platform/upstream/llvm.git
3 years ago[AMDGPU] Allow no-modifier operands in cvtDPP
Joe Nash [Mon, 24 May 2021 20:36:45 +0000 (16:36 -0400)]
[AMDGPU] Allow no-modifier operands in cvtDPP

NFC, since no instructions have their AsmMatchConverter
changed, but prepares for that to happen.

Reviewed By: rampitec

Differential Revision: https://reviews.llvm.org/D103046

Change-Id: I6afefad899076de7b9a412374d09b95b29e012fa

3 years ago[CostModel][X86] Improve accuracy of vXi64 vector non-uniform shift costs on AVX2...
Simon Pilgrim [Tue, 25 May 2021 14:24:27 +0000 (15:24 +0100)]
[CostModel][X86] Improve accuracy of vXi64 vector non-uniform shift costs on AVX2+ targets

rG1ad4f887bd7692a9e63fb42586f0ece366f2fe01 incorrectly assumed that vXi64 non-uniform shifts were slow like vXi32 were - but llvm-mca (+Agner) both confirm that Haswell/Broadwell are full rate.

3 years ago[X86][SSE] Regenerate vector shift codegen tests. NFCI.
Simon Pilgrim [Tue, 25 May 2021 11:01:03 +0000 (12:01 +0100)]
[X86][SSE] Regenerate vector shift codegen tests. NFCI.

3 years agoReland "Do not create LLVM IR `constant`s for objects with dynamic initialisation"
Momchil Velikov [Tue, 25 May 2021 13:46:39 +0000 (14:46 +0100)]
Reland "Do not create LLVM IR `constant`s for objects with dynamic initialisation"

This relands commit 13dd65b3a1a3ac049b5f3a9712059f7c61649bea.

The original commit contained a test, which failed when compiled
for a MACH-O target.

This patch changes the test to run for x86_64-linux instead of
`%itanium_abi_triple`, to avoid having invalid syntax for MACH-O
sections. The patch itself does not care about section attribute
syntax and a x86 backend does not even need to be included in the
build.

Differential Revision: https://reviews.llvm.org/D102693

3 years ago[clang][ARM] When handling multiple -mimplicit-it mark all as used
David Spickett [Tue, 25 May 2021 13:54:03 +0000 (13:54 +0000)]
[clang][ARM] When handling multiple -mimplicit-it mark all as used

Since 4468e5b8999291cc84b78f33f207dcd0e58146bc clang will prefer
the last one it finds of "-mimplicit-it" or "-Wa,-mimplicit-it".

Due to a mistake in that patch the compiler argument "-mimplicit-it"
was never marked as used, even if it was the last one and was passed
to llvm.

Move the Claim call back to the start of the loop and update
the testing to check we don't get any unused argument warnings.

Reviewed By: mstorsjo

Differential Revision: https://reviews.llvm.org/D103086

3 years ago[AMDGPU] More accurate names for dpp operand types
Joe Nash [Mon, 24 May 2021 21:05:40 +0000 (17:05 -0400)]
[AMDGPU] More accurate names for dpp operand types

NFC. Renames the variable in the dpp input operand generators
from DstRC to OldRC, because that is what it actually sets.

Also documents the importance of setting HasModifiers = 0 in the
dpp8 asm string.

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D103047

Change-Id: Ice69ae38f644de7f228a75ca47c43e88b1f7d9e1

3 years ago[InstSimplify] Transform X * Y % Y --> 0
David Goldblatt [Tue, 25 May 2021 14:15:27 +0000 (10:15 -0400)]
[InstSimplify] Transform X * Y % Y --> 0

simplifyDiv already handles the case X * Y / Y --> X (barring overflow).
This adds the equivalent handling to simplifyRem.

Correctness:
https://alive2.llvm.org/ce/z/J2cUbS
https://alive2.llvm.org/ce/z/us9NUM
https://alive2.llvm.org/ce/z/AvaDGJ
https://alive2.llvm.org/ce/z/kq9ige

Extending the situations in which we apply this transform would not be
correct:
https://alive2.llvm.org/ce/z/Lf9V63
https://alive2.llvm.org/ce/z/6RPQK3
https://alive2.llvm.org/ce/z/p9UdxC
https://alive2.llvm.org/ce/z/A2zlhE
https://alive2.llvm.org/ce/z/vHTtLw
https://alive2.llvm.org/ce/z/lvpH42

Differential Revision: https://reviews.llvm.org/D102864

3 years ago[VectorCombine] Add test that combines load & store scalarization.
Florian Hahn [Thu, 13 May 2021 15:01:56 +0000 (16:01 +0100)]
[VectorCombine] Add test that combines load & store scalarization.

3 years ago[Headers][WASM] adjust test that runs the optimizer; NFC
Sanjay Patel [Tue, 25 May 2021 13:15:58 +0000 (09:15 -0400)]
[Headers][WASM] adjust test that runs the optimizer; NFC

This broke with the LLVM change in 0bab0f616119

3 years ago[VectorCombine] Use constant range info for index scalarization legality.
Florian Hahn [Tue, 25 May 2021 12:54:55 +0000 (13:54 +0100)]
[VectorCombine] Use constant range info for index scalarization legality.

We can only scalarize memory accesses if we know the index is valid.

This patch adjusts canScalarizeAcceess to fall back to
computeConstantRange to check if the index is known to be valid.

Reviewed By: nlopes

Differential Revision: https://reviews.llvm.org/D102476

3 years ago[mlir][doc] Fix links and references in documentation of Dialects
Markus Böck [Tue, 25 May 2021 12:50:59 +0000 (14:50 +0200)]
[mlir][doc] Fix links and references in documentation of Dialects

This patch is the first in a series of patches fixing markdown links and references inside the mlir documentation. I chose to split it in a few reviews to be able to iterate quicker and to ease review.

This patch addresses all broken references to other markdown files and sections inside the Dialects folder.

One change that was also done was to insert '/' between the markdown files and section:
Example:
Builtin.md#integertype
was changed to:
Builtin.md/#integertype

After compilation, hugo then translates the later to jump directly to the integer type section, but not the former. Not inserting the slash would simply jump to just the Builtin page, instead of the integertype section. I therefore changed occurrences of the former version to the later as well.

Differential Revision: https://reviews.llvm.org/D103011

3 years ago[mlir] Support buffer hoisting on allocas
Tres Popp [Tue, 18 May 2021 09:51:00 +0000 (11:51 +0200)]
[mlir] Support buffer hoisting on allocas

This adds support for hoisting allocas in both BufferHoisting and
BufferLoopHoisting.

Differential Revision: https://reviews.llvm.org/D102681

3 years ago[mlir][doc] Fix links and references in documentation of Rationale
Markus Böck [Tue, 25 May 2021 12:48:07 +0000 (14:48 +0200)]
[mlir][doc] Fix links and references in documentation of Rationale

This patch is the second in a series of patches fixing markdown links and references inside the mlir documentation.

This patch addresses all broken references to other markdown files and sections inside the Rationale folder.

In addition to fixing the links and references like in the previous patch, I also changed references which are URLs to the mlir.llvm.org/docs website, to proper relative markdown references instead.

Differential Revision: https://reviews.llvm.org/D103013

3 years ago[InstCombine] canonicalize cast before unary shuffle
Sanjay Patel [Tue, 25 May 2021 11:52:48 +0000 (07:52 -0400)]
[InstCombine] canonicalize cast before unary shuffle

We could go either direction on this transform. VectorCombine already goes this
way for bitcasts (and handles more complicated cases using the cost model), so
let's try cast-first.

Deferring completely to VectorCombine is another possibility. But the backend
should be able to invert this easily when the vectors have the same shape, so
it doesn't seem like a transform that we need to avoid.

The motivating example from https://llvm.org/PR49081 has an int-to-float
sandwiched between 2 shuffles, and the backend currently does not reduce that,
so on x86, we get something like:

  pshufd $249, %xmm0, %xmm0]
  cvtdq2ps %xmm0, %xmm0
  shufps $144, %xmm0, %xmm0

...instead of just a single conversion instruction.

Differential Revision: https://reviews.llvm.org/D103038

3 years ago[InstCombine] add tests for cast-of-shuffle; NFC
Sanjay Patel [Mon, 24 May 2021 18:10:46 +0000 (14:10 -0400)]
[InstCombine] add tests for cast-of-shuffle; NFC

3 years ago[mlir] Disallow certain transfer ops in VectorToSCF
Matthias Springer [Tue, 25 May 2021 12:30:25 +0000 (21:30 +0900)]
[mlir] Disallow certain transfer ops in VectorToSCF

Disallow transfer ops that change the element type of the transfer. Such transfers could be supported in the future, if needed.

Differential Revision: https://reviews.llvm.org/D102746

3 years ago[Dexter] Remove erroneously added diff file
Tom Weaver [Tue, 25 May 2021 12:36:11 +0000 (13:36 +0100)]
[Dexter] Remove erroneously added diff file

Delete d.diff from debuginfo-tests/dexter directory.

3 years ago[NFC] [Coroutines] Remove unused variable: UnreachableCache
Chuanqi Xu [Tue, 25 May 2021 12:33:46 +0000 (20:33 +0800)]
[NFC] [Coroutines] Remove unused variable: UnreachableCache

3 years ago[dexter] Change --source-root-dir and add --debugger-use-relative-paths
OCHyams [Tue, 25 May 2021 12:09:14 +0000 (13:09 +0100)]
[dexter] Change --source-root-dir and add --debugger-use-relative-paths

We want to use `DexDeclareFile` to specify paths relative to a project root
directory. The option `--source-root-dir`, prior to this patch, causes dexter
to strip the path prefix from commands before passing them to a debugger, and
appends the prefix to file paths returned from a debugger. This patch changes
the behviour of `--source-root-dir`. Relative paths in commands, made possible
with `DexDeclareFile(relative/path)`, are appended to the `--source-root-dir`
directory.

A new option, `--debugger-use-relative-paths`, can be used alongside
`--source-root-dir` to reproduce the old behaviour: all paths passed to the
debugger will be made relative to `--source-root-dir`.

I've added a regression test source_root_dir.dex for this new behaviour, and
modified the existing `--source-root-dir` regression and unit tests to use
`--debugger-use-relative-paths`.

Reviewed By: jmorse

Differential Revision: https://reviews.llvm.org/D100307

3 years ago[LoopIdiom] Support 'left-shift until zero' idiom
Roman Lebedev [Tue, 25 May 2021 12:00:20 +0000 (15:00 +0300)]
[LoopIdiom] Support 'left-shift until zero' idiom

This adds support for the "count active bits" pattern, i.e.:
```
int countBits(unsigned val) {
    int cnt = 0;
    for( ; (val << cnt) != 0; ++cnt)
        ;
    return cnt;
}
```
but a somewhat more general one:
```
int countBits(unsigned val, int start, int off) {
    int cnt;
    for (cnt = start; val << (cnt + off); cnt++)
        ;
    return cnt;
}
```

alive2 is happy with all the tests there.

Note that, again, much like with the right-shift cases,
we don't require the `val != 0` guard.

This is the last pattern that was supported by
`detectShiftUntilZeroIdiom()`, which now becomes obsolete.

3 years ago[NFC][LoopIdiom] Add tests for 'left-shift until zero' idiom
Roman Lebedev [Tue, 25 May 2021 11:58:44 +0000 (14:58 +0300)]
[NFC][LoopIdiom] Add tests for 'left-shift until zero' idiom

3 years ago[AMDGPU][Libomptarget] Mark lambda_by_value test as XFAIL
Pushpinder Singh [Tue, 25 May 2021 11:13:46 +0000 (11:13 +0000)]
[AMDGPU][Libomptarget] Mark lambda_by_value test as XFAIL

Reason: Missing printf definition

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D103078

3 years ago[AArch64][SVE] Add fixed length codegen for FP_TO_{S,U}INT/{S,U}INT_TO_FP
Bradley Smith [Tue, 18 May 2021 12:49:27 +0000 (13:49 +0100)]
[AArch64][SVE] Add fixed length codegen for FP_TO_{S,U}INT/{S,U}INT_TO_FP

Depends on D102607

Differential Revision: https://reviews.llvm.org/D102777

3 years ago[Dexter] Add DexDeclareFile command to Dexter
Tom Weaver [Tue, 25 May 2021 11:47:16 +0000 (12:47 +0100)]
[Dexter] Add DexDeclareFile command to Dexter

DexDeclareFile allows test producers to write test files with .dex extensions
that contain pure dexter commands.

.dex file commands do not need to be commented out like they do when written
inline within test source files.

DexDeclareFile commands are declarative in behaviour, they state that any
Dexter command seen from this point on will have its path attribute set to the
path declared in the DexDeclareFile command.

Differential Revision: https://reviews.llvm.org/D99651

3 years ago[lldb] Fix that LLDB doesn't print NaN's sign on Darwin
Raphael Isemann [Tue, 25 May 2021 11:27:38 +0000 (13:27 +0200)]
[lldb] Fix that LLDB doesn't print NaN's sign on Darwin

It seems std::ostringstream ignores NaN signs on Darwin while it prints them on
Linux. This causes that LLDB behaves differently on those platforms which is
both confusing for users and it also means we have to deal with that in our
tests.

This patch manually implements the NaN/Inf printing (which are apparently
implementation defined) to make LLDB print the same thing on all platforms. The
only output difference in practice seems to be that we now print negative NaNs
as `-nan`, but this potentially also changes the output on other systems I
haven't tested this on.

Reviewed By: JDevlieghere

Differential Revision: https://reviews.llvm.org/D102845

3 years ago[LoopIdiom] Support 'arithmetic right-shift until zero' idiom
Roman Lebedev [Tue, 25 May 2021 11:06:00 +0000 (14:06 +0300)]
[LoopIdiom] Support 'arithmetic right-shift until zero' idiom

This adds support for the "count active bits" pattern, i.e.:
```
int countActiveBits(signed val) {
    int cnt = 0;
    for( ; (val >> cnt) != 0; ++cnt)
        ;
    return cnt;
}
```
but a somewhat more general one:
```
int countActiveBits(signed val, int start, int off) {
    int cnt;
    for (cnt = start; val >> (cnt + off); cnt++)
        ;
    return cnt;
}
```

This directly matches the existing 'logical right-shift until zero' idiom.
alive2 is happy with all the tests there.

Note that, again, much like with the original unsigned case,
we don't require the `val != 0` guard.

The old `detectShiftUntilZeroIdiom()` already supports this pattern,
the idea here is that the `val` must be positive (have at least one
leading zero), because otherwise the loop is non-terminating,
but since it is not `while(1)`, that would have been UB.

3 years ago[NFC][LoopIdiom] Add tests for 'arithmetic right-shift until zero' idiom
Roman Lebedev [Tue, 25 May 2021 11:04:04 +0000 (14:04 +0300)]
[NFC][LoopIdiom] Add tests for 'arithmetic right-shift until zero' idiom

3 years ago[lldb][NFC] Remove misleading ModulePass base class for IRForTarget
Raphael Isemann [Tue, 25 May 2021 11:11:45 +0000 (13:11 +0200)]
[lldb][NFC] Remove misleading ModulePass base class for IRForTarget

IRForTarget is never used by a pass manager or any other interface that requires
this class to inherit from `Pass`.

Also IRForTarget doesn't implement the current interface correctly because it
uses the `runOnModule` return value to indicate success/failure instead of
changed/not-changed, so if this ever ends up being used as a pass it would most
likely not work as intended.

Reviewed By: JDevlieghere

Differential Revision: https://reviews.llvm.org/D102677

3 years ago[lldb] X-FAIL TestCPPStaticMembers on Windows
Raphael Isemann [Tue, 25 May 2021 11:09:45 +0000 (13:09 +0200)]
[lldb] X-FAIL TestCPPStaticMembers on Windows

This was originally failed because of llvm.org/pr21765 which describes that
LLDB can't call a debugee's functions, but I removed the (unnecessary)
function call in the rewrite. It seems that the actual bug here is that we
can't lookup static members at all, so let's X-FAIL the test for the right
reason.

3 years ago[SanitizeCoverage] Add support for NoSanitizeCoverage function attribute
Marco Elver [Tue, 25 May 2021 10:29:00 +0000 (12:29 +0200)]
[SanitizeCoverage] Add support for NoSanitizeCoverage function attribute

We really ought to support no_sanitize("coverage") in line with other
sanitizers. This came up again in discussions on the Linux-kernel
mailing lists, because we currently do workarounds using objtool to
remove coverage instrumentation. Since that support is only on x86, to
continue support coverage instrumentation on other architectures, we
must support selectively disabling coverage instrumentation via function
attributes.

Unfortunately, for SanitizeCoverage, it has not been implemented as a
sanitizer via fsanitize= and associated options in Sanitizers.def, but
rolls its own option fsanitize-coverage. This meant that we never got
"automatic" no_sanitize attribute support.

Implement no_sanitize attribute support by special-casing the string
"coverage" in the NoSanitizeAttr implementation. To keep the feature as
unintrusive to existing IR generation as possible, define a new negative
function attribute NoSanitizeCoverage to propagate the information
through to the instrumentation pass.

Fixes: https://bugs.llvm.org/show_bug.cgi?id=49035

Reviewed By: vitalybuka, morehouse

Differential Revision: https://reviews.llvm.org/D102772

3 years ago[NFC][SanitizeCoverage] Test always_inline functions work
Marco Elver [Tue, 25 May 2021 10:28:50 +0000 (12:28 +0200)]
[NFC][SanitizeCoverage] Test always_inline functions work

Test that always_inline functions are instrumented as expected.

Reviewed By: vitalybuka

Differential Revision: https://reviews.llvm.org/D102929

3 years ago[NFC][CodeGenOptions] Refactor checking SanitizeCoverage options
Marco Elver [Tue, 25 May 2021 10:28:36 +0000 (12:28 +0200)]
[NFC][CodeGenOptions] Refactor checking SanitizeCoverage options

Refactor checking SanitizeCoverage options into
CodeGenOptions::hasSanitizeCoverage().

Reviewed By: vitalybuka

Differential Revision: https://reviews.llvm.org/D102927

3 years agoFix MSVC "truncation of constant value" warning. NFCI.
Simon Pilgrim [Tue, 25 May 2021 10:32:19 +0000 (11:32 +0100)]
Fix MSVC "truncation of constant value" warning. NFCI.

3 years ago[CostModel][X86] Improve accuracy of vXi8/vXi16 vector non-uniform shift costs on...
Simon Pilgrim [Mon, 24 May 2021 17:26:16 +0000 (18:26 +0100)]
[CostModel][X86] Improve accuracy of vXi8/vXi16 vector non-uniform shift costs on AVX2/AVX512 targets

Determined from llvm-mca analysis, AVX2+ capable targets have a higher throughput for VPBLENDVB and VPMOVZX ops, making it cheaper to perform shift+select patterns for vXi8 shifts or extend/shift/truncate for vXi16 shifts. Similarly AVX512BW can perform vXi8 as extend/shift/truncate patterns.

3 years ago[AMDGPU] Remove dead declaration (NFC).
Christudasan Devadasan [Tue, 25 May 2021 10:17:42 +0000 (15:47 +0530)]
[AMDGPU] Remove dead declaration (NFC).

3 years ago[MLIR][Affine][LICM] Mark users of `iter_args` variant
Vinayaka Bandishti [Tue, 25 May 2021 09:49:15 +0000 (15:19 +0530)]
[MLIR][Affine][LICM] Mark users of `iter_args` variant

Prevent users of `iter_args` of an affine for loop from being hoisted
out of it. Otherwise, LICM leads to a violation of the SSA dominance
(as demonstrated in the added test case).

Fixes: https://bugs.llvm.org/show_bug.cgi?id=50103

Reviewed By: bondhugula, ayzhuang

Differential Revision: https://reviews.llvm.org/D102984

3 years ago[mlir] Fold memref.dim of OffsetSizeAndStrideOpInterface outputs
Tres Popp [Tue, 25 May 2021 09:35:14 +0000 (11:35 +0200)]
[mlir] Fold memref.dim of OffsetSizeAndStrideOpInterface outputs

This previously handled memref::SubviewOp, but this can be extended to
all ops implementing the interface.

Differential Revision: https://reviews.llvm.org/D103076

3 years ago[AArch64] Add tests for lowering of vector load + single extract.
Florian Hahn [Tue, 25 May 2021 09:50:08 +0000 (10:50 +0100)]
[AArch64] Add tests for lowering of vector load + single extract.

Currently the vector load + extract gets lowered to a single scalar
store, not accounting for the fact that the index could be
out-of-bounds, which is poison, not UB.

See PR50382.

3 years ago[lldb] Disable minimal import mode for RecordDecls that back FieldDecls
Raphael Isemann [Tue, 25 May 2021 09:53:30 +0000 (11:53 +0200)]
[lldb] Disable minimal import mode for RecordDecls that back FieldDecls

Clang adds a Decl in two phases to a DeclContext. First it adds it invisible and
then it makes it visible (which will add it to the lookup data structures). It's
important that we can't do lookups into the DeclContext we are currently adding
the Decl to during this process as once the Decl has been added, any lookup will
automatically build a new lookup map and add the added Decl to it. The second
step would then add the Decl a second time to the lookup which will lead to
weird errors later one. I made adding a Decl twice to a lookup an assertion
error in D84827.

In the first step Clang also does some computations on the added Decl if it's
for example a FieldDecl that is added to a RecordDecl.

One of these computations is checking if the FieldDecl is of a record type
and the record type has a deleted constexpr destructor which will delete
the constexpr destructor of the record that got the FieldDecl.

This can lead to a bug with the way we implement MinimalImport in LLDB
and the following code:

```
struct Outer {
  typedef int HookToOuter;
  struct NestedClass {
    HookToOuter RefToOuter;
  } NestedClassMember; // We are adding this.
};
```

1. We just imported `Outer` minimally so far.
2. We are now asked to add `NestedClassMember` as a FieldDecl.
3. We import `NestedClass` minimally.
4. We add `NestedClassMember` and clang does a lookup for a constexpr dtor in
   `NestedClass`. `NestedClassMember` hasn't been added to the lookup.
5. The lookup into `NestedClass` will now load the members of `NestedClass`.
6. We try to import the type of `RefToOuter` which will try to import the `HookToOuter` typedef.
7. We import the typedef and while importing we check for conflicts in `Outer` via a lookup.
8. The lookup into `Outer` will cause the invisible `NestedClassMember` to be added to the lookup.
9. We continue normally until we get back to the `addDecl` call in step 2.
10. We now add `NestedClassMember` to the lookup even though we already did that in step 8.

The fix here is disabling the minimal import for RecordTypes from FieldDecls. We
actually already did this, but so far we only force the definition of the type
to be imported *after* we imported the FieldDecl. This just moves that code
*before* we import the FieldDecl so prevent the issue above.

Reviewed By: shafik, aprantl

Differential Revision: https://reviews.llvm.org/D102993

3 years ago[lldb] Re-eanble and rewrite TestCPPStaticMembers
Raphael Isemann [Tue, 25 May 2021 09:43:24 +0000 (11:43 +0200)]
[lldb] Re-eanble and rewrite TestCPPStaticMembers

It's not clear why the whole test got disabled, but the linked bug report
has since been fixed and the only part of it that still fails is the test
for the too permissive lookup. This re-enables the test, rewrites it to use
the modern test functions we have and splits the failing part into its
own test that we can skip without disabling the rest.

3 years ago[IR] Allow Value::replaceUsesWithIf() to process constants
Stanislav Mekhanoshin [Mon, 24 May 2021 21:55:49 +0000 (14:55 -0700)]
[IR] Allow Value::replaceUsesWithIf() to process constants

The change is currently NFC, but exploited by the depending D102954.
Code to handle constants is borrowed from the general implementation
of Value::doRAUW().

Differential Revision: https://reviews.llvm.org/D103051

3 years ago[llvm-exegesis] Loop unrolling for loop snippet repetitor mode
Roman Lebedev [Tue, 25 May 2021 08:48:43 +0000 (11:48 +0300)]
[llvm-exegesis] Loop unrolling for loop snippet repetitor mode

I really needed this, like, factually, yesterday,
when verifying dependency breaking idioms for AMD Zen 3 scheduler model.

Consider the following example:
```
$ ./bin/llvm-exegesis --mode=inverse_throughput --snippets-file=/tmp/snippet.s --num-repetitions=1000000 --repetition-mode=duplicate
Check generated assembly with: /usr/bin/objdump -d /tmp/snippet-4a7e50.o
---
mode:            inverse_throughput
key:
  instructions:
    - 'VPXORYrr YMM0 YMM0 YMM0'
  config:          ''
  register_initial_values: []
cpu_name:        znver3
llvm_triple:     x86_64-unknown-linux-gnu
num_repetitions: 1000000
measurements:
  - { key: inverse_throughput, value: 0.31025, per_snippet_value: 0.31025 }
error:           ''
info:            ''
assembled_snippet: C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C5FDEFC0C3
...

```
What does it tell us?
So wait, it can only execute ~3 x86 AVX YMM PXOR zero-idioms per cycle?
That doesn't seem right. That's even less than there are pipes supporting this type of op.

Now, second example:
```
$ ./bin/llvm-exegesis --mode=inverse_throughput --snippets-file=/tmp/snippet.s --num-repetitions=1000000 --repetition-mode=loop
Check generated assembly with: /usr/bin/objdump -d /tmp/snippet-2418b5.o
---
mode:            inverse_throughput
key:
  instructions:
    - 'VPXORYrr YMM0 YMM0 YMM0'
  config:          ''
  register_initial_values: []
cpu_name:        znver3
llvm_triple:     x86_64-unknown-linux-gnu
num_repetitions: 1000000
measurements:
  - { key: inverse_throughput, value: 1.00011, per_snippet_value: 1.00011 }
error:           ''
info:            ''
assembled_snippet: 49B80800000000000000C5FDEFC0C5FDEFC04983C0FF75F2C3
...
```
Now that's just worse. Due to the looping, the throughput completely plummeted,
and now we can only do a single instruction/cycle!?

That's not great.
And final example:
```
$ ./bin/llvm-exegesis --mode=inverse_throughput --snippets-file=/tmp/snippet.s --num-repetitions=1000000 --repetition-mode=loop --loop-body-size=1000
Check generated assembly with: /usr/bin/objdump -d /tmp/snippet-c402e2.o
---
mode:            inverse_throughput
key:
  instructions:
    - 'VPXORYrr YMM0 YMM0 YMM0'
  config:          ''
  register_initial_values: []
cpu_name:        znver3
llvm_triple:     x86_64-unknown-linux-gnu
num_repetitions: 1000000
measurements:
  - { key: inverse_throughput, value: 0.167087, per_snippet_value: 0.167087 }
error:           ''
info:            ''
assembled_snippet: 49B80800000000000000C5FDEFC0C5FDEFC04983C0FF75F2C3
...
```

So if we merge the previous two approaches, do duplicate this single-instruction snippet 1000x
(loop-body-size/instruction count in snippet), and run a loop with 1000 iterations
over that duplicated/unrolled snippet, the measured throughput goes through the roof,
up to 5.9 instructions/cycle, which finally tells us that this idiom is zero-cycle!

Reviewed By: courbet

Differential Revision: https://reviews.llvm.org/D102522

3 years ago[ARM][NEON] Combine base address updates for vld1x intrinsics
Kristina Bessonova [Tue, 25 May 2021 08:59:38 +0000 (10:59 +0200)]
[ARM][NEON] Combine base address updates for vld1x intrinsics

Reviewed By: dmgreen

Differential Revision: https://reviews.llvm.org/D102855

3 years ago[clang][ARM] Remove non-existent arm9312 CPU
David Spickett [Mon, 24 May 2021 14:00:08 +0000 (14:00 +0000)]
[clang][ARM] Remove non-existent arm9312 CPU

I cannot find documentation on this CPU, and it
is not supported by the Arm Compiler 5 product either.

It was likely a mistake or a different name for the
"ep9312", which is an Arm based Cirrus Logic chip.

Reviewed By: peter.smith

Differential Revision: https://reviews.llvm.org/D103024

3 years ago[llvm][ARM] Remove non-existent arm1176j-s CPU
David Spickett [Mon, 24 May 2021 13:52:00 +0000 (13:52 +0000)]
[llvm][ARM] Remove non-existent arm1176j-s CPU

This was removed in https://reviews.llvm.org/D52594 for clang.

The one test using it has been updated to use the mpcore
CPU as the linked clang change does.

This is part of fixing https://bugs.llvm.org/show_bug.cgi?id=50454.

Reviewed By: peter.smith

Differential Revision: https://reviews.llvm.org/D103022

3 years ago[GlobalISel] Silence unused variable warning in Release builds. NFC.
Benjamin Kramer [Tue, 25 May 2021 08:55:00 +0000 (10:55 +0200)]
[GlobalISel] Silence unused variable warning in Release builds. NFC.

3 years ago[clang][ARM] Remove non-existent arm1136jz-s CPU
David Spickett [Mon, 24 May 2021 13:33:08 +0000 (13:33 +0000)]
[clang][ARM] Remove non-existent arm1136jz-s CPU

There is an ARM1136JF-S and an ARM1136J-S but I could find
no references to an ARM1136JZ-S. In CPU manuals or the manual
for Arm Compiler 5.

See:
https://developer.arm.com/documentation/ddi0211/latest/
https://developer.arm.com/documentation/dui0472/latest/

Using this CPU you get:
$ ./bin/clang --target=arm-linux-gnueabihf -march=armv3m -mcpu=arm1136jz-s -c /tmp/test.c -o /tmp/test.o
'arm1136jz-s' is not a recognized processor for this target (ignoring processor)

Since the llvm target does not know what it is.

This is part of fixing https://bugs.llvm.org/show_bug.cgi?id=50454.

Reviewed By: peter.smith

Differential Revision: https://reviews.llvm.org/D103019

3 years ago[mlir] Check only last dim stride in transfer op lowering
Matthias Springer [Tue, 25 May 2021 08:42:49 +0000 (17:42 +0900)]
[mlir] Check only last dim stride in transfer op lowering

Lower a 1D vector transfer op to LLVM if the last dim stride is 1. Also fixes a bug in the original unit stride computation.

Differential Revision: https://reviews.llvm.org/D102897

3 years ago[TRE] Reland: allow TRE for non-capturing calls.
Alexey Lapshin [Fri, 26 Mar 2021 16:16:26 +0000 (19:16 +0300)]
[TRE] Reland: allow TRE for non-capturing calls.

The D82085 "allow TRE for non-capturing calls" caused failure during bootstrap.
This patch does the same as D82085 plus fixes bootstrap error.

The problem with D82085 is that it does not create copies for byval
operands, while replacing function call with a branch.

Consider following example:

```
    int zoo ( S p1 );

    int foo ( int count, S p1 ) {
      if ( count > 10 )
        return zoo(p1);

      // temporarily variable created for passing byvalue parameter
      // p1 could be used when zoo(p1) is called(after TRE is done).
      // lifetime.start p1.byvalue.temp
      return foo(count+1, p1);
      // lifetime.end p1.byvalue.temp
    }
```

After recursive call to foo is replaced with a jump into
start of the function, its parameters could be passed to
zoo function. i.e. temporarily variable created for byvalue
parameter "p1" could be passed to zoo. Finally zoo receives
broken operand:

```
    int foo ( int count, S p1 ) {
    :tailrecurse
      p1_tr = phi p1, p1.byvalue.temp
      if ( count > 10 )
        return zoo(p1_tr);

      // temporarily variable created for passing byvalue parameter
      // p1 could be used when zoo(p1) is called(after TRE is done).
      lifetime.start p1.byvalue.temp
      memcpy (p1.byvalue.temp, p1_tr)
      count = count + 1
      lifetime.end p1.byvalue.temp
      br tailrecurse
    }
```

To prevent using p1.byvalue.temp after its scope finished by
lifetime.end marker this patch copies value from p1.byvalue.temp
into another temporarily variable and then copies this variable
into the input parameter for next iteration.

This patch passes bootstrap build and bootstrap build with AddressSanitizer.

Differential Revision: https://reviews.llvm.org/D85614

3 years ago[libomptarget][nfc] Accept callable for hsa iterate_symbols
Jon Chesterfield [Tue, 25 May 2021 08:29:10 +0000 (09:29 +0100)]
[libomptarget][nfc] Accept callable for hsa iterate_symbols

[libomptarget][nfc] Accept callable for hsa iterate_symbols
Candidate refactor to simplify D102692

Reviewed By: pdhaliwal

Differential Revision: https://reviews.llvm.org/D103030

3 years ago[analyzer][ctu] Reland "Avoid parsing invocation list again and again..
Ella Ma [Tue, 25 May 2021 07:42:16 +0000 (09:42 +0200)]
[analyzer][ctu] Reland "Avoid parsing invocation list again and again..

..during on-demand parsing of CTU"

During CTU, the *on-demand parsing* will read and parse the invocation
list to know how to compile the file being imported. However, it seems
that the invocation list will be parsed again if a previous parsing
has failed.
Then, parse again and fail again. This patch tries to overcome the
problem by storing the error code during the first parsing, and
re-create the stored error during the later parsings.

Reland without test.

Reviewed By: steakhal

Patch By: OikawaKirie!

Differential Revision: https://reviews.llvm.org/D101763

3 years ago[GlobalISel] Fix MachineIRBuilder not using the DstOp argument for G_SHUFFLE_VECTOR.
Amara Emerson [Mon, 24 May 2021 22:07:00 +0000 (15:07 -0700)]
[GlobalISel] Fix MachineIRBuilder not using the DstOp argument for G_SHUFFLE_VECTOR.

3 years agoRevert "[analyzer][ctu] Avoid parsing invocation list again and again during on-deman...
Balazs Benics [Tue, 25 May 2021 07:28:58 +0000 (09:28 +0200)]
Revert "[analyzer][ctu] Avoid parsing invocation list again and again during on-demand parsing of CTU"

This reverts commit db8af0f21dc9aad4d336754c857c24470afe53e3.

clang-x86_64-debian-fast fails on this.

+ : 'RUN: at line 4'
+ /usr/bin/ccache
/b/1/clang-x86_64-debian-fast/llvm.src/clang/test/Analysis/ctu-on-demand-parsing-multiple-invocation-list-parsing.cpp
-fPIC -shared -o
/b/1/clang-x86_64-debian-fast/llvm.obj/tools/clang/test/Analysis/Output/ctu-on-demand-parsing-multiple-invocation-list-parsing.cpp.tmp/mock_open.so
ccache: error: execv of
/b/1/clang-x86_64-debian-fast/llvm.src/clang/test/Analysis/ctu-on-demand-parsing-multiple-invocation-list-parsing.cpp
failed: Permission denied

3 years ago[analyzer][ctu] Avoid parsing invocation list again and again during on-demand parsin...
Ella Ma [Tue, 25 May 2021 07:19:14 +0000 (09:19 +0200)]
[analyzer][ctu] Avoid parsing invocation list again and again during on-demand parsing of CTU

During CTU, the *on-demand parsing* will read and parse the invocation
list to know how to compile the file being imported. However, it seems
that the invocation list will be parsed again if a previous parsing
has failed.
Then, parse again and fail again. This patch tries to overcome the
problem by storing the error code during the first parsing, and
re-create the stored error during the later parsings.

Reviewed By: steakhal

Patch By: OikawaKirie!

Differential Revision: https://reviews.llvm.org/D101763

3 years ago[RISCV] Optimize xor/or with immediate in the zbs extension
Ben Shi [Tue, 25 May 2021 06:14:09 +0000 (14:14 +0800)]
[RISCV] Optimize xor/or with immediate in the zbs extension

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D102893

3 years ago[JITLink] Suppress expect-death test in release mode.
Lang Hames [Tue, 25 May 2021 05:56:17 +0000 (22:56 -0700)]
[JITLink] Suppress expect-death test in release mode.

3 years ago[LoopDeletion] Break backedge if we can prove that the loop is exited on 1st iteration
Max Kazantsev [Tue, 25 May 2021 05:22:41 +0000 (12:22 +0700)]
[LoopDeletion] Break backedge if we can prove that the loop is exited on 1st iteration

This patch handles one particular case of one-iteration loops for which SCEV
cannot straightforwardly prove BECount = 1. The idea of the optimization is to
symbolically execute conditional branches on the 1st iteration, moving in topoligical
order, and only visiting blocks that may be reached on the first iteration. If we find out
that we never reach header via the latch, then the backedge can be broken.

Differential Revision: https://reviews.llvm.org/D102615
Reviewed By: reames

3 years ago[Test] Add test for unreachable backedge with duplicating predecessors
Max Kazantsev [Tue, 25 May 2021 05:10:31 +0000 (12:10 +0700)]
[Test] Add test for unreachable backedge with duplicating predecessors

3 years agoAMDGPU/GlobalISel: Legalize G_[SU]DIVREM instructions
Christudasan Devadasan [Mon, 12 Apr 2021 10:19:47 +0000 (15:49 +0530)]
AMDGPU/GlobalISel: Legalize G_[SU]DIVREM instructions

Reviewed By: arsenm

Differential Revision: https://reviews.llvm.org/D100726

3 years ago[JITLink] Enable creation and management of mutable block content.
Lang Hames [Tue, 25 May 2021 03:19:32 +0000 (20:19 -0700)]
[JITLink] Enable creation and management of mutable block content.

This patch introduces new operations on jitlink::Blocks: setMutableContent,
getMutableContent and getAlreadyMutableContent. The setMutableContent method
will set the block content data and size members and flag the content as
mutable. The getMutableContent method will return a mutable copy of the existing
content value, auto-allocating and populating a new mutable copy if the existing
content is marked immutable. The getAlreadyMutableMethod asserts that the
existing content is already mutable and returns it.

setMutableContent should be used when updating the block with totally new
content backed by mutable memory. It can be used to change the size of the
block. The argument value should *not* be shared with any other block.

getMutableContent should be used when clients want to modify the existing
content and are unsure whether it is mutable yet.

getAlreadyMutableContent should be used when clients want to modify the existing
content and know from context that it must already be immutable.

These operations reduce copy-modify-update boilerplate and unnecessary copies
introduced when clients couldn't me sure whether the existing content was
mutable or not.

3 years ago[cfe] Support target-specific escaped character in inline asm
Min-Yih Hsu [Fri, 21 May 2021 22:15:11 +0000 (15:15 -0700)]
[cfe] Support target-specific escaped character in inline asm

GCC allows each target to define a set of non-letter and non-digit
escaped characters for inline assembly that will be replaced by another
string (They call this "punctuation" characters. The existing "%%" and
"%{" -- replaced by '%' and '{' at the end -- can be seen as special
cases shared by all targets).
This patch implements this feature by adding a new hook in `TargetInfo`.

Differential Revision: https://reviews.llvm.org/D103036

3 years ago[Sema] Always search the full function scope context if a potential availability...
Logan Smith [Tue, 25 May 2021 04:13:30 +0000 (21:13 -0700)]
[Sema] Always search the full function scope context if a potential availability violation is encountered

This fixes both https://bugs.llvm.org/show_bug.cgi?id=50309 and https://bugs.llvm.org/show_bug.cgi?id=50310.

Previously, lambdas inside functions would mark their own bodies for later analysis when encountering a potentially unavailable decl, without taking into consideration that the entire lambda itself might be correctly guarded inside an @available check. The same applied to inner class member functions. Blocks happened to work as expected already, since Sema::getEnclosingFunction() skips through block scopes.

This patch instead simply and conservatively marks the entire outermost function scope for search, and removes some special-case logic that prevented DiagnoseUnguardedAvailabilityViolations from traversing down into lambdas and nested functions. This correctly accounts for arbitrarily nested lambdas, inner classes, and blocks that may be inside appropriate @available checks at any ancestor level. It also treats all potential availability violations inside functions consistently, without being overly sensitive to the current DeclContext, which previously caused issues where e.g. nested struct members were warned about twice.

DiagnoseUnguardedAvailabilityViolations now has more work to do in some cases, particularly in functions with many (possibly deeply) nested lambdas and classes, but the big-O is the same, and the simplicity of the approach and the fact that it fixes at least two bugs feels like a strong win.

Differential Revision: https://reviews.llvm.org/D102338

3 years ago[lld:elf] Weaken the requirement for a computed binding to be STB_LOCAL
Nathan Lanza [Tue, 16 Mar 2021 08:33:50 +0000 (04:33 -0400)]
[lld:elf] Weaken the requirement for a computed binding to be STB_LOCAL

Given the following scenario:

```
// Cat.cpp
struct Animal { virtual void makeNoise() const = 0; };
struct Cat : Animal { void makeNoise() const override; };

extern "C" int puts(char const *);
void Cat::makeNoise() const { puts("Meow"); }
void doThingWithCat(Animal *a) { static_cast<Cat *>(a)->makeNoise(); }

// CatUser.cpp
struct Animal { virtual void makeNoise() const = 0; };
struct Cat : Animal { void makeNoise() const override; };

void doThingWithCat(Animal *a);

void useDoThingWithCat() {
  Cat *d = new Cat;
  doThingWithCat(d);
}

// cat.ver
{
  global: _Z17useDoThingWithCatv;
  local: *;
};

$ clang++ Cat.cpp CatUser.cpp -fpic -flto=thin -fwhole-program-vtables
-shared -O3 -fuse-ld=lld -Wl,--lto-whole-program-visibility
-Wl,--version-script,cat.ver
```

We cannot devirtualize `Cat::makeNoise`. The issue is complex:

Due to `-fsplit-lto-unit` and usage of type metadata, we place the Cat
vtable declaration into module 0 and the Cat vtable definition with type
metadata into module 1, causing duplicate entries (Undefined followed by
Defined) in the `lto::InputFile::symbols()` output.
In `BitcodeFile::parse`, after processing the `Undefined` then the
`Defined`, the final state is `Defined`.
In `BitcodeCompiler::add`, for the first symbol, `computeBinding`
returns `STB_LOCAL`, then we reset it to `Undefined` because it is
prevailing (`versionId` is `preserved`). For the second symbol, because
the state is now `Undefined`, `computeBinding` returns `STB_GLOBAL`,
causing `ExportDynamic` to be true and suppressing devirtualization.

In D77280, the `computeBinding` change used a stricter `isDefined()`
condition to make weak``Lazy` symbol work.
This patch relaxes the condition to weaker `!isLazy()` to keep it
working while making the devirtualization work as well.

Differential Revision: https://reviews.llvm.org/D98686

3 years agoMaking Instrumentation aware of LoopNest Pass
Arthur Eubanks [Tue, 25 May 2021 01:44:14 +0000 (18:44 -0700)]
Making Instrumentation aware of LoopNest Pass

Intrumentation callbacks are not made aware of LoopNest passes. From the loop pass manager, we can pass the outermost loop of the LoopNest to instrumentation in case of LoopNest passes.

The current patch made the change in two places in StandardInstrumentation.cpp. I will submit a proper patch where the OuterMostLoop is passed from the LoopPassManager to the call backs. That way we will avoid making changes at multiple places in StandardInstrumentation.cpp.

A testcase also will be submitted.

Reviewed By: aeubanks

Differential Revision: https://reviews.llvm.org/D102463

3 years agoRevert "[LoopUnrollAndJam] Change LoopUnrollAndJamPass to LoopNest pass"
maekawatoshiki [Tue, 25 May 2021 02:39:49 +0000 (11:39 +0900)]
Revert "[LoopUnrollAndJam] Change LoopUnrollAndJamPass to LoopNest pass"

This reverts commit d65c32fb41b03a35a2a16330ba1ea15cf6818f04.

3 years ago[libomptarget] [amdgpu] Added LDS usage to the kernel trace
Dhruva Chakrabarti [Mon, 24 May 2021 23:35:29 +0000 (16:35 -0700)]
[libomptarget] [amdgpu] Added LDS usage to the kernel trace

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D103059

3 years agoRevert "Do not create LLVM IR `constant`s for objects with dynamic initialisation"
Nico Weber [Tue, 25 May 2021 01:22:07 +0000 (21:22 -0400)]
Revert "Do not create LLVM IR `constant`s for objects with dynamic initialisation"

This reverts commit 13dd65b3a1a3ac049b5f3a9712059f7c61649bea.
Breaks check-clang on macOS, see https://reviews.llvm.org/D102693

3 years ago[NFC][scudo] Add paramenters DCHECKs
Vitaly Buka [Mon, 24 May 2021 20:12:47 +0000 (13:12 -0700)]
[NFC][scudo] Add paramenters DCHECKs

Reviewed By: hctim

Differential Revision: https://reviews.llvm.org/D103042

3 years agolld-coff: Simplify a few lambda uses after 7975dd033cb9
David Blaikie [Mon, 24 May 2021 23:51:31 +0000 (16:51 -0700)]
lld-coff: Simplify a few lambda uses after 7975dd033cb9

3 years agoAdd a range-based wrapper for std::unique(begin, end, binary_predicate)
David Blaikie [Mon, 24 May 2021 23:48:41 +0000 (16:48 -0700)]
Add a range-based wrapper for std::unique(begin, end, binary_predicate)

3 years ago[NFC][OMP] Fix 'unused' warning
Vitaly Buka [Tue, 25 May 2021 00:14:17 +0000 (17:14 -0700)]
[NFC][OMP] Fix 'unused' warning

3 years ago[NFC][scudo] Avoid cast in test
Vitaly Buka [Tue, 25 May 2021 00:13:29 +0000 (17:13 -0700)]
[NFC][scudo] Avoid cast in test

3 years ago[dsymutil] Emit an error when the Mach-O exceeds the 4GB limit.
Jonas Devlieghere [Mon, 24 May 2021 23:24:16 +0000 (16:24 -0700)]
[dsymutil] Emit an error when the Mach-O exceeds the 4GB limit.

The Mach-O object file format is limited to 4GB because its used of
32-bit offsets in the header. It is possible for dsymutil to (silently)
emit an invalid binary. Instead of having consumers deal with this, emit
an error instead.

3 years ago[dsymutil] Use EXIT_SUCCESS and EXIT_FAILURE (NFC)
Jonas Devlieghere [Mon, 24 May 2021 21:55:52 +0000 (14:55 -0700)]
[dsymutil] Use EXIT_SUCCESS and EXIT_FAILURE (NFC)

3 years ago[dsymutil] Compute the output location once per input file (NFC)
Jonas Devlieghere [Mon, 24 May 2021 21:49:14 +0000 (14:49 -0700)]
[dsymutil] Compute the output location once per input file (NFC)

Compute the location of the output file just once outside the loop over
the different architectures.

3 years agoPR50456: Properly handle multiple escaped newlines in a '*/'.
Richard Smith [Mon, 24 May 2021 23:06:28 +0000 (16:06 -0700)]
PR50456: Properly handle multiple escaped newlines in a '*/'.

3 years ago[scudo] Add unmapTestOnly() to secondary.
Mitch Phillips [Mon, 24 May 2021 23:08:57 +0000 (16:08 -0700)]
[scudo] Add unmapTestOnly() to secondary.

When trying to track down a vaddr-poisoning bug, I found that that the
secondary cache isn't emptied on test teardown. We should probably do
that to make the tests hermetic. Otherwise, repeating the tests lots of
times using --gtest_repeat fails after the mmap vaddr space is
exhausted.

To repro:
$ ninja check-scudo_standalone # build
$ ./projects/compiler-rt/lib/scudo/standalone/tests/ScudoUnitTest-x86_64-Test \
--gtest_filter=ScudoSecondaryTest.*:-ScudoSecondaryTest.SecondaryCombinations \
--gtest_repeat=10000

Reviewed By: cryptoad

Differential Revision: https://reviews.llvm.org/D102874

3 years ago[mlir-opt] Don't enable `printOpOnDiagnostic` if it was explicitly disabled.
River Riddle [Mon, 24 May 2021 22:56:22 +0000 (15:56 -0700)]
[mlir-opt] Don't enable `printOpOnDiagnostic` if it was explicitly disabled.

We are currently explicitly setting the flag solely based on the value of `-verify`, which ends up ignoring the situation where the user explicitly disabled this option from the command line.

Differential Revision: https://reviews.llvm.org/D102952

3 years ago[SLP] Fix "gathering" of insertelement instructions
Anton Afanasyev [Tue, 18 May 2021 08:30:03 +0000 (11:30 +0300)]
[SLP] Fix "gathering" of insertelement instructions

For rare exceptional case vector tree node (insertelements for now only)
is marked as `NeedToGather`, this case is processed by patch. Follow-up
of D98714 to fix bug reported here https://reviews.llvm.org/D98714#2764135.

Differential Revision: https://reviews.llvm.org/D102675

3 years ago[OpenMP] Fix crashing critical section with hint clause
Hansang Bae [Fri, 21 May 2021 23:13:36 +0000 (18:13 -0500)]
[OpenMP] Fix crashing critical section with hint clause

Runtime was using the default lock type without using the hint.

Differential Revision: https://reviews.llvm.org/D102955

3 years ago[libomptarget] [amdgpu] Fix copy-paste error setting NumThreads for a corner case.
Dhruva Chakrabarti [Sat, 22 May 2021 02:35:03 +0000 (19:35 -0700)]
[libomptarget] [amdgpu] Fix copy-paste error setting NumThreads for a corner case.

Fix the case where NumTeams was set incorrectly instead of NumThreads

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D103037

3 years ago[lldb][NFC] Remove unused header from Target
Alex Langford [Mon, 24 May 2021 22:13:06 +0000 (15:13 -0700)]
[lldb][NFC] Remove unused header from Target

Should have been removed with 4c0b0de904a5622c33e3ed97e86c6792fbc13feb
but I forgot to do so.

3 years ago[mlir] Lower sm version for TensorCore intergration tests
thomasraoux [Mon, 24 May 2021 21:32:56 +0000 (14:32 -0700)]
[mlir] Lower sm version for TensorCore intergration tests

Those tests only require sm70, this allows to run those integration
tests on more hardware.

Differential Revision: https://reviews.llvm.org/D103049

3 years ago[compiler-rt][scudo] Fix sign-compare warnings
Jinsong Ji [Mon, 24 May 2021 21:32:50 +0000 (21:32 +0000)]
[compiler-rt][scudo] Fix sign-compare warnings

Fix buildbot failure
https://lab.llvm.org/buildbot/#/builders/57/builds/6542/steps/6/logs/stdio

/llvm-project/llvm/utils/unittest/googletest/include/gtest/gtest.h:1629:28:
error: comparison of integers of different signs: 'const unsigned long'
and 'const int' [-Werror,-Wsign-compare]
GTEST_IMPL_CMP_HELPER_(GT, >);
~~~~~~~~~~~~~~~~~~~~~~~~~~^~
/llvm-project/llvm/utils/unittest/googletest/include/gtest/gtest.h:1609:12:
note: expanded from macro 'GTEST_IMPL_CMP_HELPER_'
  if (val1 op val2) {\
      ~~~~ ^  ~~~~
/llvm-project/compiler-rt/lib/scudo/standalone/tests/common_test.cpp:30:3:
note: in instantiation of function template specialization
'testing::internal::CmpHelperGT<unsigned long, int>' requested here
  EXPECT_GT(OnStart, 0);
  ^

Reviewed By: vitalybuka

Differential Revision: https://reviews.llvm.org/D103029

3 years ago[libc++] Assume that __wrap_iter always wraps a fancy pointer.
Arthur O'Dwyer [Wed, 19 May 2021 15:54:31 +0000 (11:54 -0400)]
[libc++] Assume that __wrap_iter always wraps a fancy pointer.

Not only do we conscientiously avoid using `__wrap_iter` for non-contiguous
iterators (in vector, string, span...) but also we make the assumption
(in regex) that `__wrap_iter<_Iter>` is contiguous for all `_Iter`.

So `__wrap_iter<reverse_iterator<int*>>` should be considered IFNDR,
and every `__wrap_iter` should correctly advertise contiguity in C++20.

Drive-by simplify some type traits.

Reviewed as part of https://reviews.llvm.org/D102781

3 years agoDo not create LLVM IR `constant`s for objects with dynamic initialisation
Momchil Velikov [Mon, 24 May 2021 20:40:47 +0000 (21:40 +0100)]
Do not create LLVM IR `constant`s for objects with dynamic initialisation

When a const-qualified object has a section attribute, that
section is set to read-only and clang outputs a LLVM IR constant
for that object. This is incorrect for dynamically initialised
objects.

For example:

    int init() { return 15; }

    __attribute__((section("SA")))
    const int a = init();

a is allocated to a read-only section and is left
unintialised (zero-initialised).

This patch adds checks if an initialiser is a constant expression
and allocates objects to sections as follows:

* const-qualified objects
  - no initialiser or constant initialiser: .rodata
  - dynamic initializer: .bss
* non const-qualified objects
  - no initialiser or dynamic initialiser: .bss
  - constant initialiser: .data

(".rodata", ".data", and ".bss" names used just for explanatory
purpose)

Differential Revision: https://reviews.llvm.org/D102693

3 years ago[lldb] Move ClangModulesDeclVendor ownership to ClangPersistentVariables from Target
Alex Langford [Wed, 19 May 2021 20:54:14 +0000 (13:54 -0700)]
[lldb] Move ClangModulesDeclVendor ownership to ClangPersistentVariables from Target

More decoupling of plugins and non-plugins. Target doesn't need to
manage ClangModulesDeclVendor and ClangPersistentVariables is always available
in situations where you need ClangModulesDeclVendor.

Differential Revision: https://reviews.llvm.org/D102811

3 years ago[flang][cmake] Set the default for FLANG_BUILD_NEW_DRIVER for oot builds
Andrzej Warzynski [Mon, 24 May 2021 20:10:11 +0000 (20:10 +0000)]
[flang][cmake] Set the default for FLANG_BUILD_NEW_DRIVER for oot builds

For out-of-tree builds of Flang, FLANG_BUILD_NEW_DRIVER is not inherited
from llvm-project/llvm/CMakeLists.txt. Instead, a separate definition is
required (but only for out-of-tree builds).

Differential Revision: https://reviews.llvm.org/D102323

3 years ago[NFC][CSSPGO]llvm-profge] Fix Build warning dueo to an attrbute usage.
Hongtao Yu [Mon, 24 May 2021 19:58:25 +0000 (12:58 -0700)]
[NFC][CSSPGO]llvm-profge] Fix Build warning dueo to an attrbute usage.

3 years ago[GreedyPatternRewriter] Introduce a config object that allows controlling internal...
Chris Lattner [Sun, 23 May 2021 18:24:59 +0000 (11:24 -0700)]
[GreedyPatternRewriter] Introduce a config object that allows controlling internal parameters. NFC.

This exposes the iterations and top-down processing as flags, and also
allows controlling whether region simplification is desirable for a client.
This allows deleting some duplicated entrypoints to
applyPatternsAndFoldGreedily.

This also deletes the Constant Preprocessing pass, which isn't worth it
on balance.

All defaults are all kept the same, so no one should see a behavior change.

Differential Revision: https://reviews.llvm.org/D102988

3 years ago[CSSPGO][llvm-profgen] Report samples for untrackable frames.
Hongtao Yu [Sat, 22 May 2021 00:44:56 +0000 (17:44 -0700)]
[CSSPGO][llvm-profgen] Report samples for untrackable frames.

Fixing an issue where samples collected for an untrackable frame is not reported. An untrackable frame refers to a frame whose caller is untrackable due to missing debug info or pseudo probe. Though the frame is connected to its parent frame through the frame pointer chain at runtime, the compiler cannot build the connection without debug info or pseudo probe. In such case we just need to report the untrackable frame as the base frame and all of its child frames.

With more samples reported I'm seeing this improves the performance of an internal benchmark by 2.5%.

Reviewed By: wenlei, wlei

Differential Revision: https://reviews.llvm.org/D102961

3 years agofix up test from D102742
Nick Desaulniers [Mon, 24 May 2021 19:06:49 +0000 (12:06 -0700)]
fix up test from D102742

In D102742, I mistakenly put the split file designator above a bunch of
CHECK lines, which unintentionally removed the CHECKs from actually
being verified.

This can be verified by observing:
<build dir>/test/CodeGen/X86/Output/stack-protector-3.ll.tmp/main.ll

3 years agoSurface clone APIs in CAPI
George [Mon, 24 May 2021 18:52:41 +0000 (11:52 -0700)]
Surface clone APIs in CAPI

Reviewed By: mehdi_amini

Differential Revision: https://reviews.llvm.org/D102987

3 years ago[gn build] Port b510e4cf1b96
LLVM GN Syncbot [Mon, 24 May 2021 18:48:17 +0000 (18:48 +0000)]
[gn build] Port b510e4cf1b96

3 years ago[RISCV] Add a vsetvli insert pass that can be extended to be aware of incoming VL...
Craig Topper [Mon, 24 May 2021 17:25:27 +0000 (10:25 -0700)]
[RISCV] Add a vsetvli insert pass that can be extended to be aware of incoming VL/VTYPE from other basic blocks.

This is a replacement for D101938 for inserting vsetvli
instructions where needed. This new version changes how
we track the information in such a way that we can extend
it to be aware of VL/VTYPE changes in other blocks. Given
how much it changes the previous patch, I've decided to
abandon the previous patch and post this from scratch.

For now the pass consists of a single phase that assumes
the incoming state from other basic blocks is unknown. A
follow up patch will extend this with a phase to collect
information about how VL/VTYPE change in each block and
a second phase to propagate this information to the entire
function. This will be used by a third phase to do the
vsetvli insertion.

Reviewed By: frasercrmck

Differential Revision: https://reviews.llvm.org/D102737

3 years ago[gn build] Port a64ebb863727
LLVM GN Syncbot [Mon, 24 May 2021 18:36:50 +0000 (18:36 +0000)]
[gn build] Port a64ebb863727

3 years ago[WebAssembly] Add NullifyDebugValueLists pass
Heejin Ahn [Sun, 23 May 2021 09:09:17 +0000 (02:09 -0700)]
[WebAssembly] Add NullifyDebugValueLists pass

`WebAssemblyDebugValueManager` does not currently handle
`DBG_VALUE_LIST`, which is a recent addition to LLVM. We tried to
nullify them within the constructor of `WebAssemblyDebugValueManager` in
D102589, but it made the class error-prone to use because it deletes
instructions within the constructor and thus invalidates existing
iterators within the BB, so the user of the class should take special
care not to use invalidated iterators. This actually caused a bug in
ExplicitLocals pass.

Instead of trying to fix ExplicitLocals pass to make the iterator usage
correct, which is possible but error-prone, this adds
NullifyDebugValueLists pass that nullifies all `DBG_VALUE_LIST`
instructions before we run WebAssembly specific passes in the backend.
We can remove this pass after we implement handlers for
`DBG_VALUE_LIST`s in `WebAssemblyDebugValueManager` and elsewhere.

Fixes https://github.com/emscripten-core/emscripten/issues/14255.

Reviewed By: dschuff

Differential Revision: https://reviews.llvm.org/D102999

3 years ago[dfsan] Add function that prints origin stack trace to buffer
George Balatsouras [Fri, 21 May 2021 17:56:45 +0000 (10:56 -0700)]
[dfsan] Add function that prints origin stack trace to buffer

Reviewed By: stephan.yichao.zhao

Differential Revision: https://reviews.llvm.org/D102451

3 years ago[CUDA] Work around compatibility issue with libstdc++ 11.1.0
Artem Belevich [Fri, 21 May 2021 17:53:28 +0000 (10:53 -0700)]
[CUDA]  Work around compatibility issue with libstdc++ 11.1.0

libstdc++ redeclares __failed_assertion multiple times and that results in the
function declared with conflicting set of attributes when we include <complex>
with __host__ __device__ attributes force-applied to all functions.

In order to work around the issue, we rename __failed_assertion within the
region with forced attributes.

See https://bugs.llvm.org/show_bug.cgi?id=50383 for the details.

Differential Revision: https://reviews.llvm.org/D102936

3 years agoEnable MLIR Python bindings for TOSA.
Stella Laurenzo [Mon, 24 May 2021 16:41:38 +0000 (16:41 +0000)]
Enable MLIR Python bindings for TOSA.

Differential Revision: https://reviews.llvm.org/D103035