platform/upstream/llvm.git
18 months agoDon't loop unswitch vector selects
Valentin Churavy [Sun, 30 Apr 2023 23:19:17 +0000 (19:19 -0400)]
Don't loop unswitch vector selects

Otherwise we could produce `br <2x i1>` which are of course not legal.

```
Branch condition is not 'i1' type!
  br <2 x i1> %cond.fr1, label %entry.split.us, label %entry.split
  %cond.fr1 = freeze <2 x i1> %cond
LLVM ERROR: Broken module found, compilation aborted!
PLEASE submit a bug report to https://github.com/llvm/llvm-project/issues/ and include the crash backtrace.
Stack dump:
0. Program arguments: /home/vchuravy/builds/llvm/bin/opt -passes=simple-loop-unswitch<nontrivial> -S
```

Fixes change introduced by https://reviews.llvm.org/D138526

Reviewed By: caojoshua

Differential Revision: https://reviews.llvm.org/D149560

18 months ago[clang-format] Recognize Verilog type dimension in module header
sstwcw [Sun, 30 Apr 2023 22:26:41 +0000 (22:26 +0000)]
[clang-format] Recognize Verilog type dimension in module header

We had the function `verilogGroupDecl` for that.  However, the type
name would be incorrectly annotated in `isStartOfName` when it was not
a C++ keyword and followed another identifier.

Reviewed By: HazardyKnusperkeks, owenpan, MyDeveloperDay

Differential Revision: https://reviews.llvm.org/D149352

18 months ago[clang-format] Correctly format goto labels followed by blocks
sstwcw [Sun, 30 Apr 2023 22:22:31 +0000 (22:22 +0000)]
[clang-format] Correctly format goto labels followed by blocks

There doesn't seem to be an issue on GitHub.  But previously, a space
would be inserted before the goto colon in the code below.

    switch (x) {
    case 0:
    goto_0: {
      action();
      break;
    }
    }

Previously, the colon following a goto label would be annotated as
`TT_InheritanceColon`.  A goto label followed by an opening brace
wasn't recognized.  It is easy to add another line to have
`spaceRequiredBefore` function recognize the case, but I believed it
is more proper to avoid doing the same thing in `UnwrappedLineParser`
and `TokenAnnotator`.  So now the label colons would be labeled in
`UnwrappedLineParser`, and `spaceRequiredBefore` would rely on that.

Previously we had the type `TT_GotoLabelColon` intended for both goto
labels and case labels.  But since handling of goto labels and case
labels differ somewhat, I split it into separate types for goto and
case labels.

This patch doesn't change the behavior for case labels.  I added the
lines annotating case labels because they would previously be
mistakenly annotated as `TT_InheritanceColon` just like goto labels.
And since I added the annotations, the checks for the `case` and
`default` keywords in `spaceRequiredBefore` are not necessary anymore.

Reviewed By: MyDeveloperDay

Differential Revision: https://reviews.llvm.org/D148484

18 months agoRemove bug fix release note for a bug that was never released.
Richard Smith [Sun, 30 Apr 2023 22:13:22 +0000 (15:13 -0700)]
Remove bug fix release note for a bug that was never released.

In its place, add a description of a bug that the change introducing
the bug in question inadvertently fixed.

See BZ 62361 for more details.

18 months ago[gn build] Port 68ee014b15d3
LLVM GN Syncbot [Sun, 30 Apr 2023 20:27:26 +0000 (20:27 +0000)]
[gn build] Port 68ee014b15d3

18 months ago[libc++][PSTL] Implement std::for_each{, _n}
Nikolas Klauser [Thu, 19 Jan 2023 19:10:11 +0000 (20:10 +0100)]
[libc++][PSTL] Implement std::for_each{, _n}

Reviewed By: ldionne, #libc

Spies: libcxx-commits

Differential Revision: https://reviews.llvm.org/D149381

18 months ago[LV] Clarify comment for selectVectorizationFactor (NFC).
Florian Hahn [Sun, 30 Apr 2023 20:12:15 +0000 (21:12 +0100)]
[LV] Clarify comment for selectVectorizationFactor (NFC).

The comment is stale, as UserVF is handled before selectVectorizationFactor
is called. Clarify the comment by remove the mention of UserVF.

Suggested as independent improvement in D143938.

18 months ago[LV] Remove loop arg from CM::isCandidateForEpilogueVectorization (NFC)
Florian Hahn [Sun, 30 Apr 2023 20:10:55 +0000 (21:10 +0100)]
[LV] Remove loop arg from CM::isCandidateForEpilogueVectorization (NFC)

LVP operates on the loop it stores in TheLoop. Use it instead of the
argument, to be in line with other member functions.

Suggested as independent improvement in D143938.

18 months ago[LV] Document selectEpilogueVectorizationFactor (NFC).
Florian Hahn [Sun, 30 Apr 2023 20:09:24 +0000 (21:09 +0100)]
[LV] Document selectEpilogueVectorizationFactor (NFC).

Add missing documentation for selectEpilogueVectorizationFactor.

Suggested as independent improvement in D143938.

18 months ago[clang-repl] Correctly disambiguate dtor declarations from statements.
Vassil Vassilev [Sun, 30 Apr 2023 19:43:38 +0000 (19:43 +0000)]
[clang-repl] Correctly disambiguate dtor declarations from statements.

Differential revision: https://reviews.llvm.org/D148425

18 months ago[clang-tidy] Ignore declarations in bugprone-exception-escape
Piotr Zegar [Sun, 30 Apr 2023 18:22:44 +0000 (18:22 +0000)]
[clang-tidy] Ignore declarations in bugprone-exception-escape

Warnings will now only be printed for function definitions, not declarations

Reviewed By: isuckatcs

Differential Revision: https://reviews.llvm.org/D148462

18 months ago[LVP] Fix comment for hasPlanWithVF (NFC).
Florian Hahn [Sun, 30 Apr 2023 18:13:52 +0000 (19:13 +0100)]
[LVP] Fix comment for hasPlanWithVF (NFC).

The function checks if there's a plan with the specified VF. Update the
comment to match the implementation.

Pointed out as independent improvement in D143938.

18 months ago[ValueTracking] Use knownbits interface for determining if `div`/`rem` are safe to...
Noah Goldstein [Sun, 30 Apr 2023 15:16:55 +0000 (10:16 -0500)]
[ValueTracking] Use knownbits interface for determining if `div`/`rem` are safe to speculate

This just replaces the exact constant requirements with known-bits
which can prove better results.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149423

18 months ago[ValueTracking] Add tests for checking whether `div`/`rem` is safe to speculate; NFC
Noah Goldstein [Fri, 28 Apr 2023 04:21:29 +0000 (23:21 -0500)]
[ValueTracking] Add tests for checking whether `div`/`rem` is safe to speculate; NFC

Differential Revision: https://reviews.llvm.org/D149422

18 months ago[ValueTracking] Slight refactor to avoid unnecessary work; NFC
Noah Goldstein [Sun, 30 Apr 2023 14:38:43 +0000 (09:38 -0500)]
[ValueTracking] Slight refactor to avoid unnecessary work; NFC

Two changes:
    1) Make some cases that conditionally returned unconditional.
    2) In cases of `Op0 != 0 || Op1 != 0` its better check `Op1` first
       as its more likely to be a constant due to canonicalization (so
       faster to get knownbits of).

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149419

18 months ago[ValueTracking] Add additional cases for `isKnownNonZero(mul X, Y)`
Noah Goldstein [Sun, 30 Apr 2023 14:38:38 +0000 (09:38 -0500)]
[ValueTracking] Add additional cases for `isKnownNonZero(mul X, Y)`

If either `X` or `Y` is odd and the other is non-zero, the result is
non-zero.

Alive2 Link:
    https://alive2.llvm.org/ce/z/9V7-es

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149418

18 months ago[ValueTracking] Add logic for `isKnownNonZero(smin/smax X, Y)`
Noah Goldstein [Sun, 30 Apr 2023 14:33:22 +0000 (09:33 -0500)]
[ValueTracking] Add logic for `isKnownNonZero(smin/smax X, Y)`

For `smin` if either `X` or `Y` is negative, the result is non-zero.
For `smax` if either `X` or `Y` is strictly positive, the result is
non-zero.

For both if `X != 0` and `Y != 0` the result is non-zero.

Alive2 Link:
    https://alive2.llvm.org/ce/z/7yvbgN
    https://alive2.llvm.org/ce/z/zizbvq

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149417

18 months ago[ValueTracking] Add logic for `isKnownNonZero(umin X, Y)`
Noah Goldstein [Sun, 30 Apr 2023 14:33:18 +0000 (09:33 -0500)]
[ValueTracking] Add logic for `isKnownNonZero(umin X, Y)`

`(umin X, Y) != 0` -> `X != 0 && Y != 0`

Alive2 Link:
    https://alive2.llvm.org/ce/z/AQh67i

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149416

18 months ago[ValueTracking] Add logic for `isKnownNonZero(umax X, Y)`
Noah Goldstein [Sun, 30 Apr 2023 14:33:14 +0000 (09:33 -0500)]
[ValueTracking] Add logic for `isKnownNonZero(umax X, Y)`

`(umax X, Y) != 0` -> `X != 0 || Y != 0`

Alive2 Link:
    https://alive2.llvm.org/ce/z/_Z9AUT

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149415

18 months ago[ValueTracking] Add logic for `isKnownNonZero(sadd.sat X, Y)`
Noah Goldstein [Sun, 30 Apr 2023 14:33:10 +0000 (09:33 -0500)]
[ValueTracking] Add logic for `isKnownNonZero(sadd.sat X, Y)`

The logic here is the same for `add` so reuse the existing helper
`isNonZeroAdd`

Alive2 Link:
    https://alive2.llvm.org/ce/z/mhKvC5

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149414

18 months ago[ValueTracking] Add logic for `isKnownNonZero(ssub.sat X, Y)`
Noah Goldstein [Sun, 30 Apr 2023 14:33:03 +0000 (09:33 -0500)]
[ValueTracking] Add logic for `isKnownNonZero(ssub.sat X, Y)`

The logic here is the same for normal `(sub X, Y)`, so just reused
`isNonZeroSub`.

Alive2 Link:
    https://alive2.llvm.org/ce/z/9kSkMv

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149412

18 months ago[ValueTracking] Add logic for `isKnownNonZero(sshl.sat/ushl.sat X, Y)`
Noah Goldstein [Sun, 30 Apr 2023 14:33:00 +0000 (09:33 -0500)]
[ValueTracking] Add logic for `isKnownNonZero(sshl.sat/ushl.sat X, Y)`

`(sshl/ushl X, Y) != 0` -> `X != 0`

Alive2 Links
    https://alive2.llvm.org/ce/z/4WLM2p
    https://alive2.llvm.org/ce/z/BHFng4

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149411

18 months ago[ValueTracking] Add logic for `isKnownNonZero(ctlz/cttz X)`
Noah Goldstein [Sun, 30 Apr 2023 14:32:57 +0000 (09:32 -0500)]
[ValueTracking] Add logic for `isKnownNonZero(ctlz/cttz X)`

for `cttz` if `X[0]` is non-zero, then the expression is non-zero.
for `ctlz` if `X[SignBit]` is non-zero, then the expression in
non-zero.

Alive2 Links:
    cttz (false): https://alive2.llvm.org/ce/z/ySQzbg
    cttz (true): https://alive2.llvm.org/ce/z/auiTCJ
    ctlz (false): https://alive2.llvm.org/ce/z/yk3sTJ
    ctlz (true): https://alive2.llvm.org/ce/z/-JuDty

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149410

18 months ago[ValueTracking] Handle bitcasts between vec-int-ptr in `isKnownNonZero`
Noah Goldstein [Sun, 30 Apr 2023 14:32:54 +0000 (09:32 -0500)]
[ValueTracking] Handle bitcasts between vec-int-ptr in `isKnownNonZero`

We where missing these cases so something like:
`(bitcast to i32 (or v216 x, <2, 1>))`

would not be found to be non-zero.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149409

18 months ago[ValueTracking] Add more tests for `isKnownNonZero` cases; NFC
Noah Goldstein [Sun, 30 Apr 2023 14:32:51 +0000 (09:32 -0500)]
[ValueTracking] Add more tests for `isKnownNonZero` cases; NFC

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149408

18 months ago[ValueTracking] Pull out logic for detecting if `(add X, Y)` is non-zero; NFC
Noah Goldstein [Sun, 30 Apr 2023 14:32:49 +0000 (09:32 -0500)]
[ValueTracking] Pull out logic for detecting if `(add X, Y)` is non-zero; NFC

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149407

18 months ago[ValueTracking] Pull out logic for detecting if `(sub X, Y)` is non-zero; NFC
Noah Goldstein [Sun, 30 Apr 2023 14:32:44 +0000 (09:32 -0500)]
[ValueTracking] Pull out logic for detecting if `(sub X, Y)` is non-zero; NFC

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149406

18 months ago[libc++] Fixes CI breakage.
Mark de Wever [Sun, 30 Apr 2023 12:38:05 +0000 (14:38 +0200)]
[libc++] Fixes CI breakage.

Reviewed By: #libc, Mordante

Differential Revision: https://reviews.llvm.org/D149545

18 months ago[libc++][format] Removes vector dependency.
Mark de Wever [Thu, 20 Apr 2023 18:39:00 +0000 (20:39 +0200)]
[libc++][format] Removes vector dependency.

During the review of D140653 it was suggested to use vector in
__retarget_buffer instead of manually managing the memory. Due to the
requirements of the Standard it turns out format needs to include vector
leading to a cycle. Therefore switching back to manual memory
management.

This is a preparation to fix https://llvm.org/PR61314

Reviewed By: #libc, ldionne

Differential Revision: https://reviews.llvm.org/D148826

18 months ago[LV] Sink collect* calls to LVP::plan() (NFC).
Florian Hahn [Sun, 30 Apr 2023 10:41:22 +0000 (11:41 +0100)]
[LV] Sink collect* calls to LVP::plan() (NFC).

Move calls of collect* helpers closer to where the cost-model is used.
Should help simplifying D142669 & D142670.

Differential Revision: https://reviews.llvm.org/D142674

18 months agoRevert "[InstCombine] Generate better code for std::bit_floor from libstdc++"
Nuno Lopes [Sun, 30 Apr 2023 08:56:34 +0000 (09:56 +0100)]
Revert "[InstCombine] Generate better code for std::bit_floor from libstdc++"

This reverts commit d775fc390d3c78cc81872e276c4b1314f19af577.

The patch is wrong wrt undef and the author didn't fix it after 2 weeks.

18 months agoAdd an action for each iteration of the GreedyPatternRewriteDriver
Mehdi Amini [Thu, 23 Feb 2023 05:43:45 +0000 (22:43 -0700)]
Add an action for each iteration of the GreedyPatternRewriteDriver

Differential Revision: https://reviews.llvm.org/D149101

18 months ago[LegalizeVectorOps][RISCV] Support condition code legalization for ISD::STRICT_FSETCC...
Craig Topper [Sun, 30 Apr 2023 05:55:41 +0000 (22:55 -0700)]
[LegalizeVectorOps][RISCV] Support condition code legalization for ISD::STRICT_FSETCC/FSETCCS during LegalizeVectorOps.

Switch RISC-V to legalize during LegalizeVectorOps instead of
LegalizeDAG. LegalizeDAG uses the OpVT for legalize action while
LegalizeVectorOps uses the result VT. We really should fix that.

18 months ago[RISCV] Prevent lowerVectorStrictFSetcc from creatin an ISD::AND with identical operands.
Craig Topper [Sun, 30 Apr 2023 04:42:45 +0000 (21:42 -0700)]
[RISCV] Prevent lowerVectorStrictFSetcc from creatin an ISD::AND with identical operands.

This AND immediately gets legalized to RISCVISD::VMAND_VL and we don't
yet have DAG combine to optimize that away. So this is a quick fix to
improve generated code.

18 months ago[SimpleLoopUnswitch] unswitch selects
Joshua Cao [Sat, 15 Apr 2023 23:04:58 +0000 (16:04 -0700)]
[SimpleLoopUnswitch] unswitch selects

The old LoopUnswitch pass unswitched selects, but the changes were never
ported to the new SimpleLoopUnswitch.

We unswitch by turning:

```
S = select %cond, %a, %b
```

into:

```
head:
br %cond, label %then, label %tail

then:
br label %tail

tail:
S = phi [ %a, %then ], [ %b, %head ]
```

Unswitch selects are always nontrivial, since the successors do not exit
the loop and the loop body always needs to be cloned.

Differential Revision: https://reviews.llvm.org/D138526

Co-authored-by: Sergey Kachkov <sergey.kachkov@syntacore.com>
18 months ago[SimpleLoopUnswitch] Add unswitch select tests
Joshua Cao [Mon, 17 Apr 2023 00:25:05 +0000 (17:25 -0700)]
[SimpleLoopUnswitch] Add unswitch select tests

18 months ago[TableGen] Forward declare CodeGenRegister et al. (NFC)
Sergei Barannikov [Sun, 30 Apr 2023 04:00:20 +0000 (07:00 +0300)]
[TableGen] Forward declare CodeGenRegister et al. (NFC)

18 months ago[gn build] Port ebc05b93a11b
LLVM GN Syncbot [Sun, 30 Apr 2023 03:41:51 +0000 (03:41 +0000)]
[gn build] Port ebc05b93a11b

18 months ago[gn build] Port 8af259e8d90b
LLVM GN Syncbot [Sun, 30 Apr 2023 03:41:50 +0000 (03:41 +0000)]
[gn build] Port 8af259e8d90b

18 months ago[libc++][PSTL] Implement std::{any, all, none}_of
Nikolas Klauser [Sat, 14 Jan 2023 01:08:27 +0000 (02:08 +0100)]
[libc++][PSTL] Implement std::{any, all, none}_of

Reviewed By: ldionne, #libc

Spies: arichardson, libcxx-commits, miyuki

Differential Revision: https://reviews.llvm.org/D143161

18 months ago[libc++][PSTL] Implement <execution> contents
Nikolas Klauser [Fri, 13 Jan 2023 21:48:23 +0000 (22:48 +0100)]
[libc++][PSTL] Implement <execution> contents

Reviewed By: ldionne, #libc

Spies: libcxx-commits, arichardson

Differential Revision: https://reviews.llvm.org/D141888

18 months agoRevert "[NFC][LLD] Disambiguate RISCV and llvm::RISCV"
Vitaly Buka [Sun, 30 Apr 2023 02:20:41 +0000 (19:20 -0700)]
Revert "[NFC][LLD] Disambiguate RISCV and llvm::RISCV"

This reverts commit 176cc70abe8d85df9aae223e0b35ce65238c4333.

18 months ago[NFC][LLD] Disambiguate RISCV and llvm::RISCV
Vitaly Buka [Sun, 30 Apr 2023 02:06:18 +0000 (19:06 -0700)]
[NFC][LLD] Disambiguate RISCV and llvm::RISCV

Build is broken after ee9cbe35.

18 months agoRevert "[RISCV] Move RISCV::RVVBitsPerBlock from TargetParser to Support/RISCVISAInfo.h."
Craig Topper [Sun, 30 Apr 2023 01:05:56 +0000 (18:05 -0700)]
Revert "[RISCV] Move RISCV::RVVBitsPerBlock from TargetParser to Support/RISCVISAInfo.h."

This reverts commit ee9cbe3548cded885c6409d6dd8a616b515a06d3.

I've been told this cauess a namespace clash in lld.

18 months ago[RISCV] Use TokenFactor instead of MERGE_VALUES to combine chains.
Craig Topper [Sun, 30 Apr 2023 01:05:10 +0000 (18:05 -0700)]
[RISCV] Use TokenFactor instead of MERGE_VALUES to combine chains.

18 months ago[NFC][HWASAN] Use pointercast instead of bitcast
Vitaly Buka [Sun, 30 Apr 2023 00:40:33 +0000 (17:40 -0700)]
[NFC][HWASAN] Use pointercast instead of bitcast

18 months ago[NFC][HWASAN] Add cont to parameter
Vitaly Buka [Sun, 30 Apr 2023 00:22:45 +0000 (17:22 -0700)]
[NFC][HWASAN] Add cont to parameter

18 months ago[NFC][HWASAN] Fix comment
Vitaly Buka [Sun, 30 Apr 2023 00:22:12 +0000 (17:22 -0700)]
[NFC][HWASAN] Fix comment

18 months ago[NFC][HWASAN] Remove unused parameter
Vitaly Buka [Sun, 30 Apr 2023 00:14:39 +0000 (17:14 -0700)]
[NFC][HWASAN] Remove unused parameter

18 months ago[M68k] Add instruction selection support for zext with PCD addressing
Ian Douglas Scott [Sat, 29 Apr 2023 23:26:17 +0000 (16:26 -0700)]
[M68k] Add instruction selection support for zext with PCD addressing

Instruction selection was failing when trying to zero extend a value
loaded from a PC-relative address. This adds support for zero extension
using the "program counter indirect with displacement" addressing mode.
It also adds a test with code that was previously failing to compile.

This fixes a compile error in Rust's libcore.

Differential Revision: https://reviews.llvm.org/D149034

18 months ago[X86] Update some stale comments. NFC
Craig Topper [Sat, 29 Apr 2023 22:38:31 +0000 (15:38 -0700)]
[X86] Update some stale comments. NFC

18 months ago[mlir][vector][NFC] Update `vector.flat_transpose` description
Jakub Kuderski [Sat, 29 Apr 2023 22:34:33 +0000 (18:34 -0400)]
[mlir][vector][NFC] Update `vector.flat_transpose` description

Change the example to match how this op is actually printed. Do not use
'smart quotes' in the description.

Reviewed By: jpienaar

Differential Revision: https://reviews.llvm.org/D149535

18 months ago[TargetLowering] Stop passing an ISD::CondCode to isOperationLegalOrCustom.
Craig Topper [Sat, 29 Apr 2023 22:10:29 +0000 (15:10 -0700)]
[TargetLowering] Stop passing an ISD::CondCode to isOperationLegalOrCustom.

ISD::CondCode is a separate num space from opcodes. isOperationLegalOrCustom
should take an opcode.

Reviewed By: barannikov88

Differential Revision: https://reviews.llvm.org/D149528

18 months ago[RISCV] Move RISCV::RVVBitsPerBlock from TargetParser to Support/RISCVISAInfo.h.
Craig Topper [Sat, 29 Apr 2023 22:04:52 +0000 (15:04 -0700)]
[RISCV] Move RISCV::RVVBitsPerBlock from TargetParser to Support/RISCVISAInfo.h.

RISCVTargetParser.h has a dependency on a tablegen generated file.

Using RISCVISAInfo.h instead avoids this dependency.

We just need this constant somewhere visible to the frontend and
backend and I'm trying to avoid adding a header just for it.

18 months ago[RISCV][Sema] Add a build dependency on RISCVTargetParserTableGen after D145088.
Craig Topper [Sat, 29 Apr 2023 20:33:02 +0000 (13:33 -0700)]
[RISCV][Sema] Add a build dependency on RISCVTargetParserTableGen after D145088.

Need to find a better home for RISCV::RVVBitsPerBlock to remove
this dependency. This just a quick fix to fix the build.

18 months ago[SCEV] Use object size for globals to sharpen ranges.
Florian Hahn [Sat, 29 Apr 2023 20:33:06 +0000 (21:33 +0100)]
[SCEV] Use object size for globals to sharpen ranges.

The highest address the object can start is ObjSize bytes before the
end (unsigned max value). If this value is not a multiple of the
alignment, the last possible start value is the next lowest multiple
of the alignment. Note: The computations cannot overflow,
because if they would there's no possible start address for the
object.

At the moment, this is limited to GlobalVariables, because I could not
find a API similar to getObjectSize to also get the alignment of the
object. With such an API, this can be generalized to general addresses.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149483

18 months agoGlobalOpt: Add some test coverage for addrspacecast treatment
Matt Arsenault [Sat, 29 Apr 2023 19:56:15 +0000 (15:56 -0400)]
GlobalOpt: Add some test coverage for addrspacecast treatment

18 months ago[ARM] Add predicated shift patterns
David Green [Sat, 29 Apr 2023 19:32:54 +0000 (20:32 +0100)]
[ARM] Add predicated shift patterns

This uses the patterns defined in MVE_TwoOpPattern to add predicated patterns
for vshls/u instructions.

Differnetial Revision: https://reviews.llvm.org/D149366

18 months agoAMDGPU: Don't need pointer bitcast in AMDGPULowerKernelArguments
Matt Arsenault [Sat, 29 Apr 2023 16:27:22 +0000 (12:27 -0400)]
AMDGPU: Don't need pointer bitcast in AMDGPULowerKernelArguments

18 months ago[SelectionDAG] Rename ADDCARRY/SUBCARRY to UADDO_CARRY/USUBO_CARRY (NFC)
Sergei Barannikov [Sat, 29 Apr 2023 18:07:04 +0000 (21:07 +0300)]
[SelectionDAG] Rename ADDCARRY/SUBCARRY to UADDO_CARRY/USUBO_CARRY (NFC)

This will make them consistent with other overflow-aware nodes.

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D148196

18 months ago[emacs] Highlight 'ptr' keyword as a type
Noah Goldstein [Sat, 29 Apr 2023 18:49:14 +0000 (13:49 -0500)]
[emacs] Highlight 'ptr' keyword as a type

Seems natural to highlight 'ptr' same as 'void', 'i1', etc...

Reviewed By: artagnon, nikic

Differential Revision: https://reviews.llvm.org/D149519

18 months ago[InstCombine] Fold `(cmp eq/ne (umax X, Y),0)` -> `(cmp eq/ne (or X, Y),0)`
Noah Goldstein [Sat, 29 Apr 2023 16:26:33 +0000 (11:26 -0500)]
[InstCombine] Fold `(cmp eq/ne (umax X, Y),0)` -> `(cmp eq/ne (or X, Y),0)`

`or` is almost always preferable.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D149426

18 months ago[InstCombine] Don't fold `uadd.sat` to `or` if it increase instruction count
Noah Goldstein [Fri, 28 Apr 2023 04:52:12 +0000 (23:52 -0500)]
[InstCombine] Don't fold `uadd.sat` to `or` if it increase instruction count

In the `(cmp eq/ne (uadd.sat X, Y),0)` case, we where missing a
`hasOneUse` check.

Differential Revision: https://reviews.llvm.org/D149425

18 months ago[InstCombine] Add tests `(cmp eq/ne (umax/uadd.sat X, Y), 0)`; NFC
Noah Goldstein [Fri, 28 Apr 2023 04:09:05 +0000 (23:09 -0500)]
[InstCombine] Add tests `(cmp eq/ne (umax/uadd.sat X, Y), 0)`; NFC

Differential Revision: https://reviews.llvm.org/D149424

18 months ago[TargetLowering] Don't use ISD::SELECT_CC in expandFP_TO_INT_SAT.
Craig Topper [Sat, 29 Apr 2023 17:06:22 +0000 (10:06 -0700)]
[TargetLowering] Don't use ISD::SELECT_CC in expandFP_TO_INT_SAT.

This function gets called for vectors and ISD::SELECT_CC was never
intended to support vectors. Some updates were made to support
it when this function started getting used for vectors.

Overall, using separate ISD::SETCC and ISD::SELECT looks like an
improvement even for scalar.

Reviewed By: RKSimon

Differential Revision: https://reviews.llvm.org/D149481

18 months ago[lldb][ObjectFileELF] Support AArch32 in ApplyRelocations
Stefan Gränitz [Sat, 29 Apr 2023 15:29:28 +0000 (17:29 +0200)]
[lldb][ObjectFileELF] Support AArch32 in ApplyRelocations

Allow the ObjectFileELF plugin to resolve R_ARM_ABS32 relocations from AArch32 object files. This fixes https://github.com/llvm/llvm-project/issues/61948

The existing architectures work with RELA-type relocation records that read addend from the relocation entry. REL-type relocations in AArch32 store addend in-place.
The new function doesn't re-use ELFRelocation::RelocAddend32(), because the interface doesn't match: in addition to the relocation entry we need the actual target section memory.

Reviewed By: labath

Differential Revision: https://reviews.llvm.org/D147642

18 months agoAMDGPU: Don't need pointer bitcasts in AMDGPURewriteOutArguments
Matt Arsenault [Sat, 29 Apr 2023 14:18:34 +0000 (10:18 -0400)]
AMDGPU: Don't need pointer bitcasts in AMDGPURewriteOutArguments

18 months agoAMDGPU: Remove code to create pointer bitcasts in atomicrmw expansion
Matt Arsenault [Sat, 29 Apr 2023 14:14:30 +0000 (10:14 -0400)]
AMDGPU: Remove code to create pointer bitcasts in atomicrmw expansion

18 months agoAMDGPU: Use getTypeStoreSizeInBits
Matt Arsenault [Sat, 29 Apr 2023 14:12:11 +0000 (10:12 -0400)]
AMDGPU: Use getTypeStoreSizeInBits

18 months agoAMDGPU: Don't create a pointer bitcast in AMDGPULateCodeGenPrepare
Matt Arsenault [Sat, 29 Apr 2023 14:11:21 +0000 (10:11 -0400)]
AMDGPU: Don't create a pointer bitcast in AMDGPULateCodeGenPrepare

18 months agoAMDGPU: Don't try to create pointer bitcasts in kernarg lowering
Matt Arsenault [Sat, 29 Apr 2023 14:04:06 +0000 (10:04 -0400)]
AMDGPU: Don't try to create pointer bitcasts in kernarg lowering

18 months agoAMDGPU: Don't try to create pointer bitcasts in load widening
Matt Arsenault [Sat, 29 Apr 2023 14:01:37 +0000 (10:01 -0400)]
AMDGPU: Don't try to create pointer bitcasts in load widening

18 months agoGlobalOpt: Drop code to handle typed pointers
Matt Arsenault [Wed, 26 Apr 2023 19:27:54 +0000 (15:27 -0400)]
GlobalOpt: Drop code to handle typed pointers

Fixes assert with pointers with different address spaces. We
could keep looking through addrspacecast, but it would require
checking for null handling of the access address space.

Fixes #62384

18 months ago[libc] Add support for global ctors / dtors for AMDGPU
Joseph Huber [Thu, 27 Apr 2023 23:29:10 +0000 (18:29 -0500)]
[libc] Add support for global ctors / dtors for AMDGPU

This patch makes the necessary changes to support calling global
constructors and destructors on the GPU. The patch in D149340 allows the
`lld` linker to create the symbols pointing us to these globals. These
should be executed by a single thread, which is more difficult on the
GPU because all threads are active. I chose to use an atomic counter to
sync every thread on the GPU. This is very slow if you use more than a
few thousand threads, but for testing purposes it should be sufficient.

Depends on D149340 D149363

Reviewed By: sivachandra

Differential Revision: https://reviews.llvm.org/D149398

18 months ago[AMDGPU] Place global constructors in .init_array and .fini_array
Joseph Huber [Thu, 27 Apr 2023 13:00:01 +0000 (08:00 -0500)]
[AMDGPU] Place global constructors in .init_array and .fini_array

For the GPU, we emit external kernels that call the initializers and
constructors, however if we had a persistent kernel like in the `_start`
kernel for the `libc` project, we could initialize the standard way of
calling constructors. This patch adds new global variables containing
pointers to the constructors to be called. If these are placed in the
`.init_array` and `.fini_array` sections, then the backend will handle
them specially. The linker will then provide the `__init_array_` and
`__fini_array_` sections to traverse them. An implementation would look
like this.

```
extern uintptr_t __init_array_start[];
extern uintptr_t __init_array_end[];
extern uintptr_t __fini_array_start[];
extern uintptr_t __fini_array_end[];

using InitCallback = void(int, char **, char **);
using FiniCallback = void(void);

extern "C" [[gnu::visibility("protected"), clang::amdgpu_kernel]] void
_start(int argc, char **argv, char **envp) {
  uint64_t init_array_size = __init_array_end - __init_array_start;
  for (uint64_t i = 0; i < init_array_size; ++i)
    reinterpret_cast<InitCallback *>(__init_array_start[i])(argc, argv, env);
  uint64_t fini_array_size = __fini_array_end - __fini_array_start;
  for (uint64_t i = 0; i < fini_array_size; ++i)
    reinterpret_cast<FiniCallback *>(__fini_array_start[i])();
}
```

Reviewed By: yaxunl

Differential Revision: https://reviews.llvm.org/D149340

18 months agoLangRef: Add "dynamic" option to "denormal-fp-math"
Matt Arsenault [Tue, 6 Dec 2022 14:25:33 +0000 (09:25 -0500)]
LangRef: Add "dynamic" option to "denormal-fp-math"

This is stricter than the default "ieee", and should probably be the
default. This patch leaves the default alone. I can change this in a
future patch.

There are non-reversible transforms I would like to perform which are
legal under IEEE denormal handling, but illegal with flushing zero
behavior. Namely, conversions between llvm.is.fpclass and fcmp with
zeroes.

Under "ieee" handling, it is legal to translate between
llvm.is.fpclass(x, fcZero) and fcmp x, 0.

Under "preserve-sign" handling, it is legal to translate between
llvm.is.fpclass(x, fcSubnormal|fcZero) and fcmp x, 0.

I would like to compile and distribute some math library functions in
a mode where it's callable from code with and without denormals
enabled, which requires not changing the compares with denormals or
zeroes.

If an IEEE function transforms an llvm.is.fpclass call into an fcmp 0,
it is no longer possible to call the function from code with denormals
enabled, or write an optimization to move the function into a denormal
flushing mode. For the original function, if x was a denormal, the
class would evaluate to false. If the function compiled with denormal
handling was converted to or called from a preserve-sign function, the
fcmp now evaluates to true.

This could also be of use for strictfp handling, where code may be
changing the denormal mode.

Alternative name could be "unknown".

Replaces the old AMDGPU custom inlining logic with more conservative
logic which tries to permit inlining for callees with dynamic handling
and avoids inlining other mismatched modes.

18 months ago[mlir][bytecode] Allow client to specify a desired version.
Jacques Pienaar [Sat, 29 Apr 2023 12:35:53 +0000 (05:35 -0700)]
[mlir][bytecode] Allow client to specify a desired version.

Add method to set a desired bytecode file format to generate. Change
write method to be able to return status including the minimum bytecode
version needed by reader. This enables generating an older version of
the bytecode (not dialect ops, attributes or types). But this does not
guarantee that an older version can always be generated, e.g., if a
dialect uses a new encoding only available at later bytecode version.
This clamps setting to at most current version.

Differential Revision: https://reviews.llvm.org/D146555

18 months ago[clangd] Hover: Add CalleeArgInfo for constructor expressions
Tom Praschan [Sat, 8 Apr 2023 12:57:34 +0000 (14:57 +0200)]
[clangd] Hover: Add CalleeArgInfo for constructor expressions

Differential Revision: https://reviews.llvm.org/D147847

18 months ago[CMake] Serialize `build_native_tool`
NAKAMURA Takumi [Mon, 24 Apr 2023 13:59:56 +0000 (22:59 +0900)]
[CMake] Serialize `build_native_tool`

To prevent race in `NATIVE`, let each action depend on preceding action.
At the moment, this is restricted only for "Visual Studio".

For example,

  - `llvm-tblgen-host` depends on the target's `llvm-tblgen`.
  - `clang-tblgen-host` depends on the target's `clang-tblgen` and `llvm-tblgen-host`.

This is rework for D54153.
`build_native_tool` has been introduced since D60024.

Differential Revision: https://reviews.llvm.org/D149072

18 months ago[clang-format] Add BracedInitializerIndentWidth option
Jon Phillips [Sat, 29 Apr 2023 07:22:22 +0000 (00:22 -0700)]
[clang-format] Add BracedInitializerIndentWidth option

The option allows users to specify how many columns to use to indent
the contents of initializer lists.

Closes #51070.

Differential Revision: https://reviews.llvm.org/D146101

18 months ago[X86] Fix the vnni machine combine issue.
Luo, Yuanke [Sat, 29 Apr 2023 05:14:54 +0000 (13:14 +0800)]
[X86] Fix the vnni machine combine issue.

The previous patch (D148980) didn't set the InstrIdxForVirtReg correctly
in genAlternativeDpCodeSequence(). It causes vnni lit test failure when
LLVM_ENABLE_EXPENSIVE_CHECKS is on.

18 months agoRevert "Reland D147337 "[tsan] Add debugging interfaces into interface header.""
Hans Wennborg [Sat, 29 Apr 2023 05:06:16 +0000 (07:06 +0200)]
Revert "Reland D147337 "[tsan] Add debugging interfaces into interface header.""

This broke the lit tests on Mac, see comment on the code review.

> This change the types to match the ones used in:
> Darwin/debug_external.cpp
> debugging.cpp
>
> Reviewed By: vitalybuka
>
> Differential Revision: https://reviews.llvm.org/D148214

This reverts commit ea7d6e658e37363c61d2f9ce64de0b223a39be50.

18 months ago[RISCV] Disable strict node mutation and use correct lowering action of strict_fsetcc(s).
Yeting Kuo [Thu, 27 Apr 2023 07:37:39 +0000 (15:37 +0800)]
[RISCV] Disable strict node mutation and use correct lowering action of strict_fsetcc(s).

The patch disables strict node mutation by setting IsStrictFPEnabled to true.
It also fixes wrong lowering action of strict_fsetcc(s).

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D149320

18 months agoGOFFObjectFile.cpp: Suppress a warning in -Asserts (D98437)
NAKAMURA Takumi [Sat, 29 Apr 2023 02:31:27 +0000 (11:31 +0900)]
GOFFObjectFile.cpp: Suppress a warning in -Asserts (D98437)

18 months ago[libc] Fix printing on the GPU when given a `cpp::string_ref`
Joseph Huber [Fri, 28 Apr 2023 22:44:17 +0000 (17:44 -0500)]
[libc] Fix printing on the GPU when given a `cpp::string_ref`

The implementation of the test printing currently expects a null
terminated C-string. However, the `write_to_stderr` interface uses a
string view, which doesn't need to be null terminated. This patch
changes the printing interface to directly use `fwrite` instead rather
than relying on a null terminator.

Reviewed By: sivachandra

Differential Revision: https://reviews.llvm.org/D149493

18 months ago[NFC][clang] Fix static analyzer concerns about AUTO_CAUSES_COPY
Manna, Soumi [Sat, 29 Apr 2023 01:38:07 +0000 (18:38 -0700)]
[NFC][clang] Fix static analyzer concerns about AUTO_CAUSES_COPY

Reported by Coverity:

AUTO_CAUSES_COPY
Unnecessary object copies can affect performance

1. Inside "ODRHash.cpp" file, in clang::ODRHash::AddCXXRecordDecl(clang::CXXRecordDecl const *): Using the auto keyword without an & causes the copy of an object of type CXXBaseSpecifier.

2. Inside "Tokens.cpp" file, in clang::syntax::TokenBuffer::dumpForTests[abi:cxx11](): Using the auto keyword without an & causes the copy of an object of type DenseMapPair.

3. Inside "TargetID.cpp" file, in clang::getCanonicalTargetID[abi:cxx11](llvm::StringRef, llvm::StringMap<bool, llvm::MallocAllocator> const &): Using the auto keyword without an & causes the copy of an object of type pair.

Reviewed By: tahonermann, aaron.ballman

Differential Revision: https://reviews.llvm.org/D148639

18 months agoSimplify affine bound min or max with operands
Uday Bondhugula [Sat, 29 Apr 2023 01:35:56 +0000 (07:05 +0530)]
Simplify affine bound min or max with operands

Add canonicalization for affine.for bounds to use operand info (when
operands are outer loop affine.for IVs) to simplify bounds: redundant
bound expressions are eliminated in specific cases that are easy to
check and well-suited for op canonicaliation. If the lowest or the
highest value the affine expression can take is already covered by other
constant bounds, the expression can be removed.

Eg:
`min (d0) -> (32 * d0 + 32, 32)(%i) where 0 <= %i < 2`

The first expression can't be less than 32 and can be simplified away
with a lightweight local rewrite.

This simplification being part of canonicalization only handles simple
expressions, specifically, a sum of products of operands with constants.
This is a very common and a dominant case where such simplification is
desired. These can be flattened without any local variables.

Reviewed By: dcaballe, springerm

Differential Revision: https://reviews.llvm.org/D149007

18 months ago[HWASAN] Support short granules in __hwasan_test_shadow
Vitaly Buka [Sat, 29 Apr 2023 01:18:31 +0000 (18:18 -0700)]
[HWASAN] Support short granules in __hwasan_test_shadow

Reviewed By: thurston

Differential Revision: https://reviews.llvm.org/D149430

18 months agoRevert "[HWASAN] Support short granules in __hwasan_test_shadow"
Vitaly Buka [Sat, 29 Apr 2023 01:31:03 +0000 (18:31 -0700)]
Revert "[HWASAN] Support short granules in __hwasan_test_shadow"

Acidentally relanded without the fix.

This reverts commit 008c3934ab6b3a6dfcbb81f6258b9e03e9b61c1c.

18 months ago[HWASAN] Support short granules in __hwasan_test_shadow
Vitaly Buka [Sat, 29 Apr 2023 01:18:31 +0000 (18:18 -0700)]
[HWASAN] Support short granules in __hwasan_test_shadow

Reviewed By: thurston

Differential Revision: https://reviews.llvm.org/D149430

18 months agoRefactor and generalize AArch64 watchpoint support in debugserver
Jason Molenda [Sat, 29 Apr 2023 01:20:40 +0000 (18:20 -0700)]
Refactor and generalize AArch64 watchpoint support in debugserver

Refactor the debugserver watchpiont support in anticipating of
adding support for AArch64 MASK hardware watchpoints to watch
larger regions of memory.  debugserver already had support for
handling a request to watch an unaligned region of memory up
to 8 bytes using Byte Address Select watchpoints - it would split
an unaligned watch request into two aligned doublewords that
could be watched with two hardware watchpoints using the BAS
specification.

This patch generalizes that code for properly aligning, and
possibly splitting, a watchpoint request into two hardware watchpoints
to handle any size request.  And separates out the specifics
about BAS watchpoints into its own method, so a sibling method
for MASK watchpoints can be dropped in next.

Differential Revision: https://reviews.llvm.org/D149040
rdar://108233371

18 months agoRemove i386 and armv7 native support in debugserver
Jason Molenda [Sat, 29 Apr 2023 01:18:43 +0000 (18:18 -0700)]
Remove i386 and armv7 native support in debugserver

i386 and armv7 macOS/iOS cannot be built with current Xcode
any longer; we cannot build or test the support code for running
debugserver on these targets.  Remove the code.

Differential Revision: https://reviews.llvm.org/D149503

18 months ago[lldb] Refactor host::OpenFileInExternalEditor
Jonas Devlieghere [Fri, 28 Apr 2023 20:36:00 +0000 (13:36 -0700)]
[lldb] Refactor host::OpenFileInExternalEditor

This patch refactors the macOS implementation of
OpenFileInExternalEditor. It fixes an AppleEvent memory leak, the
caching of LLDB_EXTERNAL_EDITOR and speculatively fixes a crash when
CFURL is NULL (rdar://108633464). The new code also improves error
handling, readability and documents calls to the CoreFoundation Launch
Services APIs.

A bunch of the Launch Services APIs have been deprecated
(LSFindApplicationForInfo, LSOpenURLsWithRole). The preferred API is
LSOpenCFURLRef but it doesn't specifying the "location" Apple Event
which is used to highlight the current line and switching over would
regress the existing behavior.

rdar://108633464

Differential revision: https://reviews.llvm.org/D149482

18 months agoRevert "[HWASAN] Support short granules in __hwasan_test_shadow"
Vitaly Buka [Sat, 29 Apr 2023 00:38:11 +0000 (17:38 -0700)]
Revert "[HWASAN] Support short granules in __hwasan_test_shadow"

Break internal tests, but it supposed to be almost NFC.

This reverts commit bf12b746372a25c5a15fa9bd452bd469cc047b47.

18 months ago[RISCV] Add a DAG combine to fold (add (xor (setcc X, Y), 1) -1)->(neg (setcc X,...
Craig Topper [Fri, 28 Apr 2023 23:52:55 +0000 (16:52 -0700)]
[RISCV] Add a DAG combine to fold (add (xor (setcc X, Y), 1) -1)->(neg (setcc X, Y)).

18 months ago[mlir][Tensor] Dont drop attributes during tiling.
Mahesh Ravishankar [Fri, 28 Apr 2023 23:45:02 +0000 (23:45 +0000)]
[mlir][Tensor] Dont drop attributes during tiling.

Reviewed By: hanchung

Differential Revision: https://reviews.llvm.org/D149120

18 months ago[gn build] Port 2fa6bcf02f15
LLVM GN Syncbot [Fri, 28 Apr 2023 23:32:00 +0000 (23:32 +0000)]
[gn build] Port 2fa6bcf02f15

18 months agoFix use of undeclared identifier 'EC' (NFC)
Jie Fu [Fri, 28 Apr 2023 23:30:35 +0000 (07:30 +0800)]
Fix use of undeclared identifier 'EC' (NFC)

/Users/jiefu/llvm-project/llvm/lib/Object/GOFFObjectFile.cpp:199:3: error: use of undeclared identifier 'EC'
  EC = ConverterEBCDIC::convertToUTF8(SymbolName, SymbolNameConverted);
  ^
1 error generated.

18 months ago[msan] Improve handling of Intrinsic::is_fpclass after c55fffe
Vitaly Buka [Fri, 28 Apr 2023 22:13:52 +0000 (15:13 -0700)]
[msan] Improve handling of Intrinsic::is_fpclass after c55fffe

c55fffe replaced fcmp with fpclass.

```
declare i1 @llvm.is.fpclass(<fptype> <op>, i32 <test>)
declare <N x i1> @llvm.is.fpclass(<vector-fptype> <op>, i32 <test>)
```

Perfect fix will require checking bits of <op> corresponding to <test>
argument. For now just propagate shadow without reporting before
intrinsic. Still existing handling of fcmp is also simple OR, so it's
not making it worse.

Reviewed By: eugenis

Differential Revision: https://reviews.llvm.org/D149491

18 months ago[msan] Precommit test for regression from c55fffe
Vitaly Buka [Fri, 28 Apr 2023 21:52:27 +0000 (14:52 -0700)]
[msan] Precommit test for regression from c55fffe

Msan uses strict approach for unknown intrinsics like fpclass. So any
uninitialized bits in the value trigger msan report.