platform/upstream/llvm.git
2 years agoRevert "[Attributor] Ensure to use the proper liveness AA"
Mitch Phillips [Fri, 10 Jun 2022 20:56:29 +0000 (13:56 -0700)]
Revert "[Attributor] Ensure to use the proper liveness AA"

This reverts commit a3273c0c06003b9e5f53a856844fee42a17e6b7b.

Reason: Broke the ASan buildbots with a memory leak. See
https://reviews.llvm.org/rG94841c713fdd2bce3276015d1e946d414bb74ee8 for
more information.

2 years ago[InstCombine] switch synthetic unreachable to use undef instead of poison (NFC)
Nuno Lopes [Fri, 10 Jun 2022 20:54:09 +0000 (21:54 +0100)]
[InstCombine] switch synthetic unreachable to use undef instead of poison (NFC)

2 years ago[lldb/crashlog] Show help when the command is called without any argument
Med Ismail Bennani [Wed, 8 Jun 2022 23:21:24 +0000 (16:21 -0700)]
[lldb/crashlog] Show help when the command is called without any argument

This patch changes the `crashlog` command behavior to print the help
message if no argument was provided with the command.

rdar://94576026

Differential Revision: https://reviews.llvm.org/D127362

Signed-off-by: Med Ismail Bennani <medismail.bennani@gmail.com>
2 years ago[mlir][py-bindings] Fix include issue introduced by D127352
agostini01 [Fri, 10 Jun 2022 20:42:16 +0000 (20:42 +0000)]
[mlir][py-bindings] Fix include issue introduced by D127352

Using:
      -DMLIR_ENABLE_BINDINGS_PYTHON=ON

Resulted in a failed build due to changes implemented by
https://reviews.llvm.org/D127352

This updates the include line

Reviewed By: Mogball

Differential Revision: https://reviews.llvm.org/D127523

2 years ago[libc++] Granularize <iterator> includes
Nikolas Klauser [Fri, 10 Jun 2022 17:53:10 +0000 (19:53 +0200)]
[libc++] Granularize <iterator> includes

Reviewed By: ldionne, #libc

Spies: libcxx-commits, wenlei

Differential Revision: https://reviews.llvm.org/D127445

2 years ago[libc++] Simplify type_traits and use more builtins
Nikolas Klauser [Fri, 10 Jun 2022 09:38:04 +0000 (11:38 +0200)]
[libc++] Simplify type_traits and use more builtins

Reviewed By: ldionne, #libc

Spies: libcxx-commits

Differential Revision: https://reviews.llvm.org/D127226

2 years agoFix interaction of CFI instructions with MachineOutliner.
Eli Friedman [Fri, 10 Jun 2022 20:37:49 +0000 (13:37 -0700)]
Fix interaction of CFI instructions with MachineOutliner.

1. When checking if a candidate contains a CFI instruction, actually
iterate over all of the instructions, instead of stopping halfway
through.
2. Make sure copied CFI directives refer to the correct instruction.

Fixes https://github.com/llvm/llvm-project/issues/55842

Differential Revision: https://reviews.llvm.org/D126930

2 years ago[Lex] Fix `fixits` for typo-corrections of preprocessing directives within skipped...
Argyrios Kyrtzidis [Tue, 7 Jun 2022 20:35:17 +0000 (13:35 -0700)]
[Lex] Fix `fixits` for typo-corrections of preprocessing directives within skipped blocks

The `EndLoc` parameter was always unset so no fixit was emitted. But it is also unnecessary for determining the range so we can remove it.

Differential Revision: https://reviews.llvm.org/D127251

2 years ago[RISCV] Refine costs for i1 reductions
Philip Reames [Fri, 10 Jun 2022 20:21:52 +0000 (13:21 -0700)]
[RISCV] Refine costs for i1 reductions

Our actual lowering for i1 reductions uses ctpop combined with possibly a vector negate and possibly a logic op afterwards. I believe ctpop to be low cost on all reasonable hardware.

The default costing implementation here was returning quite inconsistent costs. and/or were returning very high costs (because we seem to think moving into scalar registers is very expensive?) and others were returning lower but still too high (because of the assumed tree reduce strategy). While we should probably improve the generic costing strategy for i1 vectors, let's start by fixing the immediate problem.

Differential Revision: https://reviews.llvm.org/D127511

2 years ago[RISCV] Implement isElementTypeLegalForScalableVector TTI hook
Philip Reames [Fri, 10 Jun 2022 20:15:11 +0000 (13:15 -0700)]
[RISCV] Implement isElementTypeLegalForScalableVector TTI hook

This brings us into alignment with AArch64, and in the process fixes a compiler crash bug in uniform store handling in the vectorizer.

Before the recent invalid cost bailout work, this would have also avoided crashes on invalid costs in some cases. I honestly think the vectorizer should gracefully bailout on uniform stores it can't use a scatter for, but it doesn't, so lets take the path of least resistance here. It's also possible that there are other vectorizer bugs AArch64 isn't seeing because of this hook; we don't want to be finding them either.

Differential Revision: https://reviews.llvm.org/D127514

2 years ago[PS4/PS5] Don't inherit base class alignment
Paul Robinson [Fri, 10 Jun 2022 20:02:17 +0000 (13:02 -0700)]
[PS4/PS5] Don't inherit base class alignment

2 years ago[RISCV] Minor test improvements for scalable scatter/gather tests added in 275b2e524
Philip Reames [Fri, 10 Jun 2022 20:13:43 +0000 (13:13 -0700)]
[RISCV] Minor test improvements for scalable scatter/gather tests added in 275b2e524

2 years ago[libc][Obvious] Change all __builtin_clz* calls to clz in builtin_wrappers.h.
Tue Ly [Fri, 10 Jun 2022 20:00:13 +0000 (16:00 -0400)]
[libc][Obvious] Change all __builtin_clz* calls to clz in builtin_wrappers.h.

2 years ago[RISCV] Don't require loop simplify form in RISCVGatherScatterLowering.
Craig Topper [Fri, 10 Jun 2022 19:51:01 +0000 (12:51 -0700)]
[RISCV] Don't require loop simplify form in RISCVGatherScatterLowering.

We need a preheader and a single latch, but we don't need a dedicated
exit.

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D127513

2 years ago[RISCV] Add test case showing failure to convert gather/scatter to strided load/store...
Craig Topper [Fri, 10 Jun 2022 19:50:52 +0000 (12:50 -0700)]
[RISCV] Add test case showing failure to convert gather/scatter to strided load/store. NFC

Our optimization pass checks for loop simplify form, before doing
the transform. The loops here aren't in loop simplify form because
the exit block has two predecessors.

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D127451

2 years ago[RISCV] Add cost model coverage for scalable scatter/gather
Philip Reames [Fri, 10 Jun 2022 19:45:54 +0000 (12:45 -0700)]
[RISCV] Add cost model coverage for scalable scatter/gather

2 years agoAdd documentation of new sanitizer-specific GV attributes.
Mitch Phillips [Fri, 10 Jun 2022 19:45:26 +0000 (12:45 -0700)]
Add documentation of new sanitizer-specific GV attributes.

Reviewed By: vitalybuka

Differential Revision: https://reviews.llvm.org/D126922

2 years agoAdd sanitizer-specific GlobalValue attributes.
Mitch Phillips [Fri, 10 Jun 2022 19:24:36 +0000 (12:24 -0700)]
Add sanitizer-specific GlobalValue attributes.

Plan is the migrate the global variable metadata for sanitizers, that's
currently carried around generally in the 'llvm.asan.globals' section,
onto the global variable itself.

This patch adds the attribute and plumbs it through the LLVM IR and
bitcode formats, but is a no-op other than that so far.

Reviewed By: vitalybuka, kstoimenov

Differential Revision: https://reviews.llvm.org/D126100

2 years agoRevert "[cmake] Don't export `LLVM_TOOLS_INSTALL_DIR` anymore"
John Ericson [Fri, 10 Jun 2022 19:26:12 +0000 (19:26 +0000)]
Revert "[cmake] Don't export `LLVM_TOOLS_INSTALL_DIR` anymore"

This reverts commit d5daa5c5b091cafb9b7ffd19b5dfa2daadef3229.

2 years ago[Clang] Added missing doc for minsize attribute
Dávid Bolvanský [Fri, 10 Jun 2022 19:17:52 +0000 (21:17 +0200)]
[Clang] Added missing doc for minsize attribute

Fixes https://github.com/llvm/llvm-project/issues/53226

2 years ago[libc++] Make sure we install the modulemap file with install-cxx-headers
Louis Dionne [Fri, 10 Jun 2022 19:14:57 +0000 (15:14 -0400)]
[libc++] Make sure we install the modulemap file with install-cxx-headers

2 years ago[clang][dataflow] Don't `assert` full LHS coverage in `optional` model
Sam Estep [Fri, 10 Jun 2022 19:08:41 +0000 (19:08 +0000)]
[clang][dataflow] Don't `assert` full LHS coverage in `optional` model

Followup to D127434.

Reviewed By: ymandel, sgatev

Differential Revision: https://reviews.llvm.org/D127502

2 years ago[JITLink][ELF][AArch64] Implement eh frame handling.
Sunho Kim [Fri, 10 Jun 2022 18:34:49 +0000 (03:34 +0900)]
[JITLink][ELF][AArch64] Implement eh frame handling.

Implements eh frame handling by using generic EHFrame passes. The c++ exception handling works correctly with this change.

Reviewed By: lhames

Differential Revision: https://reviews.llvm.org/D127063

2 years ago[RISCV] Add cost model coverage for mask reductions requiring legalization
Philip Reames [Fri, 10 Jun 2022 19:01:16 +0000 (12:01 -0700)]
[RISCV] Add cost model coverage for mask reductions requiring legalization

2 years ago[mlir][AMDGPU] Set ABI version constant when linking device libs
Krzysztof Drewniak [Thu, 2 Jun 2022 20:19:41 +0000 (20:19 +0000)]
[mlir][AMDGPU] Set ABI version constant when linking device libs

Currently, linking the device libraries requires setting a constant
that indicates the code object ABI version the compilation is
targeting.

This fixes the MLIR linking process by setting this constant to 400,
which is the value corresponding to the current code object ABI
default, version 4.

Reviewed By: Mogball

Differential Revision: https://reviews.llvm.org/D126913

2 years ago[OpenMP][IRBuilder] Add final clause to task
Shraiysh Vaishay [Fri, 10 Jun 2022 17:37:18 +0000 (23:07 +0530)]
[OpenMP][IRBuilder] Add final clause to task

This patch adds final clause to OpenMP IR Builder.

Reviewed By: Meinersbur

Differential Revision: https://reviews.llvm.org/D126626

2 years ago[libc++] Remove uses of __two in type_traits
Nikolas Klauser [Fri, 10 Jun 2022 10:42:46 +0000 (12:42 +0200)]
[libc++] Remove uses of __two in type_traits

Reviewed By: ldionne, #libc

Spies: libcxx-commits

Differential Revision: https://reviews.llvm.org/D127483

2 years agoRevert "[Sanitizers] Cleanup handling of stat64/statfs64"
Mariusz Borsa [Fri, 10 Jun 2022 16:40:17 +0000 (09:40 -0700)]
Revert "[Sanitizers] Cleanup handling of stat64/statfs64"

This reverts commit 6d890a0fb8b1962cf7b58e9e16ff4545f1f2b84f.

2 years agoUpdate *_TMPRING_SIZE.WAVESIZE for GFX11
Jay Foad [Tue, 5 Apr 2022 12:38:57 +0000 (13:38 +0100)]
Update *_TMPRING_SIZE.WAVESIZE for GFX11

The encoding of COMPUTE_TMPRING_SIZE.WAVESIZE and
SPI_TMPRING_SIZE.WAVESIZE has changed in GFX11: it is now in units
of 64 dwords instead of 256 dwords, and the field has been widened
from 13 bits to 15 bits.

Depends on D126989

Reviewed By: rampitec, arsenm, #amdgpu

Differential Revision: https://reviews.llvm.org/D127248

2 years ago[mlir][vector] Add patterns for vector distribution
Thomas Raoux [Mon, 6 Jun 2022 20:56:50 +0000 (20:56 +0000)]
[mlir][vector] Add patterns for vector distribution

Add pattern to hoist scalar code outside of warp distribute region as
those cannot be distributed and we would want to execute them on all
the lanes.
Add patterns to distribute transfer_write ops. Those operations can be
distributed in different ways and it is control by user.

Differential Revision: https://reviews.llvm.org/D127152

2 years ago[SystemZ/z/OS] Set DWARF version to 4 for z/OS.
Kai Nacke [Thu, 9 Jun 2022 16:54:40 +0000 (12:54 -0400)]
[SystemZ/z/OS] Set DWARF version to 4 for z/OS.

The DWARF version was raised to 5 for all platforms which do not opt
out. Default to DWARF version to 4 for z/OS again.

Reviewed By: abhina.sreeskantharajan, uweigand

Differential Revision: https://reviews.llvm.org/D127498

2 years ago[AMDGPU] gfx11 add bits to COMPUTE_PGM_RSRC3
Joe Nash [Mon, 6 Jun 2022 18:05:52 +0000 (14:05 -0400)]
[AMDGPU] gfx11 add bits to COMPUTE_PGM_RSRC3

Contributors:
Konstantin Zhuravlyov <kzhuravl_dev@outlook.com>

Patch 21/N for upstreaming of AMDGPU gfx11 architecture

Depends on D127143

Reviewed By: rampitec, #amdgpu, kzhuravl

Differential Revision: https://reviews.llvm.org/D127241

2 years ago[NFC] Suggest Release mode in clang GettingStarted.html
Shivam Gupta [Fri, 10 Jun 2022 05:29:50 +0000 (10:59 +0530)]
[NFC] Suggest Release mode in clang GettingStarted.html

This fix https://github.com/llvm/llvm-project/issues/23841.
Lots of beginners are not of aware of this option do suggesting it here
would be helpful.

2 years ago[clang][tablegen] adds human documentation to `WarningOption`
Christopher Di Bella [Wed, 1 Jun 2022 22:12:38 +0000 (22:12 +0000)]
[clang][tablegen] adds human documentation to `WarningOption`

Building on D126796, this commit adds the infrastructure for being able
to print out descriptions of what each warning does.

Differential Revision: https://reviews.llvm.org/D126832

2 years ago[AMDGPU] NFC. Comment change to GFX10+ in AsmParser
Joe Nash [Thu, 9 Jun 2022 14:32:27 +0000 (10:32 -0400)]
[AMDGPU] NFC. Comment change to GFX10+ in AsmParser

2 years ago[AMDGPU] gfx11 SRC_POPS_EXISTING_WAVE_ID is removed
Joe Nash [Mon, 6 Jun 2022 19:33:15 +0000 (15:33 -0400)]
[AMDGPU] gfx11 SRC_POPS_EXISTING_WAVE_ID is removed

2 years ago[libc] add buffering to FILE writes
Michael Jones [Wed, 1 Jun 2022 21:42:57 +0000 (14:42 -0700)]
[libc] add buffering to FILE writes

Previously all FILE objects were fully buffered, this patch adds line
buffering and unbuffered output, as well as applying them to stdout and
stderr.

Reviewed By: sivachandra

Differential Revision: https://reviews.llvm.org/D126829

2 years ago[libc][math] Add EXP_MANT_MASK when long double is double or quad
Benjamin Kramer [Fri, 10 Jun 2022 16:47:40 +0000 (18:47 +0200)]
[libc][math] Add EXP_MANT_MASK when long double is double or quad

2 years ago[InstCombine] add narrowing transform for low-masked binop with zext operand (2nd...
Sanjay Patel [Fri, 10 Jun 2022 14:44:13 +0000 (10:44 -0400)]
[InstCombine] add narrowing transform for low-masked binop with zext operand (2nd try)

The 1st try ( afa192cfb6049a15c55 ) was reverted because it could
cause an infinite loop with constant expressions.

A test for that and an extra condition to enable the transform
are added now. I also added code comments to better describe
the transform and the existing, related transform.

Original commit message:
https://alive2.llvm.org/ce/z/hRy3rE

As shown in D123408, we can produce this pattern when moving
casts around, and we already have a related fold for a binop
with a constant operand.

2 years ago[InstCombine] add test for mask op with constant expression; NFC
Sanjay Patel [Fri, 10 Jun 2022 14:36:41 +0000 (10:36 -0400)]
[InstCombine] add test for mask op with constant expression; NFC

This is reduced from the post-commit C example in:
afa192cfb6049a15

That was reverted with:
6fedc6a2b41e458f

This will infinite loop unless that transform or the ones
with casts that can "EvaluateInDifferentType" are limited
to ignore constant expressions.

2 years ago[SystemZ][z/OS] Add llvm.read_register() intrinsic support for zOS
Yusra Syeda [Fri, 10 Jun 2022 14:52:05 +0000 (10:52 -0400)]
[SystemZ][z/OS] Add llvm.read_register() intrinsic support for zOS

Differential Revision: https://reviews.llvm.org/D127412

2 years ago[RISCV] move `isFaultFirstLoad` into `RISCVInstrInfo`
Shao-Ce SUN [Fri, 10 Jun 2022 16:11:13 +0000 (00:11 +0800)]
[RISCV] move `isFaultFirstLoad` into `RISCVInstrInfo`

Fix build errors in D126794

```
ld.lld: error: undefined symbol: llvm::MachineInstr::getNumExplicitDefs() const
>>> referenced by RISCVBaseInfo.cpp
>>>               RISCVBaseInfo.cpp.o:(llvm::isFaultFirstLoad(llvm::MachineInstr const&)) in archive lib/libLLVMRISCVDesc.a

ld.lld: error: undefined symbol: llvm::MachineInstr::findRegisterDefOperandIdx(llvm::Register, bool, bool, llvm::TargetRegisterInfo const*) const
>>> referenced by RISCVBaseInfo.cpp
>>>               RISCVBaseInfo.cpp.o:(llvm::isFaultFirstLoad(llvm::MachineInstr const&)) in archive lib/libLLVMRISCVDesc.a
clang-15: error: linker command failed with exit code 1 (use -v to see invocation)
```

Reviewed By: fakepaper56, craig.topper

Differential Revision: https://reviews.llvm.org/D127477

2 years ago[AMDGPU] Restore documentation of .amdhsa_shared_vgpr_count
Jay Foad [Fri, 10 Jun 2022 16:05:29 +0000 (17:05 +0100)]
[AMDGPU] Restore documentation of .amdhsa_shared_vgpr_count

This was accidentally lost in D127402.

2 years agoRevert "[RISCV] move `isFaultFirstLoad` into `RISCVInstrInfo`"
Shao-Ce SUN [Fri, 10 Jun 2022 16:02:45 +0000 (00:02 +0800)]
Revert "[RISCV] move `isFaultFirstLoad` into `RISCVInstrInfo`"

This reverts commit e018e493c1ac514504bbaa1d1396aec025142a31.

There are some problems with this commit,
related revision: https://reviews.llvm.org/D127477

2 years ago[RISCV] Fix accidental deletion of test lines in 2247e4d
Philip Reames [Fri, 10 Jun 2022 15:57:27 +0000 (08:57 -0700)]
[RISCV] Fix accidental deletion of test lines in 2247e4d

Apparently, update_analyze_test_checks.sh does *not* warn on conflicting CHECKs, it just silently drops those lines from the generated test.  That is.. less than helpful.

2 years ago[PS5] Use linker scripting to find profiling data, like PS4
Paul Robinson [Thu, 9 Jun 2022 18:17:16 +0000 (11:17 -0700)]
[PS5] Use linker scripting to find profiling data, like PS4

2 years ago[RISCV] Mark FSIN and other math functions as Expand for scalable vectors.
Craig Topper [Fri, 10 Jun 2022 15:36:31 +0000 (08:36 -0700)]
[RISCV] Mark FSIN and other math functions as Expand for scalable vectors.

This prevents them from being assumed legal by the cost model.

This matches what is done for AArch64 SVE.

Reviewed By: reames

Differential Revision: https://reviews.llvm.org/D123799

2 years ago[RISCV] Use common prefixes to reduce duplication in cost model tests
Philip Reames [Fri, 10 Jun 2022 15:27:12 +0000 (08:27 -0700)]
[RISCV] Use common prefixes to reduce duplication in cost model tests

2 years ago[DCE] Update tests to use opaque pointers (NFC)
Nikita Popov [Fri, 10 Jun 2022 15:31:10 +0000 (17:31 +0200)]
[DCE] Update tests to use opaque pointers (NFC)

2 years ago[flang][lowering] Ignore compiler directives
Andrzej Warzynski [Fri, 10 Jun 2022 15:26:13 +0000 (15:26 +0000)]
[flang][lowering] Ignore compiler directives

This patch simply replaces a `TODO` with a warning.

This is part of the upstreaming effort from the `fir-dev` branch in [1].

[1] https://github.com/flang-compiler/f18-llvm-project

Co-authored-by: Eric Schweitz <eschweitz@nvidia.com>
Reviewed By: kiranchandramohan

Differential Revision: https://reviews.llvm.org/D127415

2 years ago[NFC][Alignment] Simplify code
Guillaume Chatelet [Fri, 10 Jun 2022 15:20:07 +0000 (15:20 +0000)]
[NFC][Alignment] Simplify code

2 years ago[LLDB][Docs] Fix formatting of example code-block
Adrian Prantl [Fri, 10 Jun 2022 15:16:30 +0000 (08:16 -0700)]
[LLDB][Docs] Fix formatting of example code-block

Tested by building the docs-lldb-html target and
confirming the code-block renders properly with
the fix.

Patch by Michael Buch!

Differential Revision: https://reviews.llvm.org/D127437

2 years ago[InstSimplify] Convert tests to opaque pointers (NFC)
Nikita Popov [Fri, 10 Jun 2022 15:07:27 +0000 (17:07 +0200)]
[InstSimplify] Convert tests to opaque pointers (NFC)

The only interesting test change is in @PR31262, where the following
fold is now performed, while it previously was not:
https://alive2.llvm.org/ce/z/a5Qmr6

llvm/test/Transforms/InstSimplify/ConstProp/gep.ll has not been
updated, because there is a tradeoff between folding and inrange
preservation there that we may want to discuss.

Updates have been performed using:
https://gist.github.com/nikic/98357b71fd67756b0f064c9517b62a34

2 years ago[NFC] Remove deadcode
Guillaume Chatelet [Fri, 10 Jun 2022 15:13:25 +0000 (15:13 +0000)]
[NFC] Remove deadcode

2 years ago[RISCV] Broaden cost model coverage for fixed vectors w/i1 element type
Philip Reames [Fri, 10 Jun 2022 15:07:55 +0000 (08:07 -0700)]
[RISCV] Broaden cost model coverage for fixed vectors w/i1 element type

2 years ago[DAG] SimplifyDemandedVectorElts - adding SimplifyMultipleUseDemandedVectorElts handl...
Simon Pilgrim [Fri, 10 Jun 2022 14:49:49 +0000 (15:49 +0100)]
[DAG] SimplifyDemandedVectorElts - adding SimplifyMultipleUseDemandedVectorElts handling to ISD::CONCAT_VECTORS

Attempt to look through multiple use operands of ISD::CONCAT_VECTORS nodes

Another minor improvement for D127115

2 years ago[InstSimplify] Name variables/labels in test (NFC)
Nikita Popov [Fri, 10 Jun 2022 14:56:13 +0000 (16:56 +0200)]
[InstSimplify] Name variables/labels in test (NFC)

Run the test through -instnamer. Also drop irrelevant uwtable
attribute.

2 years ago[InstSimplify] Regenerate test checks (NFC)
Nikita Popov [Fri, 10 Jun 2022 14:52:53 +0000 (16:52 +0200)]
[InstSimplify] Regenerate test checks (NFC)

2 years ago[clang][dataflow] In `optional` model, match call return via hasType
Sam Estep [Fri, 10 Jun 2022 14:50:22 +0000 (14:50 +0000)]
[clang][dataflow] In `optional` model, match call return via hasType

Currently the unchecked-optional-access model fails on this example:

    #include <memory>
    #include <optional>

    void foo() {
      std::unique_ptr<std::optional<float>> x;
      *x = std::nullopt;
    }

You can verify the failure by saving the file as `foo.cpp` and running this command:

    clang-tidy -checks='-*,bugprone-unchecked-optional-access' foo.cpp -- -std=c++17

The failing `assert` is in the `transferAssignment` function of the `UncheckedOptionalAccessModel.cpp` file:

    assert(OptionalLoc != nullptr);

The symptom can be treated by replacing that `assert` with an early `return`:

    if (OptionalLoc == nullptr)
      return;

That would be better anyway since we cannot expect to always cover all possible LHS expressions, but it is out of scope for this patch and left as a followup.

Note that the failure did not occur on this very similar example:

    #include <optional>

    template <typename T>
    struct smart_ptr {
      T& operator*() &;
      T* operator->();
    };

    void foo() {
      smart_ptr<std::optional<float>> x;
      *x = std::nullopt;
    }

The difference is caused by the `isCallReturningOptional` matcher, which was previously checking the `functionDecl` of the `callee`. This patch changes it to instead use `hasType` directly on the call expression, fixing the failure for the `std::unique_ptr` example above.

Reviewed By: sgatev

Differential Revision: https://reviews.llvm.org/D127434

2 years ago[cmake] Don't export `LLVM_TOOLS_INSTALL_DIR` anymore
John Ericson [Sat, 8 Jan 2022 09:00:53 +0000 (09:00 +0000)]
[cmake] Don't export `LLVM_TOOLS_INSTALL_DIR` anymore

First of all, `LLVM_TOOLS_INSTALL_DIR` put there breaks our NixOS
builds, because `LLVM_TOOLS_INSTALL_DIR` defined the same as
`CMAKE_INSTALL_BINDIR` becomes an *absolute* path, and then when
downstream projects try to install there too this breaks because our
builds always install to fresh directories for isolation's sake.

Second of all, note that `LLVM_TOOLS_INSTALL_DIR` stands out against the
other specially crafted `LLVM_CONFIG_*` variables substituted in
`llvm/cmake/modules/LLVMConfig.cmake.in`.

@beanz added it in d0e1c2a550ef348aae036d0fe78cab6f038c420c to fix a
dangling reference in `AddLLVM`, but I am suspicious of how this
variable doesn't follow the pattern.

Those other ones are carefully made to be build-time vs install-time
variables depending on which `LLVMConfig.cmake` is being generated, are
carefully made relative as appropriate, etc. etc. For my NixOS use-case
they are also fine because they are never used as downstream install
variables, only for reading not writing.

To avoid the problems I face, and restore symmetry, I deleted the
exported and arranged to have many `${project}_TOOLS_INSTALL_DIR`s.
`AddLLVM` now instead expects each project to define its own, and they
do so based on `CMAKE_INSTALL_BINDIR`. `LLVMConfig` still exports
`LLVM_TOOLS_BINARY_DIR` which is the location for the tools defined in
the usual way, matching the other remaining exported variables.

For the `AddLLVM` changes, I tried to copy the existing pattern of
internal vs non-internal or for LLVM vs for downstream function/macro
names, but it would good to confirm I did that correctly.

Reviewed By: nikic

Differential Revision: https://reviews.llvm.org/D117977

2 years ago[AMDGPU] gfx11 EXECZ and VCCZ are no longer allowed to be used as
Joe Nash [Mon, 6 Jun 2022 19:33:15 +0000 (15:33 -0400)]
[AMDGPU] gfx11 EXECZ and VCCZ are no longer allowed to be used as

sources to SALU and VALU instructions.

Contributors:
Baptiste Saleil <baptiste.saleil@amd.com>

Patch 20/N for upstreaming of AMDGPU gfx11 architecture

Depends on D126989

Reviewed By: rampitec, foad, #amdgpu

Differential Revision: https://reviews.llvm.org/D127143

2 years ago[NFC] remove 'br undef' from InstCombine test cases
Nuno Lopes [Fri, 10 Jun 2022 14:28:57 +0000 (15:28 +0100)]
[NFC] remove 'br undef' from InstCombine test cases
This is UB and allows the compiler to give any result, so these tests weren't meaningful
InstCombine tests are now clean of 'br undef'

2 years ago[Clang,ARM] Add release note for D127197.
Simon Tatham [Fri, 10 Jun 2022 14:18:06 +0000 (15:18 +0100)]
[Clang,ARM] Add release note for D127197.

I should have put that in the original commit, but @lenary only just
reminded me that it needed to be there.

2 years ago[InstCombine] Add additional cost of phi tests (NFC)
Nikita Popov [Fri, 10 Jun 2022 13:48:34 +0000 (15:48 +0200)]
[InstCombine] Add additional cost of phi tests (NFC)

These check for undesirable folds for a cast of a loop phi.

2 years ago[libc++][NFC] Fix indentation of some #if directives
Louis Dionne [Fri, 10 Jun 2022 14:11:24 +0000 (10:11 -0400)]
[libc++][NFC] Fix indentation of some #if directives

2 years ago[AMDGPU] Basic implementation of isExtractSubvectorCheap
Jay Foad [Tue, 7 Jun 2022 16:21:00 +0000 (17:21 +0100)]
[AMDGPU] Basic implementation of isExtractSubvectorCheap

Add a basic implementation of isExtractSubvectorCheap that only
considers extracts at offset 0.

Differential Revision: https://reviews.llvm.org/D127385

2 years ago[TableGen][CodeEmitterGen] Do not crash on insufficient positional instruction operands.
Ivan Kosarev [Fri, 10 Jun 2022 13:18:29 +0000 (14:18 +0100)]
[TableGen][CodeEmitterGen] Do not crash on insufficient positional instruction operands.

Reviewed By: foad

Differential Revision: https://reviews.llvm.org/D126288

2 years ago[SimplifyCFG] Share code to compute switch density between ShouldBuildLookupTable...
Hans Wennborg [Wed, 8 Jun 2022 13:08:46 +0000 (15:08 +0200)]
[SimplifyCFG] Share code to compute switch density between ShouldBuildLookupTable() and ReduceSwitchRange()

They're computing the same thing. No functionality change.

Differential revision: https://reviews.llvm.org/D127482

2 years ago[clang] Add support for __builtin_memset_inline
Guillaume Chatelet [Tue, 7 Jun 2022 09:51:32 +0000 (09:51 +0000)]
[clang] Add support for __builtin_memset_inline

In the same spirit as D73543 and in reply to https://reviews.llvm.org/D126768#3549920 this patch is adding support for `__builtin_memset_inline`.

The idea is to get support from the compiler to easily write efficient memory function implementations.

This patch could be split in two:
 - one for the LLVM part adding the `llvm.memset.inline.*` intrinsics.
 - and another one for the Clang part providing the instrinsic as a builtin.

Differential Revision: https://reviews.llvm.org/D126903

2 years ago[NFC] Add 3 more -inseltpoison.ll test variations
Nuno Lopes [Fri, 10 Jun 2022 13:06:32 +0000 (14:06 +0100)]
[NFC] Add 3 more -inseltpoison.ll test variations

2 years ago[RISCV] move `isFaultFirstLoad` into `RISCVInstrInfo`
Shao-Ce SUN [Fri, 10 Jun 2022 08:53:34 +0000 (16:53 +0800)]
[RISCV] move `isFaultFirstLoad` into `RISCVInstrInfo`

Fix build errors in D126794

```
ld.lld: error: undefined symbol: llvm::MachineInstr::getNumExplicitDefs() const
>>> referenced by RISCVBaseInfo.cpp
>>>               RISCVBaseInfo.cpp.o:(llvm::isFaultFirstLoad(llvm::MachineInstr const&)) in archive lib/libLLVMRISCVDesc.a

ld.lld: error: undefined symbol: llvm::MachineInstr::findRegisterDefOperandIdx(llvm::Register, bool, bool, llvm::TargetRegisterInfo const*) const
>>> referenced by RISCVBaseInfo.cpp
>>>               RISCVBaseInfo.cpp.o:(llvm::isFaultFirstLoad(llvm::MachineInstr const&)) in archive lib/libLLVMRISCVDesc.a
clang-15: error: linker command failed with exit code 1 (use -v to see invocation)
```

Reviewed By: fakepaper56

Differential Revision: https://reviews.llvm.org/D127477

2 years agoRevert "[analyzer] Deprecate `-analyzer-store region` flag"
Nico Weber [Fri, 10 Jun 2022 12:36:34 +0000 (08:36 -0400)]
Revert "[analyzer] Deprecate `-analyzer-store region` flag"

This reverts commit d50d9946d1d7e5f20881f0bc71fbd025040b1c96.
Broke check-clang, see comments on https://reviews.llvm.org/D126067

Also revert dependent change "[analyzer] Deprecate the unused 'analyzer-opt-analyze-nested-blocks' cc1 flag"
This reverts commit 07b4a6d0461fe64e10d30029ed3d598e49ca3810.

Also revert "[analyzer] Fix buildbots after introducing a new frontend warning"
This reverts commit 90374df15ddc58d823ca42326a76f58e748f20eb.
(See https://reviews.llvm.org/rG90374df15ddc58d823ca42326a76f58e748f20eb)

2 years ago[analyzer] Fix buildbots after introducing a new frontend warning
Balazs Benics [Fri, 10 Jun 2022 12:47:23 +0000 (14:47 +0200)]
[analyzer] Fix buildbots after introducing a new frontend warning

It seems like I should have ran the `check-clang` target after introducing a new warning diagnostic entry.
My bad. I'll run it next time; `check-clang-analysis` was not enough.
Here is a link to the broken bot: http://45.33.8.238/linux/78236/step_7.txt

This commit should fix this.

Differential Revision: https://reviews.llvm.org/D126067

2 years ago[X86] combineMulToPMADDWD - don't bitcast the source ops before splitting to ensure...
Simon Pilgrim [Fri, 10 Jun 2022 12:29:04 +0000 (13:29 +0100)]
[X86] combineMulToPMADDWD - don't bitcast the source ops before splitting to ensure we split the build vectors early

Fixes a regression on D127115 - splitting was creating extract_subvector(bitcast(build_vector())) patterns which prevented the build vectors being split before being bitcast to vXi16 types, resulting in various issues with further folding of the (now legal) build vectors

2 years ago[NFC] Use proper getSliceAlign type in SROA
Guillaume Chatelet [Fri, 10 Jun 2022 12:36:40 +0000 (12:36 +0000)]
[NFC] Use proper getSliceAlign type in SROA

2 years agoRevert "[InstCombine] add narrowing transform for low-masked binop with zext operand"
Sanjay Patel [Fri, 10 Jun 2022 12:12:02 +0000 (08:12 -0400)]
Revert "[InstCombine] add narrowing transform for low-masked binop with zext operand"

This reverts commit afa192cfb6049a15c5542d132d500b910b802c74.
This can cause an infinite loop as shown with an example in the
post-commit thread.

2 years ago[AMDGPU][GFX9][GFX10] Support base+soffset+offset SMEM atomics.
Ivan Kosarev [Fri, 10 Jun 2022 09:06:45 +0000 (10:06 +0100)]
[AMDGPU][GFX9][GFX10] Support base+soffset+offset SMEM atomics.

Resolves a part of
https://github.com/llvm/llvm-project/issues/38652

Reviewed By: dp

Differential Revision: https://reviews.llvm.org/D127314

2 years ago[InstCombine] Use +0.0 instead of -0.0 as the FP identity for some folds
David Sherwood [Tue, 31 May 2022 14:54:49 +0000 (15:54 +0100)]
[InstCombine] Use +0.0 instead of -0.0 as the FP identity for some folds

In foldSelectIntoOp we sometimes transform a select of a fadd into a
fadd of a select, where we select between data and an identity value.
For both fadd and fsub the identity is always -0.0, but if the nsz
flag is set on the select instruction we can use +0.0 instead. Doing
so then triggers other optimisations, such as when folding the select
of masked load into a new masked load.

Differential Revision: https://reviews.llvm.org/D126774

2 years ago[flang] Fix RUN line in test from 3e782ba
Diana Picus [Fri, 10 Jun 2022 11:36:52 +0000 (11:36 +0000)]
[flang] Fix RUN line in test from 3e782ba

The current RUN line leaves a parse-ir-error.s file around, which might
confuse the buildbots. Direct the output into the void instead (although
it would probably be better to direct it to FileCheck).

2 years ago[NFC] update clang-tools-extra README.txt
Shivam Gupta [Fri, 10 Jun 2022 11:25:55 +0000 (16:55 +0530)]
[NFC] update clang-tools-extra README.txt

2 years ago[flang][runtime] Remove dependency on C++ <mutex> on Windows
Markus Mützel [Fri, 10 Jun 2022 08:31:42 +0000 (08:31 +0000)]
[flang][runtime] Remove dependency on C++ <mutex> on Windows

The implementation of the Lock class on Windows currently uses C++
mutexes. That introduces a dependency on the C++ runtime on that
platform.

Use a Windows CriticalSection instead of a std::mutex to avoid that
dependency.

This works for me with MinGW (tested in a CLANG64 environment of MSYS2).

See also D126291.

Differential Revision: https://reviews.llvm.org/D127316

2 years ago[clang-tidy] Remove reference to CSA AnalysisStoreOpt field
Balazs Benics [Fri, 10 Jun 2022 11:21:28 +0000 (13:21 +0200)]
[clang-tidy] Remove reference to CSA AnalysisStoreOpt field

D126215 removed the `AnalysisStoreOpt` field, so `clang-tidy` should not
refer to it.
http://45.33.8.238/linux/78232/step_4.txt

  ../../clang-tools-extra/clang-tidy/ClangTidy.cpp:444:22: error: no
member named 'AnalysisStoreOpt' in 'clang::AnalyzerOptions'
      AnalyzerOptions->AnalysisStoreOpt = RegionStoreModel;
      ~~~~~~~~~~~~~~~~~^
  ../../clang-tools-extra/clang-tidy/ClangTidy.cpp:444:41: error: use of
undeclared identifier 'RegionStoreModel'
      AnalyzerOptions->AnalysisStoreOpt = RegionStoreModel;

This one liner patch should resolve the build failure.

Differential Revision: https://reviews.llvm.org/D126215

2 years ago[analyzer] Deprecate the unused 'analyzer-opt-analyze-nested-blocks' cc1 flag
Balazs Benics [Fri, 10 Jun 2022 11:09:37 +0000 (13:09 +0200)]
[analyzer] Deprecate the unused 'analyzer-opt-analyze-nested-blocks' cc1 flag

This flag was introduced by
https://github.com/llvm/llvm-project/commit/6818991d7165f68fe196922d9e5c6648dd57cc47
    commit 6818991d7165f68fe196922d9e5c6648dd57cc47
    Author: Ted Kremenek <kremenek@apple.com>
    Date:   Mon Dec 7 22:06:12 2009 +0000

  Add clang-cc option '-analyzer-opt-analyze-nested-blocks' to treat
  block literals as an entry point for analyzer checks.

The last reference was removed by this commit:
https://github.com/llvm/llvm-project/commit/5c32dfc5fb1cfcff8ae3671284e17daa8da3a188

    commit 5c32dfc5fb1cfcff8ae3671284e17daa8da3a188
    Author: Anna Zaks <ganna@apple.com>
    Date:   Fri Dec 21 01:19:15 2012 +0000

  [analyzer] Add blocks and ObjC messages to the call graph.
  This paves the road for constructing a better function dependency graph.
  If we analyze a function before the functions it calls and inlines,
  there is more opportunity for optimization.
  Note, we add call edges to the called methods that correspond to
  function definitions (declarations with bodies).

Consequently, we should remove this dead flag.
However, this arises a couple of burning questions.
 - Should the `cc1` frontend still accept this flag - to keep
   tools/users passing this flag directly to `cc1` (which is unsupported,
   unadvertised) working.
 - If we should remain backward compatible, how long?
 - How can we get rid of deprecated and obsolete flags at some point?

Reviewed By: martong

Differential Revision: https://reviews.llvm.org/D126067

2 years ago[mlir] add a dynamic user-after-parent-freed transform dialect check
Alex Zinenko [Fri, 10 Jun 2022 09:57:59 +0000 (11:57 +0200)]
[mlir] add a dynamic user-after-parent-freed transform dialect check

In the transform dialect, a transform IR handle may be pointing to a payload IR
operation that is an ancestor of another payload IR operation pointed to by
another handle. If such a "parent" handle is consumed by a transformation, this
indicates that the associated operation is likely rewritten, which in turn
means that the "child" handle may now be associated with a dangling pointer or
a pointer to a different operation than originally. Add a handle invalidation
mechanism to guard against such situations by reporting errors at runtime.

Reviewed By: springerm

Differential Revision: https://reviews.llvm.org/D127480

2 years ago[analyzer][NFC] Remove unused RegionStoreFeatures
Balazs Benics [Fri, 10 Jun 2022 11:02:26 +0000 (13:02 +0200)]
[analyzer][NFC] Remove unused RegionStoreFeatures

Reviewed By: martong

Differential Revision: https://reviews.llvm.org/D126216

2 years ago[flang][driver] Make `flang-new -fc1` accept MLIR files
Andrzej Warzynski [Wed, 1 Jun 2022 16:00:31 +0000 (16:00 +0000)]
[flang][driver] Make `flang-new -fc1` accept MLIR files

This relatively small change will allow Flang's frontend driver,
`flang-new -fc1`, to consume and parse MLIR files.  Semantically (i.e.
from user's perspective) this is identical to reading LLVM IR files.

Two file extensions are associated with MLIR files: .fir and .mlir. Note
that reading MLIR files makes only sense when running one of the
code-generation actions, i.e. when using one of the following action
flags: -S, -emit-obj, -emit-llvm, -emit-llvm-bc.

The majority of tests that required `tco` to run are updated to also run
with `flang-new -fc1`. A few tests are updated to use `fir-opt` instead
of `tco` (that's the preferred choice when testing a particular MLIR
pass). basic-program.fir is not updated as that test is intended to
verify the behaviour of `tco` specifically.

Differential Revision: https://reviews.llvm.org/D126890

2 years ago[analyzer] Deprecate `-analyzer-store region` flag
Balazs Benics [Fri, 10 Jun 2022 10:57:15 +0000 (12:57 +0200)]
[analyzer] Deprecate `-analyzer-store region` flag

I'm trying to remove unused options from the `Analyses.def` file, then
merge the rest of the useful options into the `AnalyzerOptions.def`.
Then make sure one can set these by an `-analyzer-config XXX=YYY` style
flag.
Then surface the `-analyzer-config` to the `clang` frontend;

After all of this, we can pursue the tablegen approach described
https://discourse.llvm.org/t/rfc-tablegen-clang-static-analyzer-engine-options-for-better-documentation/61488

In this patch, I'm proposing flag deprecations.
We should support deprecated analyzer flags for exactly one release. In
this case I'm planning to drop this flag in `clang-16`.

In the clang frontend, now we won't pass this option to the cc1
frontend, rather emit a warning diagnostic reminding the users about
this deprecated flag, which will be turned into error in clang-16.

Unfortunately, I had to remove all the tests referring to this flag,
causing a mass change. I've also added a test for checking this warning.

I've seen that `scan-build` also uses this flag, but I think we should
remove that part only after we turn this into a hard error.

Reviewed By: martong

Differential Revision: https://reviews.llvm.org/D126215

2 years ago[FuncSpec]Fix assertion failure when value is not added to solver
Bin Cheng [Fri, 10 Jun 2022 10:40:59 +0000 (18:40 +0800)]
[FuncSpec]Fix assertion failure when value is not added to solver

This patch improves the fix in D110529 to prevent from crashing on value
with byval attribute that is not added in SCCP solver.

Authored-by: sinan.lin@linux.alibaba.com
Reviewed By: ChuanqiXu

Differential Revision: https://reviews.llvm.org/D126355

2 years ago[AMDGPU][MC][GFX1013] Enable image_msaa_load
Dmitry Preobrazhensky [Fri, 10 Jun 2022 10:37:57 +0000 (13:37 +0300)]
[AMDGPU][MC][GFX1013] Enable image_msaa_load

Differential Revision: https://reviews.llvm.org/D127198

2 years ago[gn build] Port ff6d5dee713c
LLVM GN Syncbot [Fri, 10 Jun 2022 10:37:02 +0000 (10:37 +0000)]
[gn build] Port ff6d5dee713c

2 years ago[gn build] Port 1972d1e86a62
LLVM GN Syncbot [Fri, 10 Jun 2022 10:37:01 +0000 (10:37 +0000)]
[gn build] Port 1972d1e86a62

2 years ago[flang][driver] Fix support for `-x`
Andrzej Warzynski [Mon, 6 Jun 2022 17:57:33 +0000 (17:57 +0000)]
[flang][driver] Fix support for `-x`

Until now, `-x` wasn't really taken into account in Flang's compiler and
frontend drivers. `flang-new` and `flang-new -fc1` only recently gained
powers to consume inputs other than Fortran files and that's probably
why this hasn't been noticed yet.

This patch makes sure that `-x` is supported correctly and consistently
with Clang. To this end, verification is added when reading LLVM IR
files (i.e. IR modules are verified with `llvm::verifyModule`). This
way, LLVM IR parsing errors are correctly reported to Flang users. This
also aids testing.

With the new functionality, we can verify that `-x ir` breaks
compilation for e.g. Fortran files and vice-versa. Tests are updated
accordingly.

Differential Revision: https://reviews.llvm.org/D127207

2 years ago[ARM] Fix how size-0 bitfields affect homogeneous aggregates.
Simon Tatham [Fri, 10 Jun 2022 10:25:53 +0000 (11:25 +0100)]
[ARM] Fix how size-0 bitfields affect homogeneous aggregates.

By both AAPCS32 and AAPCS64, the test for whether an aggregate
qualifies as homogeneous (either HFA or HVA) is based on the data
layout alone. So any logical member of the structure that does not
affect the data layout also should not affect homogeneity. In
particular, an empty bitfield ('int : 0') should make no difference.

In fact, clang considered it to make a difference in C but not in C++,
and justified that policy as compatible with gcc. But that's
considered a bug in gcc as well (at least for Arm targets), and it's
fixed in gcc 12.1.

This fix mimics gcc's: zero-sized bitfields are now ignored in all
languages for the Arm (32- and 64-bit) ABIs. But I've left the
previous behaviour unchanged in other ABIs, by means of adding an
ABIInfo::isZeroLengthBitfieldPermittedInHomogeneousAggregate query
method which the Arm subclasses override.

Reviewed By: lenary

Differential Revision: https://reviews.llvm.org/D127197

2 years ago[gn build] (manually) port 6bc8163c79fd
Nico Weber [Fri, 10 Jun 2022 10:26:31 +0000 (06:26 -0400)]
[gn build] (manually) port 6bc8163c79fd

2 years ago[analyzer] Print the offending function at EndAnalysis crash
Balazs Benics [Fri, 10 Jun 2022 10:21:17 +0000 (12:21 +0200)]
[analyzer] Print the offending function at EndAnalysis crash

I've faced crashes in the past multiple times when some
`check::EndAnalysis` callback caused some crash.
It's really anoying that it doesn't tell which function triggered this
callback.

This patch adds the well-known trace for that situation as well.
Example:
  1.      <eof> parser at end of file
  2.      While analyzing stack:
          #0 Calling test11

Note that this does not have tests.
I've considered `unittests` for this purpose, by using the
`ASSERT_DEATH()` similarly how we check double eval called functions in
`ConflictingEvalCallsTest.cpp`, however, that the testsuite won't invoke
the custom handlers. Only the message of the `llvm_unreachable()` will
be printed. Consequently, it's not applicable for us testing this
feature.

I've also considered using an end-to-end LIT test for this.
For that, we would need to somehow overload the `clang_analyzer_crash()`
`ExprInspection` handler, to get triggered by other events than the
`EvalCall`. I'm not saying that we could not come up with a generic way
of causing crash in a specific checker callback, but I'm not sure if
that would worth the effort.

Reviewed By: martong

Differential Revision: https://reviews.llvm.org/D127389

2 years ago[MVE] Fold fadd(select(..., +0.0)) into a predicated fadd
David Sherwood [Wed, 8 Jun 2022 08:25:00 +0000 (09:25 +0100)]
[MVE] Fold fadd(select(..., +0.0)) into a predicated fadd

We already have patterns for matching fadd(select(..., -0.0)),
but an upcoming patch will lead to patterns using +0.0 as the
identity instead of -0.0. I'm adding support for these patterns
now to avoid any regressions for MVE.

Differential Revision: https://reviews.llvm.org/D127275

2 years ago[libc++] Granularize even more of type_traits
Nikolas Klauser [Fri, 10 Jun 2022 08:48:39 +0000 (10:48 +0200)]
[libc++] Granularize even more of type_traits

Reviewed By: ldionne, #libc

Spies: libcxx-commits, mgorny

Differential Revision: https://reviews.llvm.org/D126593

2 years ago[libc++] Implement ranges::replace{, _if}
Nikolas Klauser [Tue, 7 Jun 2022 07:42:10 +0000 (09:42 +0200)]
[libc++] Implement ranges::replace{, _if}

Reviewed By: var-const, #libc

Spies: libcxx-commits, mgorny

Differential Revision: https://reviews.llvm.org/D126283

2 years ago[LoopInfo] Add getOutermostLoop() (NFC)
Nikita Popov [Sat, 10 Apr 2021 17:26:50 +0000 (19:26 +0200)]
[LoopInfo] Add getOutermostLoop() (NFC)

This is a recurring pattern, add an API function for it.

2 years ago[AggressiveInstcombine] Conditionally fold saturated fptosi to llvm.fptosi.sat
David Green [Fri, 10 Jun 2022 08:36:09 +0000 (09:36 +0100)]
[AggressiveInstcombine] Conditionally fold saturated fptosi to llvm.fptosi.sat

This adds a fold for aggressive instcombine that converts
smin(smax(fptosi(x))) into a llvm.fptosi.sat, providing that the
saturation constants are correct and the cost of the llvm.fptosi.sat is
lower.

Unfortunately, a llvm.fptosi.sat cannot always be converted back to a
smin/smax/fptosi. The llvm.fptosi.sat intrinsic is more defined that the
original, which produces poison if the original fptosi was out of range.
The llvm.fptosi.sat will saturate any value, so needs to be expanded to
a fptosi(fpmin(fpmax(x))), which can be worse for codegeneration
depending on the target.

So this change thais conditional on the backend reporting that the
llvm.fptosi.sat is cheaper that the original smin+smax+fptost.  This is
a change to the way that AggressiveInstrcombine has worked in the past.
Instead of just being a canonicalization pass, that canonicalization can
be dependant on the target in certain specific cases.

Differential Revision: https://reviews.llvm.org/D125755