13/19).

The size of the offsets in the table depends on the DWARF format.

Differential Revision: https://reviews.llvm.org/D87019

4 years ago[DebugInfo] Fix emitting pre-v5 name lookup tables in the DWARF64 format (12/19).
Igor Kudrin [Tue, 15 Sep 2020 04:31:20 +0000 (11:31 +0700)]
[DebugInfo] Fix emitting pre-v5 name lookup tables in the DWARF64 format (12/19).

The transition is done by using methods of AsmPrinter which
automatically emit values in compliance with the selected DWARF format.

Differential Revision: https://reviews.llvm.org/D87013

4 years ago[DebugInfo] Fix emitting DWARF64 .debug_aranges sections (11/19).
Igor Kudrin [Tue, 15 Sep 2020 04:31:14 +0000 (11:31 +0700)]
[DebugInfo] Fix emitting DWARF64 .debug_aranges sections (11/19).

The patch fixes calculating the size of the table and emitting
the fields which depend on the DWARF format by using methods that
choose appropriate sizes automatically.

Differential Revision: https://reviews.llvm.org/D87012

4 years ago[DebugInfo] Fix emitting DWARF64 type units (10/19).
Igor Kudrin [Tue, 15 Sep 2020 04:31:07 +0000 (11:31 +0700)]
[DebugInfo] Fix emitting DWARF64 type units (10/19).

The patch fixes emitting the offset to the type DIE. All other fields
are already fixed in previous patches.

Differential Revision: https://reviews.llvm.org/D87021

4 years ago[DebugInfo] Fix emitting DWARF64 DWO compilation units and string offset tables ...
Igor Kudrin [Tue, 15 Sep 2020 04:31:00 +0000 (11:31 +0700)]
[DebugInfo] Fix emitting DWARF64 DWO compilation units and string offset tables (9/19).

These two fixes are better to go together because llvm-dwarfdump is
unable to dump a table when another one is malformed.

Differential Revision: https://reviews.llvm.org/D87018

4 years ago[DebugInfo] Fix emitting DWARF64 .debug_str_offsets sections (8/19).
Igor Kudrin [Tue, 15 Sep 2020 04:30:53 +0000 (11:30 +0700)]
[DebugInfo] Fix emitting DWARF64 .debug_str_offsets sections (8/19).

The patch fixes calculating the size of the table and emitting the unit
length field.

Differential Revision: https://reviews.llvm.org/D87017

4 years ago[DebugInfo] Fix emitting the DW_AT_location attribute for 64-bit DWARFv3 (7/19).
Igor Kudrin [Tue, 15 Sep 2020 04:30:46 +0000 (11:30 +0700)]
[DebugInfo] Fix emitting the DW_AT_location attribute for 64-bit DWARFv3 (7/19).

The patch uses a common method to determine the appropriate form for
the value of the attribute.

Differential Revision: https://reviews.llvm.org/D87016

4 years ago[DebugInfo] Use a common method to determine a suitable form for section offsts ...
Igor Kudrin [Tue, 15 Sep 2020 04:30:38 +0000 (11:30 +0700)]
[DebugInfo] Use a common method to determine a suitable form for section offsts (6/19).

This is mostly an NFC patch because the involved methods are used when
emitting DWO files, which is incompatible with DWARFv3, or for platforms
where DWARF64 is not supported yet.

Differential Revision: https://reviews.llvm.org/D87015

4 years ago[DebugInfo] Fix emitting DWARF64 compilation units (5/19).
Igor Kudrin [Tue, 15 Sep 2020 04:30:30 +0000 (11:30 +0700)]
[DebugInfo] Fix emitting DWARF64 compilation units (5/19).

The patch also adds a method to choose an appropriate DWARF form
to represent section offsets according to the version and the format
of producing debug info.

Differential Revision: https://reviews.llvm.org/D87014

4 years ago[DebugInfo] Add the -dwarf64 switch to llc and other internal tools (4/19).
Igor Kudrin [Tue, 15 Sep 2020 04:30:18 +0000 (11:30 +0700)]
[DebugInfo] Add the -dwarf64 switch to llc and other internal tools (4/19).

The patch adds a switch to enable emitting debug info in the 64-bit
DWARF format. Most emitter for sections will be updated in the subsequent
patches, whereas for .debug_line and .debug_frame the emitters are in
the MC library, which is already updated.

For now, the switch is enabled only for 64-bit ELF targets.

Differential Revision: https://reviews.llvm.org/D87011

4 years ago[DebugInfo] Add new emitting methods for values which depend on the DWARF format...
Igor Kudrin [Tue, 15 Sep 2020 04:30:10 +0000 (11:30 +0700)]
[DebugInfo] Add new emitting methods for values which depend on the DWARF format (3/19).

These methods are going to be used in subsequent patches.

Differential Revision: https://reviews.llvm.org/D87010

4 years ago[DebugInfo] Fix DIE value emitters to be compatible with DWARF64 (2/19).
Igor Kudrin [Tue, 15 Sep 2020 04:30:02 +0000 (11:30 +0700)]
[DebugInfo] Fix DIE value emitters to be compatible with DWARF64 (2/19).

DW_FORM_sec_offset and DW_FORM_strp imply values of different sizes with
DWARF32 and DWARF64. The patch fixes DIE value classes to use correct
sizes when emitting their values. For DIELocList it ensures that the
requested DWARF form matches the current DWARF format because that class
uses a method that selects the size automatically.

Differential Revision: https://reviews.llvm.org/D87009

4 years ago[DebugInfo] Fix methods of AsmPrinter to emit values corresponding to the DWARF forma...
Igor Kudrin [Tue, 15 Sep 2020 04:29:48 +0000 (11:29 +0700)]
[DebugInfo] Fix methods of AsmPrinter to emit values corresponding to the DWARF format (1/19).

These methods are used to emit values which are 32-bit in DWARF32 and
64-bit in DWARF64. The patch fixes them so that they choose the length
automatically, depending on the DWARF format set in the Context.

Differential Revision: https://reviews.llvm.org/D87008

4 years ago[lld][WebAssembly] Allow globals imports via import_name/import_module
Sam Clegg [Tue, 15 Sep 2020 02:20:25 +0000 (19:20 -0700)]
[lld][WebAssembly] Allow globals imports via import_name/import_module

This feature already exists but was limited to function
symbols.

Differential Revision: https://reviews.llvm.org/D87666

4 years ago[DebugInfo] Remove dots from getFilenameByIndex return value
Petr Hosek [Tue, 15 Sep 2020 00:44:12 +0000 (17:44 -0700)]
[DebugInfo] Remove dots from getFilenameByIndex return value

When concatenating directory with filename in getFilenameByIndex, we
might end up with a path that contains extra dots. For example, if the
input is /path and ./example, we would return /path/./example. Run
sys::path::remove_dots on the output to eliminate unnecessary dots.

Differential Revision: https://reviews.llvm.org/D87657

4 years ago[Coroutines] Fix a typo in documentation
Xun Li [Tue, 15 Sep 2020 01:56:31 +0000 (18:56 -0700)]
[Coroutines] Fix a typo in documentation

In the example, the variable that's crossing suspend point was referred wrongly, fix it.

Differential Revision: https://reviews.llvm.org/D83563

4 years ago[GlobalISel] Add a `X, Y = G_UNMERGE(G_ZEXT Z)` -> X = G_ZEXT Z; Y = 0 combine
Quentin Colombet [Thu, 10 Sep 2020 01:03:00 +0000 (18:03 -0700)]
[GlobalISel] Add a `X, Y = G_UNMERGE(G_ZEXT Z)` -> X = G_ZEXT Z; Y = 0 combine

Add a combiner helper to transform unmerge of zext into one zext and
a constant 0

Differential Revision: https://reviews.llvm.org/D87427

4 years ago[GlobalISel] Add `X,Y<dead> = G_UNMERGE Z` -> X = G_TRUNC Z
Quentin Colombet [Sat, 5 Sep 2020 00:09:38 +0000 (17:09 -0700)]
[GlobalISel] Add `X,Y<dead> = G_UNMERGE Z` -> X = G_TRUNC Z

Add a combiner helper that replaces G_UNMERGE where all the destination lanes
are dead except the first one with a G_TRUNC.

Differential Revision: https://reviews.llvm.org/D87174

4 years ago[libc++abi] Do not declare __cxa_finalize and __cxa_atexit in <cxxabi.h>
Louis Dionne [Mon, 14 Sep 2020 21:16:46 +0000 (17:16 -0400)]
[libc++abi] Do not declare __cxa_finalize and __cxa_atexit in <cxxabi.h>

These functions are not defined by libc++abi, so they don't belong in <cxxabi.h>.

Differential Revision: https://reviews.llvm.org/D75795

4 years ago[flang] Respect BZ mode in exponent parts, too
peter klausler [Mon, 14 Sep 2020 23:11:45 +0000 (16:11 -0700)]
[flang] Respect BZ mode in exponent parts, too

The Fortran standard discusses BZ mode (treat blanks as zero digits)
explicitly in its effect on the editing of the digits prior to the
exponent part, but doesn't mention it in description of the
exponent part.  Other compilers honor BZ mode in the exponent,
so we should do so too.  So "1 e 1 " is 1.E11 in BZ mode.

Differential Revision: https://reviews.llvm.org/D87653

4 years ago[X86] Place new constant node in topological order in X86DAGToDAGISel::matchBitExtract
Craig Topper [Mon, 14 Sep 2020 23:33:23 +0000 (16:33 -0700)]
[X86] Place new constant node in topological order in X86DAGToDAGISel::matchBitExtract

Fixes PR47482

4 years agoRevert "[X86] Place new constant node in topological order in X86DAGToDAGISel::matchB...
Craig Topper [Mon, 14 Sep 2020 23:33:08 +0000 (16:33 -0700)]
Revert "[X86] Place new constant node in topological order in X86DAGToDAGISel::matchBitExtract."

I got the bug number wrong.

This reverts commit 32515938901685bcbc438d5f5bb03cb8a9f4c637.

4 years ago[flang] Allow Fortran comments after #include path
peter klausler [Mon, 14 Sep 2020 21:09:01 +0000 (14:09 -0700)]
[flang] Allow Fortran comments after #include path

C-style /*comments*/ are removed during preprocessing directive
tokenization, but Fortran !comments need to be specifically
allowed.

Fixes LLVM bugzilla 47466.

Differential Revision: https://reviews.llvm.org/D87638

4 years ago[AArch64] Statepoint support for AArch64.
Philip Reames [Mon, 14 Sep 2020 23:38:48 +0000 (16:38 -0700)]
[AArch64] Statepoint support for AArch64.

Differential Revision: https://reviews.llvm.org/D66012
Patch By: loicottet (with major rebase by me)

4 years ago[Asan] Accept __lsan_ignore_object for redzone pointer
Vitaly Buka [Mon, 14 Sep 2020 23:32:25 +0000 (16:32 -0700)]
[Asan] Accept __lsan_ignore_object for redzone pointer

The check that the pointer inside of the user part of the chunk does not
adds any value, but it's the last user of AddrIsInside.

I'd like to simplify AsanChunk in followup patches.

Reviewed By: morehouse

Differential Revision: https://reviews.llvm.org/D87642

4 years ago[GlobalISel] Add G_UNMERGE(Cst) -> Cst1, Cst2, ... combine
Quentin Colombet [Fri, 4 Sep 2020 21:27:42 +0000 (14:27 -0700)]
[GlobalISel] Add G_UNMERGE(Cst) -> Cst1, Cst2, ... combine

Add a combiner helper that replaces G_UNMERGE of big constants into direct
use of smaller constants.

Differential Revision: https://reviews.llvm.org/D87166

4 years ago[X86] Place new constant node in topological order in X86DAGToDAGISel::matchBitExtract.
Craig Topper [Mon, 14 Sep 2020 23:28:11 +0000 (16:28 -0700)]
[X86] Place new constant node in topological order in X86DAGToDAGISel::matchBitExtract.

Fixes PR47525

4 years ago[Sema] Address-space sensitive index check for unbounded arrays
Chris Hamilton [Mon, 14 Sep 2020 23:12:12 +0000 (18:12 -0500)]
[Sema] Address-space sensitive index check for unbounded arrays

Check applied to unbounded (incomplete) arrays and pointers
to spot cases where the computed address is beyond the
largest possible addressable extent of the array, based
on the address space in which the array is delcared, or
which the pointer refers to.

Check helps to avoid cases of nonsense pointer math and
array indexing which could lead to linker failures or
runtime exceptions.  Of particular interest when building
for embedded systems with small address spaces.

Reviewed By: aaron.ballman

Differential Revision: https://reviews.llvm.org/D86796

4 years ago[Hexagon] Add more detailed testcase for widening truncates
Krzysztof Parzyszek [Mon, 14 Sep 2020 21:37:41 +0000 (16:37 -0500)]
[Hexagon] Add more detailed testcase for widening truncates

4 years ago[Hexagon] Widen loads and handle any-/sign-/zero-extensions
Krzysztof Parzyszek [Mon, 14 Sep 2020 19:04:54 +0000 (14:04 -0500)]
[Hexagon] Widen loads and handle any-/sign-/zero-extensions

4 years ago[Hexagon] Some HVX DAG combines
Krzysztof Parzyszek [Mon, 14 Sep 2020 21:06:10 +0000 (16:06 -0500)]
[Hexagon] Some HVX DAG combines

1. VINSERTW0 x, undef -> x
2. VROR (VROR x, a), b) -> VROR x, a+b

4 years agoReland [docs][NewPM] Add docs for writing NPM passes
Arthur Eubanks [Tue, 1 Sep 2020 01:36:11 +0000 (18:36 -0700)]
Reland [docs][NewPM] Add docs for writing NPM passes

As to not conflict with the legacy PM example passes under
llvm/lib/Transforms/Hello, this is under HelloNew. This makes the
CMakeLists.txt and general directory structure less confusing for people
following the example.

Much of the doc structure was taken from WritinAnLLVMPass.rst.

This adds a HelloWorld pass which simply prints out each function name.

More will follow after this, e.g. passes over different units of IR, analyses.
https://llvm.org/docs/WritingAnLLVMPass.html contains a lot more.

Relanded with missing "Support" dependency in LLVMBuild.txt.

Reviewed By: ychen, asbirlea

Differential Revision: https://reviews.llvm.org/D86979

4 years ago[GISel]: Add combine for G_FABS to G_FABS
Aditya Nandakumar [Mon, 14 Sep 2020 22:43:52 +0000 (15:43 -0700)]
[GISel]: Add combine for G_FABS to G_FABS

https://reviews.llvm.org/D87554

Patch adds one new GICombinerRule for G_FABS. The combine rule folds G_FABS(G_FABS(X)) to G_FABS(X).
Patch additionally adds new combiner tests for the AArch64 target to test this new combiner rule.

Patch by mkitzan.

4 years agoRevert "[docs][NewPM] Add docs for writing NPM passes"
Arthur Eubanks [Mon, 14 Sep 2020 22:55:17 +0000 (15:55 -0700)]
Revert "[docs][NewPM] Add docs for writing NPM passes"

This reverts commit c2590de30df23ef0db39b496cdec62a83a61fbfa.

Breaks shared libs build

4 years ago[GlobalISel] Add G_UNMERGE_VALUES(G_MERGE_VALUES) combine
Quentin Colombet [Thu, 3 Sep 2020 23:06:14 +0000 (16:06 -0700)]
[GlobalISel] Add G_UNMERGE_VALUES(G_MERGE_VALUES) combine

Add the matching and applying function to the combiner helper for
G_UNMERGE_VALUES(G_MERGE_VALUES).

This combine also supports any merge-like input nodes, like G_BUILD_VECTORS
and is robust against bitcasts in between int unmerge and merge nodes.

When the input type of the merge node and the output type of the unmerge
node are not the same, but the sizes are, the combine still applies but
creates bitcasts between the sources and the destinations instead of
reusing the destinations directly.

Long term, the artifact combiner should probably reuse that helper, but
as of today, it doesn't use any outside helper, so I kept it this way.

Differential Revision: https://reviews.llvm.org/D87117

4 years ago[flang] Fix bug for forward referenced type
Peter Steinfeld [Fri, 11 Sep 2020 18:02:04 +0000 (11:02 -0700)]
[flang] Fix bug for forward referenced type

A type name in an IMPLICIT declaration that was later used in a PARAMETER
statement caused problems because the default symbol scope had not yet been
initialized.  I avoided dereferencing in the situation where the default scope
was uninitialized and added a test that triggers the problem.

Also, once I corrected the bad dereference, the compiler was putting out
misleading error messages.  The underlying error us due to violating section
7.5.10, paragraph 4, which states:
  A structure constructor shall not appear before the referenced type is
  defined.

I fixed this by testing to see if a type that is used in a structure
constructor is forward referenced.

Differential Revision: https://reviews.llvm.org/D87535

4 years ago[Take 2] [libc++] Make rotate a constexpr.
zoecarver [Mon, 14 Sep 2020 22:11:08 +0000 (18:11 -0400)]
[Take 2] [libc++] Make rotate a constexpr.

This patch makes `std::rotate` a constexpr. In doing so, this patch also
updates the internal `__move` and `__move_backward` funtions to be
constexpr.

This patch was previously reverted in ed653184ac63 because it was missing
some UNSUPPORTED markup for older compilers. This commit adds it.

Differential Revision: https://reviews.llvm.org/D65721

4 years agoRevert "[libc++] Make rotate a constexpr."
zoecarver [Mon, 14 Sep 2020 21:47:43 +0000 (14:47 -0700)]
Revert "[libc++] Make rotate a constexpr."

This reverts commit 1ec02efee9b1d01cde89f31ca9ba6a46b7662ac5.

4 years ago[flang] More clean-up of CookedSource API
peter klausler [Mon, 14 Sep 2020 20:39:52 +0000 (13:39 -0700)]
[flang] More clean-up of CookedSource API

The std::string holding the content of a CookedSource no longer
needs to be exposed in its API after the recent work that allows
the parsing context to hold multiple instances of a CookedSource.
So clean the API.  These changes were extracted from some work in
progress that was made easier by the API changes.

Differential Revision: https://reviews.llvm.org/D87635

4 years ago[PruneEH][NFC] Use CallGraphUpdater in PruneEH
Arthur Eubanks [Mon, 14 Sep 2020 19:07:17 +0000 (12:07 -0700)]
[PruneEH][NFC] Use CallGraphUpdater in PruneEH

In preparation for porting the pass to NPM.

Reviewed By: asbirlea

Differential Revision: https://reviews.llvm.org/D87632

4 years ago[docs] Update OrcV1 removal timeline.
Lang Hames [Mon, 14 Sep 2020 21:23:20 +0000 (14:23 -0700)]
[docs] Update OrcV1 removal timeline.

4 years ago[libc++][test] Portability fix of std::strstreambuf constructor test
Louis Dionne [Mon, 14 Sep 2020 21:05:49 +0000 (17:05 -0400)]
[libc++][test] Portability fix of std::strstreambuf constructor test

The standard does not require the constructor `strstreambuf(streamsize alsize_arg = 0)`
leave the stream array unallocated when called with parameter `alsize_arg > 0`.
Conformant implementations of this constructor may allocate minimal `alsize_arg`
number of bytes forcing `str()` method to return non-null pointer.

Thanks to Andrey Maksimov for the patch.

Differential Revision: https://reviews.llvm.org/D72465

4 years agoFix MLIR standalone example to properly handle namespace
Mehdi Amini [Mon, 14 Sep 2020 21:02:33 +0000 (21:02 +0000)]
Fix MLIR standalone example to properly handle namespace

ODS TableGen backend now requires to spell out which namespace they have
to be nested in, in an absolute way.

4 years ago[libc++] Finish implementing P0202R3
Nicholas-Baron [Mon, 14 Sep 2020 20:37:41 +0000 (16:37 -0400)]
[libc++] Finish implementing P0202R3

cppreference lists the support for this paper as partial.
I found 4 functions which the paper marks as `constexpr`,
but did not use the appropriate macro.

Differential Revision: https://reviews.llvm.org/D84275

4 years ago[libc++] Make rotate a constexpr.
zoecarver [Mon, 14 Sep 2020 20:52:42 +0000 (13:52 -0700)]
[libc++] Make rotate a constexpr.

This patch makes `std::rotate` a constexpr. In doing so, this patch also
updates the internal `__move` and `__move_backward` funtions to be
constexpr.

Reviewed By: ldionne

Differential Revision: https://reviews.llvm.org/D65721

4 years ago[SelectionDAG] Use Align/MaybeAlign in calls to getLoad/getStore/getExtLoad/getTruncS...
Craig Topper [Mon, 14 Sep 2020 20:54:50 +0000 (13:54 -0700)]
[SelectionDAG] Use Align/MaybeAlign in calls to getLoad/getStore/getExtLoad/getTruncStore.

The versions that take 'unsigned' will be removed in the future.

I tried to use getOriginalAlign instead of getAlign in some
places. getAlign factors in the minimum alignment implied by
the offset in the pointer info. Since we're also passing the
pointer info we can use the original alignment.

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D87592

4 years ago[AMDGPU] Add XDL resource to scheduling model
Austin Kerbow [Mon, 14 Sep 2020 05:32:48 +0000 (22:32 -0700)]
[AMDGPU] Add XDL resource to scheduling model

Reviewed By: rampitec

Differential Revision: https://reviews.llvm.org/D87621

4 years agoUpdate PowerPC backend ownership in CODE_OWNERS.TXT
Nemanja Ivanovic [Mon, 14 Sep 2020 20:45:57 +0000 (15:45 -0500)]
Update PowerPC backend ownership in CODE_OWNERS.TXT

4 years agoMake namespace handling uniform across dialect backends.
Federico Lebrón [Mon, 14 Sep 2020 20:01:07 +0000 (20:01 +0000)]
Make namespace handling uniform across dialect backends.

Now backends spell out which namespace they want to be in, instead of relying on
clients #including them inside already-opened namespaces. This also means that
cppNamespaces should be fully qualified, and there's no implicit "::mlir::"
prepended to them anymore.

Reviewed By: mehdi_amini

Differential Revision: https://reviews.llvm.org/D86811

4 years ago[Asan] Fix false leak report
Vitaly Buka [Mon, 14 Sep 2020 20:32:14 +0000 (13:32 -0700)]
[Asan] Fix false leak report

If user thread is in the allocator, the allocator
may have no pointer into future user's part of
the allocated block. AddrIsInside ignores such
pointers and lsan reports a false memory leak.

Reviewed By: morehouse

Differential Revision: https://reviews.llvm.org/D87552

4 years ago[docs][NewPM] Add docs for writing NPM passes
Arthur Eubanks [Tue, 1 Sep 2020 01:36:11 +0000 (18:36 -0700)]
[docs][NewPM] Add docs for writing NPM passes

As to not conflict with the legacy PM example passes under
llvm/lib/Transforms/Hello, this is under HelloNew. This makes the
CMakeLists.txt and general directory structure less confusing for people
following the example.

Much of the doc structure was taken from WritinAnLLVMPass.rst.

This adds a HelloWorld pass which simply prints out each function name.

More will follow after this, e.g. passes over different units of IR, analyses.
https://llvm.org/docs/WritingAnLLVMPass.html contains a lot more.

Reviewed By: ychen, asbirlea

Differential Revision: https://reviews.llvm.org/D86979

4 years ago[gn build] Port 226d80ebe20
LLVM GN Syncbot [Mon, 14 Sep 2020 20:16:21 +0000 (20:16 +0000)]
[gn build] Port 226d80ebe20

4 years ago[MemProf] Rename HeapProfiler to MemProfiler for consistency
Teresa Johnson [Mon, 14 Sep 2020 16:12:13 +0000 (09:12 -0700)]
[MemProf] Rename HeapProfiler to MemProfiler for consistency

This is consistent with the clang option added in
7ed8124d46f94601d5f1364becee9cee8538265e, and the comments on the
runtime patch in D87120.

Differential Revision: https://reviews.llvm.org/D87622

4 years ago[libc++] Add comment in atomic test to explain why part of it is disabled on Apple
Louis Dionne [Mon, 14 Sep 2020 20:12:47 +0000 (16:12 -0400)]
[libc++] Add comment in atomic test to explain why part of it is disabled on Apple

4 years ago[FastISel] Bail out of selectGetElementPtr for vector GEPs.
Craig Topper [Mon, 14 Sep 2020 19:52:54 +0000 (12:52 -0700)]
[FastISel] Bail out of selectGetElementPtr for vector GEPs.

The code that decomposes the GEP into ADD/MUL doesn't work properly
for vector GEPs. It can create bad COPY instructions or possibly
assert.

For now just bail out to SelectionDAG.

Fixes PR45906

4 years ago[ELF] Add documentation for --warn-backrefs: a GNU ld compatibility checking tool...
Fangrui Song [Tue, 8 Sep 2020 18:37:03 +0000 (11:37 -0700)]
[ELF] Add documentation for --warn-backrefs: a GNU ld compatibility checking tool (and lesser of layering detection)

Differential Revision: https://reviews.llvm.org/D86762

4 years ago[PowerPC] Implement Thread Local Storage Support for Local Exec
Kamau Bridgeman [Fri, 11 Sep 2020 14:33:33 +0000 (10:33 -0400)]
[PowerPC] Implement Thread Local Storage Support for Local Exec

This patch is the initial support for the Local Exec Thread Local
Storage model to produce code sequence and relocations correct
to the ABI for the model when using PC relative memory operations.

Patch by: Kamau Bridgeman

Differential Revision: https://reviews.llvm.org/D83404

4 years ago[AArch64] Add additional vecreduce fmax/fmin legalization tests (NFC)
Nikita Popov [Mon, 14 Sep 2020 19:11:56 +0000 (21:11 +0200)]
[AArch64] Add additional vecreduce fmax/fmin legalization tests (NFC)

Add a vector widening test with ninf flag to the existing fmax
tests, and mirror them over into fmin tests.

4 years ago[libc++] Use LLVM 11 instead of trunk on build bots
Louis Dionne [Mon, 14 Sep 2020 19:12:13 +0000 (15:12 -0400)]
[libc++] Use LLVM 11 instead of trunk on build bots

Somehow the snapshot of LLVM trunk we use was seeing failures.

4 years ago[libc] Extend MPFRMatcher to handle 2-input-1-output and support hypot function.
Tue Ly [Fri, 11 Sep 2020 14:33:33 +0000 (10:33 -0400)]
[libc] Extend MPFRMatcher to handle 2-input-1-output and support hypot function.

Differential Revision: https://reviews.llvm.org/D87514

4 years ago[Legalize][ARM][X86] Add float legalization for VECREDUCE
Nikita Popov [Sat, 12 Sep 2020 20:38:51 +0000 (22:38 +0200)]
[Legalize][ARM][X86] Add float legalization for VECREDUCE

This adds SoftenFloatRes, PromoteFloatRes and SoftPromoteHalfRes
legalizations for VECREDUCE, to fill the remaining hole in the SDAG
legalization. These legalizations simply expand the reduction and
let it be recursively legalized. For the PromoteFloatRes case at
least it is possible to do better than that, but it's pretty tricky
(because we need to consider the interaction of three different
vector legalizations and the type promotion) and probably not
really worthwhile.

I haven't added ExpandFloatRes support, as I am not familiar with
ppc_fp128.

Differential Revision: https://reviews.llvm.org/D87569

4 years ago[ARM] Add more tests for vecreduce soft float legalization (NFC)
Nikita Popov [Mon, 14 Sep 2020 18:37:28 +0000 (20:37 +0200)]
[ARM] Add more tests for vecreduce soft float legalization (NFC)

This mirrors the existing fadd tests to fmul, fmin and fmax.