platform/upstream/llvm.git
3 years ago[AArch64][NFC] Fix failing cost-model test
Irina Dobrescu [Wed, 2 Jun 2021 13:59:21 +0000 (14:59 +0100)]
[AArch64][NFC] Fix failing cost-model test

3 years ago[LV] Build and cost VPlans for scalable VFs.
Sander de Smalen [Wed, 2 Jun 2021 13:10:59 +0000 (14:10 +0100)]
[LV] Build and cost VPlans for scalable VFs.

This patch uses the calculated maximum scalable VFs to build VPlans,
cost them and select a suitable scalable VF.

Reviewed By: paulwalker-arm

Differential Revision: https://reviews.llvm.org/D98722

3 years ago[compiler-rt][lsan] Increase libdl_deadlock test timeout
David Spickett [Wed, 2 Jun 2021 13:30:37 +0000 (13:30 +0000)]
[compiler-rt][lsan] Increase libdl_deadlock test timeout

We have been seeing this test fail intermittently on our
2 stage AArch64 bot.

As far back as https://lab.llvm.org/buildbot/#/builders/53/builds/2694

Likely due to a lack of resources at certain times on the
shared machine. Up the time limit to give us some more room.

(this limit only applies to the watchdog thread, so if the
test passes then it won't take 20s)

3 years ago[PowerPC][AIX} FIx AIX bootstrap build.
Sean Fertile [Tue, 1 Jun 2021 17:39:02 +0000 (13:39 -0400)]
[PowerPC][AIX} FIx AIX bootstrap build.

A recent patch:
https://reviews.llvm.org/rGe0921655b1ff8d4ba7c14be59252fe05b705920e
changed clangs AIX bitfield handling to use 4-byte bitfield containers,
matching XLs behavior. This change triggers static assert failures when
bootstrapping. Change the macro we check to enable bitfield packing on
AIX to `__clang__` which is defined by both xlclang and clang.

Differential Revision: https://reviews.llvm.org/D103474

3 years ago[LV] NFC: Remove redundant isLegalMasked(Gather|Scatter) functions.
Sander de Smalen [Tue, 1 Jun 2021 20:05:33 +0000 (21:05 +0100)]
[LV] NFC: Remove redundant isLegalMasked(Gather|Scatter) functions.

This NFC change follows from conversation in D102437, where it was discussed
to remove these functions as a separate patch.

3 years ago[LV] NFC: Replace custom getMemInstValueType by llvm::getLoadStoreType.
Sander de Smalen [Tue, 1 Jun 2021 14:30:00 +0000 (15:30 +0100)]
[LV] NFC: Replace custom getMemInstValueType by llvm::getLoadStoreType.

llvm::getLoadStoreType was added recently and has the same implementation
as 'getMemInstValueType' in LoopVectorize.cpp. Since there is no
value in having two implementations, this patch removes the custom LV
implementation in favor of the generic one defined in Instructions.h.

3 years ago[TTI] NFC: Change getIntImmCodeSizeCost to return InstructionCost.
Daniil Fukalov [Thu, 20 May 2021 13:57:19 +0000 (16:57 +0300)]
[TTI] NFC: Change getIntImmCodeSizeCost to return InstructionCost.

This patch migrates the TTI cost interfaces to return an InstructionCost.

See this patch for the introduction of the type: https://reviews.llvm.org/D91174
See this thread for context: http://lists.llvm.org/pipermail/llvm-dev/2020-November/146408.html

Reviewed By: sdesmalen

Differential Revision: https://reviews.llvm.org/D102915

3 years ago[compiler-rt][asan] Enable unwind-tables for Arm Linux
David Spickett [Tue, 1 Jun 2021 14:54:33 +0000 (14:54 +0000)]
[compiler-rt][asan] Enable unwind-tables for Arm Linux

Since https://reviews.llvm.org/D102046 some tests have
been falling back to fast unwinding on our Thumb bot.

This fails because fast unwinding does not work on Thumb.
By adding the extra information we ensure this does not happen
during testing, but the built library can still fast unwind
as a last resort.

Since there are some situations it can work in, like if
eveything is built with clang. During testing we've got gcc
built system libs and clang built tests.

The same change was made for sanitizer-common in
https://reviews.llvm.org/D96337.

Reviewed By: zatrazz

Differential Revision: https://reviews.llvm.org/D103463

3 years ago[mlir][linalg] Update result position calculation in the Structured Op Interface...
Tobias Gysi [Wed, 2 Jun 2021 12:29:50 +0000 (12:29 +0000)]
[mlir][linalg] Update result position calculation in the Structured Op Interface (NFC).

Remove two unused methods and replace the implementation of getResultsPositionInLoopsToShapeMap. The patch is based on https://reviews.llvm.org/D103394.

Differential Revision: https://reviews.llvm.org/D103397

3 years ago[mlir][linalg] Cleanup LinalgOp usage in fusion on tensors (NFC).
Tobias Gysi [Wed, 2 Jun 2021 11:55:38 +0000 (11:55 +0000)]
[mlir][linalg] Cleanup LinalgOp usage in fusion on tensors (NFC).

Replace the uses of deprecated Structured Op Interface methods in FusionOnTensors.cpp. This patch is based on https://reviews.llvm.org/D103394.

Differential Revision: https://reviews.llvm.org/D103471

3 years ago[RISCV][NFC] Add '+mattr=+experimental-v' to RVV test
Fraser Cormack [Wed, 2 Jun 2021 10:39:53 +0000 (11:39 +0100)]
[RISCV][NFC] Add '+mattr=+experimental-v' to RVV test

3 years ago[AArch64] Optimise bitreverse lowering in ISel
Irina Dobrescu [Tue, 25 May 2021 18:19:52 +0000 (19:19 +0100)]
[AArch64] Optimise bitreverse lowering in ISel

Differential Revision: https://reviews.llvm.org/D103105

3 years ago[AMDGPU][Libomptarget][NFC] Remove bunch of dead structs
Pushpinder Singh [Wed, 2 Jun 2021 07:24:06 +0000 (07:24 +0000)]
[AMDGPU][Libomptarget][NFC] Remove bunch of dead structs

Dropped structs are atmi_machine_t, atmi_device_t and atmi_memory_t

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D103509

3 years ago[lld/mac] Implement -reexport_framework, -reexport_library, -reexport-l
Nico Weber [Wed, 2 Jun 2021 00:17:04 +0000 (20:17 -0400)]
[lld/mac] Implement -reexport_framework, -reexport_library, -reexport-l

These are slightly easier-to-use versions of -sub_library and -sub_umbrella.

Differential Revision: https://reviews.llvm.org/D103497

3 years ago[AMDGPU][Libomptarget][NFC] Remove atmi_place_t
Pushpinder Singh [Wed, 2 Jun 2021 07:22:40 +0000 (07:22 +0000)]
[AMDGPU][Libomptarget][NFC] Remove atmi_place_t

atmi_place_t has been replaced with int DeviceId.

Reviewed By: JonChesterfield

Differential Revision: https://reviews.llvm.org/D103508

3 years ago[mailmap] Use my chromium address as my canonical email address
Nico Weber [Wed, 2 Jun 2021 10:34:39 +0000 (06:34 -0400)]
[mailmap] Use my chromium address as my canonical email address

3 years agoAdd a toplevel .mailmap file
Nico Weber [Sat, 29 May 2021 16:18:15 +0000 (12:18 -0400)]
Add a toplevel .mailmap file

See "Proposal: Adding a toplevel .mailmap file" on llvm-dev:
https://lists.llvm.org/pipermail/llvm-dev/2021-May/150741.html

Differential Revision: https://reviews.llvm.org/D103360

3 years ago[SimpleLoopUnswitch] Port partially invariant unswitch from LoopUnswitch to SimpleLoo...
Jingu Kang [Fri, 21 May 2021 11:34:40 +0000 (12:34 +0100)]
[SimpleLoopUnswitch] Port partially invariant unswitch from LoopUnswitch to SimpleLoopUnswitch

This re-enables commit 107d19eb017ff6734986af077eb2e9f6600114a9 with bug fixes.

Differential Revision: https://reviews.llvm.org/D99354

3 years ago[InstCombine][msp430] Pre-commit test case for @llvm.powi and 16-bit ints
Bjorn Pettersson [Fri, 21 May 2021 11:42:50 +0000 (13:42 +0200)]
[InstCombine][msp430] Pre-commit test case for @llvm.powi and 16-bit ints

This is a pre-commit of a test case D99439 which is a patch that
updates @llvm.powi to handle different int sizes for the exponent.

Problem is that @llvm.powi is used as an IR construct that maps
to RT libcalls to __powi* functions, and those lib functions depend
on sizeof(int) to use correct type for the exponent.

The test cases show that we use i32 for the powi expenent, which
later would result in wrong type being used in libcalls (miscompile).

But there are also a couple of the negative test cases that show
that we rewrite into using powi when having a uitofp conversion
from i16, which would be wrong when doing the libcall as an
"unsigned int" isn't guaranteed to fit inside the "int" argument
in the called libcall function.

Differential Revision: https://reviews.llvm.org/D102919

3 years ago[CodeGen] Refactor libcall lookups for RTLIB::POWI_*
Bjorn Pettersson [Mon, 24 May 2021 21:17:18 +0000 (23:17 +0200)]
[CodeGen] Refactor libcall lookups for RTLIB::POWI_*

Use RuntimeLibcalls to get a common way to pick correct RTLIB::POWI_*
libcall for a given value type.

This includes a small refactoring of ExpandFPLibCall and
ExpandArgFPLibCall in SelectionDAGLegalize to share a bit of code,
plus adding an ExpandFPLibCall version that can be called directly
when expanding FPOWI/STRICT_FPOWI to ensure that we actually use
the same RTLIB::Libcall when expanding the libcall as we used when
checking the legality of such a call by doing a getLibcallName check.

Differential Revision: https://reviews.llvm.org/D103050

3 years ago[LegalizeTypes] Avoid promotion of exponent in FPOWI
Bjorn Pettersson [Fri, 21 May 2021 22:24:25 +0000 (00:24 +0200)]
[LegalizeTypes] Avoid promotion of exponent in FPOWI

The FPOWI DAG node is normally lowered to a libcall to one of the
RTLIB::POWI* runtime functions and the exponent should normally
have a type matching sizeof(int) when making the call. Thus,
type promotion of the exponent could lead to an FPOWI with a type
for the second operand that would be incorrect when doing the
libcall (a situation which would be hard to detect post-legalization
if we allow such FPOWI nodes).

This patch is changing DAGTypeLegalizer::PromoteIntOp_FPOWI to
do the rewrite into a libcall directly instead of promoting the
operand. This way we can check that the exponent is smaller than
sizeof(int) and we can let TargetLowering handle promotion as
part of making the libcall. It could be noticed here that makeLibCall
has some knowledge about targets such as 64-bit RISCV, for which the
libcall argument should be extended to a type larger than sizeof(int).

Differential Revision: https://reviews.llvm.org/D102950

3 years ago[SimplifyLibCalls] Take size of int into consideration when emitting ldexp/ldexpf
Bjorn Pettersson [Fri, 26 Mar 2021 11:49:38 +0000 (12:49 +0100)]
[SimplifyLibCalls] Take size of int into consideration when emitting ldexp/ldexpf

When rewriting
  powf(2.0, itofp(x)) -> ldexpf(1.0, x)
  exp2(sitofp(x)) -> ldexp(1.0, sext(x))
  exp2(uitofp(x)) -> ldexp(1.0, zext(x))

the wrong type was used for the second argument in the ldexp/ldexpf
libc call, for target architectures with 16 bit "int" type.
The transform incorrectly used a bitcasted function pointer with
a 32-bit argument when emitting the ldexp/ldexpf call for such
targets.

The fault is solved by using the correct function prototype
in the call, by asking TargetLibraryInfo about the size of "int".
TargetLibraryInfo by default derives the size of the int type by
assuming that it is 16 bits for 16-bit architectures, and
32 bits otherwise. If this isn't true for a target it should be
possible to override that default in the TargetLibraryInfo
initializer.

Differential Revision: https://reviews.llvm.org/D99438

3 years ago[mlir] Add DivOp lowering from Complex dialect to Standard/Math dialect.
Adrian Kuegel [Wed, 2 Jun 2021 08:34:18 +0000 (10:34 +0200)]
[mlir] Add DivOp lowering from Complex dialect to Standard/Math dialect.

Differential Revision: https://reviews.llvm.org/D103507

3 years ago[Demangle][Rust] Parse binders
Tomasz Miąsko [Wed, 2 Jun 2021 08:26:05 +0000 (10:26 +0200)]
[Demangle][Rust] Parse binders

Reviewed By: dblaikie

Differential Revision: https://reviews.llvm.org/D102729

3 years ago[RISCV] Expand unaligned fixed-length vector memory accesses
Fraser Cormack [Fri, 14 May 2021 12:56:41 +0000 (13:56 +0100)]
[RISCV] Expand unaligned fixed-length vector memory accesses

RVV vectors must be aligned to their element types, so anything less is
unaligned.

For regular loads and stores, our custom-lowering of fixed-length
vectors meant that we opted out of LegalizeDAG's built-in unaligned
expansion. This patch adds that logic in to our custom lower function.

For masked intrinsics, we declare that anything unaligned is not legal,
leaving the ScalarizeMaskedMemIntrin pass to do the expansion for us.

Note that neither of these methods can handle the expansion of
scalable-vector memory ops, so those cases are left alone by this patch.
Scalable loads and stores already go through expansion by default but
hit an assertion, and scalable masked intrinsics will silently generate
incorrect code. It may be prudent to return an error in both of these
cases.

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D102493

3 years ago[flang] Add tests for REPEAT. NFC
Diana Picus [Mon, 31 May 2021 11:42:42 +0000 (11:42 +0000)]
[flang] Add tests for REPEAT. NFC

These should already pass with the current implementation.

Differential Revision: https://reviews.llvm.org/D103402

3 years ago[NFC] Fix 'Load' name masking.
Daniil Fukalov [Tue, 1 Jun 2021 12:29:39 +0000 (15:29 +0300)]
[NFC] Fix 'Load' name masking.

Reviewed By: mkazantsev

Differential Revision: https://reviews.llvm.org/D103456

3 years ago[NFC][msan] Fix assigned-unused warning
Vitaly Buka [Wed, 2 Jun 2021 07:08:46 +0000 (00:08 -0700)]
[NFC][msan] Fix assigned-unused warning

3 years agoRevert "[NFC][msan] Fix warning on sanitizer-ppc64le-linux bot"
Vitaly Buka [Wed, 2 Jun 2021 07:01:36 +0000 (00:01 -0700)]
Revert "[NFC][msan] Fix warning on sanitizer-ppc64le-linux bot"

This fix breaks the test.

This reverts commit 6a2807bc81771d3fed397378507ecab40f8bff27.

3 years ago[mlir][linalg] Cleanup LinalgOp usage in sparse compiler (NFC).
Tobias Gysi [Wed, 2 Jun 2021 06:21:13 +0000 (06:21 +0000)]
[mlir][linalg] Cleanup LinalgOp usage in sparse compiler (NFC).

Replace the uses of deprecated Structured Op Interface methods in Sparsification.cpp. This patch is based on https://reviews.llvm.org/D103394.

Differential Revision: https://reviews.llvm.org/D103436

3 years agoResubmit D85085 after fixing the tests that were failing.
Sriraman Tallam [Tue, 1 Jun 2021 19:38:32 +0000 (12:38 -0700)]
Resubmit D85085 after fixing the tests that were failing.

D85085 was pushed earlier but broke tests on mac and win:
http://lab.llvm.org:8080/green/job/clang-stage1-RA/21182/consoleFull#-706149783d489585b-5106-414a-ac11-3ff90657619c

Recommitting it after adding mtriple to the llc commands.

Emit correct location lists with basic block sections.

This patch addresses multiple things:

1) It ensures that const_value is emitted when possible with basic block
   sections.
2) It emits location lists such that the labels are always within the
   section boundary.
3) It fixes a bug when the parameter is first used in a non-entry block
   which is in a different section from the entry block.

Differential Revision: https://reviews.llvm.org/D85085

3 years ago[lldb/API] Expose triple for SBProcessInfo.
Bruce Mitchener [Sun, 30 May 2021 10:28:09 +0000 (17:28 +0700)]
[lldb/API] Expose triple for SBProcessInfo.

This is present when doing a `platform process list` and is
tracked by the underlying code. To do something like the
process list via the SB API in the future, this must be
exposed.

Differential Revision: https://reviews.llvm.org/D103375

3 years ago[NFC][msan] Fix warning on sanitizer-ppc64le-linux bot
Vitaly Buka [Wed, 2 Jun 2021 04:03:41 +0000 (21:03 -0700)]
[NFC][msan] Fix warning on sanitizer-ppc64le-linux bot

3 years ago[scudo] Enabled MTE in tests
Vitaly Buka [Thu, 27 May 2021 17:10:50 +0000 (10:10 -0700)]
[scudo] Enabled MTE in tests

Reviewed By: pcc, hctim

Differential Revision: https://reviews.llvm.org/D103305

3 years agoRevert "Fix tmp files being left on Windows builds." for now;
Amy Huang [Wed, 2 Jun 2021 02:51:47 +0000 (19:51 -0700)]
Revert "Fix tmp files being left on Windows builds." for now;
causing some asan test failures.

This reverts commit 7daa18215905c831e130c7542f17619e9d936dfc.

3 years ago[libc++] Add a CI job to test libc++ when building for 32 bit
Louis Dionne [Wed, 2 Dec 2020 21:40:31 +0000 (16:40 -0500)]
[libc++] Add a CI job to test libc++ when building for 32 bit

Differential Revision: https://reviews.llvm.org/D92508

3 years ago[RISCV] Improve register allocation for masked vwadd(u).wv, vwsub(u).wv, vfwadd.wv...
Craig Topper [Wed, 2 Jun 2021 01:29:13 +0000 (18:29 -0700)]
[RISCV] Improve register allocation for masked vwadd(u).wv, vwsub(u).wv, vfwadd.wv, and vfwsub.wv.

The first source has the same EEW as the destination, but we're
using earlyclobber which prevents them from ever being the same
register.

To workaround this, add a special TIED pseudo to use whenever the
first source and merge operand are the same value. This allows
us to use a single operand for the merge operand and first source
which we can then tie to the destination. A tied source disables
earlyclobber for that operand.

Reviewed By: arcbbb

Differential Revision: https://reviews.llvm.org/D103211

3 years ago[gn build] Port 924ea3bb53ca
LLVM GN Syncbot [Wed, 2 Jun 2021 01:47:33 +0000 (01:47 +0000)]
[gn build] Port 924ea3bb53ca

3 years ago[libc++] NFC: Move unwrap_iter to its own header
Louis Dionne [Mon, 31 May 2021 16:08:05 +0000 (12:08 -0400)]
[libc++] NFC: Move unwrap_iter to its own header

This re-applies 9968896cd62a, which was reverted in b13edf6e907b because
it broke the build.

Differential Revision: https://reviews.llvm.org/D103369

3 years ago[mlir] Support tensor types in unrolled VectorToSCF
Matthias Springer [Wed, 2 Jun 2021 01:38:59 +0000 (10:38 +0900)]
[mlir] Support tensor types in unrolled VectorToSCF

Differential Revision: https://reviews.llvm.org/D102668

3 years ago[llvm-readobj] Print function names with `--bb-addr-map`.
Rahman Lavaee [Wed, 26 May 2021 20:36:44 +0000 (13:36 -0700)]
[llvm-readobj] Print function names with `--bb-addr-map`.

This patch uses the `getSymbolIndexForFunctionAddress` helper function to print function names for BB address map entries.

Reviewed By: jhenderson

Differential Revision: https://reviews.llvm.org/D102900

3 years ago[mlir] Support tensor types in non-unrolled VectorToSCF
Matthias Springer [Wed, 2 Jun 2021 01:29:07 +0000 (10:29 +0900)]
[mlir] Support tensor types in non-unrolled VectorToSCF

Support for tensor types in the unrolled version will follow in a separate commit.

Add a new pass option to activate lowering of transfer ops with tensor types (default: deactivated).

Differential Revision: https://reviews.llvm.org/D102666

3 years ago[CUDA][HIP] Promote const variables to constant
Yaxun (Sam) Liu [Tue, 25 May 2021 18:06:08 +0000 (14:06 -0400)]
[CUDA][HIP] Promote const variables to constant

Recently we added diagnosing ODR-use of host variables
in device functions, which includes ODR-use of const
host variables since they are not really emitted on
device side. This caused regressions since we used
to allow ODR-use of const host variables in device
functions.

This patch allows ODR-use of const variables in device
functions if the const variables can be statically initialized
and have an empty dtor. Such variables are marked with
implicit constant attrs and emitted on device side. This is
in line with what clang does for constexpr variables.

Reviewed by: Artem Belevich

Differential Revision: https://reviews.llvm.org/D103108

3 years agoMake ignore counts work as "after stop" modifiers so they play nicely with conditions
Jim Ingham [Tue, 1 Jun 2021 23:14:08 +0000 (16:14 -0700)]
Make ignore counts work as "after stop" modifiers so they play nicely with conditions

Previously ignore counts were checked when we stopped to do the sync callback in Breakpoint::ShouldStop. That meant we would do all the ignore count work even when
there is also a condition says the breakpoint should not stop.

That's wrong, lldb treats breakpoint hits that fail the thread or condition checks as "not having hit the breakpoint". So the ignore count check should happen after
the condition and thread checks in StopInfoBreakpoint::PerformAction.

The one side-effect of doing this is that if you have a breakpoint with a synchronous callback, it will run the synchronous callback before checking the ignore count.
That is probably a good thing, since this was already true of the condition and thread checks, so this removes an odd asymmetry. And breakpoints with sync callbacks
are all internal lldb breakpoints and there's not a really good reason why you would want one of these to use an ignore count (but not a condition or thread check...)

Differential Revision https://reviews.llvm.org/D103217

3 years ago[RISCV][test] Add new tests of bitwise and with constant for the Zbs extension
Ben Shi [Sat, 29 May 2021 10:18:45 +0000 (18:18 +0800)]
[RISCV][test] Add new tests of bitwise and with constant for the Zbs extension

These tests will show how (and r i) will be optimized to
(BCLRI (BCLRI r, i0), i1) or (BCLRI (ANDI r, i0), i1) by future
commits.

Reviewed By: craig.topper

Differential Revision: https://reviews.llvm.org/D103359

3 years ago[CUDA][HIP] Change default lang std to c++14
Yaxun (Sam) Liu [Thu, 27 May 2021 03:11:33 +0000 (23:11 -0400)]
[CUDA][HIP] Change default lang std to c++14

Currently clang and nvcc use c++14 as default std for C++.
gcc 11 even uses c++17 as default std for C++. However,
clang uses c++98 as default std for CUDA/HIP.

As c++14 has been well adopted and became default for
clang, it seems reasonable to use c++14 as default std
for CUDA/HIP.

Reviewed by: Artem Belevich

Differential Revision: https://reviews.llvm.org/D103221

3 years agoRemove x86 test amx-fast-tile-config.mir (by its author)
Xiang1 Zhang [Wed, 2 Jun 2021 00:29:29 +0000 (08:29 +0800)]
Remove x86 test amx-fast-tile-config.mir (by its author)

This test contains a lot of manual changes which is not convenient
to update, and the checks are duplicated with test amx-configO2toO0.ll

3 years agoFix tmp files being left on Windows builds.
Amy Huang [Tue, 4 May 2021 18:28:28 +0000 (11:28 -0700)]
Fix tmp files being left on Windows builds.

Clang writes object files by first writing to a .tmp file and then
renaming to the final .obj name. On Windows, if a compile is killed
partway through the .tmp files don't get deleted.

Currently it seems like RemoveFileOnSignal takes care of deleting the
tmp files on Linux, but on Windows we need to call
setDeleteDisposition on tmp files so that they are deleted when
closed.

This patch switches to using TempFile to create the .tmp files we write
when creating object files, since it uses setDeleteDisposition on Windows.
This change applies to both Linux and Windows for consistency.

Differential Revision: https://reviews.llvm.org/D102876

3 years ago[AMDGPU] All GWS instructions need aligned VGPR on gfx90a
Stanislav Mekhanoshin [Tue, 25 May 2021 23:37:01 +0000 (16:37 -0700)]
[AMDGPU] All GWS instructions need aligned VGPR on gfx90a

Fixes: SWDEV-288006

Differential Revision: https://reviews.llvm.org/D103197

3 years ago[OpaquePtr] Create API to make a copy of a PointerType with some address space
Arthur Eubanks [Thu, 27 May 2021 03:32:04 +0000 (20:32 -0700)]
[OpaquePtr] Create API to make a copy of a PointerType with some address space

Some existing places use getPointerElementType() to create a copy of a
pointer type with some new address space.

Reviewed By: dblaikie

Differential Revision: https://reviews.llvm.org/D103429

3 years ago[mlir-reduce] Reducer refactor.
Chia-hung Duan [Tue, 1 Jun 2021 23:00:19 +0000 (07:00 +0800)]
[mlir-reduce] Reducer refactor.

* A Reducer is a kind of RewritePattern, so it's just the same as
writing graph rewrite.
* ReductionTreePass operates on Operation rather than ModuleOp, so that
* we are able to reduce a nested structure(e.g., module in module) by
* self-nesting.

Reviewed By: jpienaar, rriddle

Differential Revision: https://reviews.llvm.org/D101046

3 years ago[InstSimplify] Treat invariant group insts as bitcasts for load operands
Arthur Eubanks [Thu, 22 Apr 2021 20:03:09 +0000 (13:03 -0700)]
[InstSimplify] Treat invariant group insts as bitcasts for load operands

We can look through invariant group intrinsics for the purposes of
simplifying the result of a load.

Since intrinsics can't be constants, but we also don't want to
completely rewrite load constant folding, we convert the load operand to
a constant. For GEPs and bitcasts we just treat them as constants. For
invariant group intrinsics, we treat them as a bitcast.

Reviewed By: lebedev.ri

Differential Revision: https://reviews.llvm.org/D101103

3 years ago[test] Precommit test for D101103
Arthur Eubanks [Tue, 1 Jun 2021 23:30:48 +0000 (16:30 -0700)]
[test] Precommit test for D101103

3 years ago[lld/mac] Make -t work correctly with -flat_namespace
Nico Weber [Mon, 31 May 2021 22:20:22 +0000 (18:20 -0400)]
[lld/mac] Make -t work correctly with -flat_namespace

We used to not print dylibs referenced by other dylibs in `-t` mode. This
affected reexports, and with `-flat_namespace` also just dylibs loaded by
dylibs. Now we print them.

Fixes PR49514.

Differential Revision: https://reviews.llvm.org/D103428

3 years ago[clang][Fuchsia] Turn on relative-vtables by default for Fuchsia
Leonard Chan [Wed, 12 May 2021 17:37:37 +0000 (10:37 -0700)]
[clang][Fuchsia] Turn on relative-vtables by default for Fuchsia

All fuchsia targets will now use the relative-vtables ABI by default.
Also remove -fexperimental-relative-c++-abi-vtables from test RUNs targeting fuchsia.

Differential Revision: https://reviews.llvm.org/D102374

3 years ago[Clang] -Wunused-but-set-parameter and -Wunused-but-set-variable
Michael Benfield [Tue, 1 Jun 2021 21:49:34 +0000 (14:49 -0700)]
[Clang] -Wunused-but-set-parameter and -Wunused-but-set-variable

These are intended to mimic warnings available in gcc.

Differential Revision: https://reviews.llvm.org/D100581

3 years ago[various] Remove or use variables which are unused but set.
Michael Benfield [Tue, 1 Jun 2021 21:48:42 +0000 (14:48 -0700)]
[various] Remove or use variables which are unused but set.

This is in preparation for the -Wunused-but-set-variable warning.

Differential Revision: https://reviews.llvm.org/D102942

3 years ago[lld-macho][nfc] Remove unnecessary use of Optional<T*>
Vy Nguyen [Tue, 1 Jun 2021 20:09:41 +0000 (16:09 -0400)]
[lld-macho][nfc] Remove unnecessary use of Optional<T*>

In all of these cases, the functions could simply return a nullptr instead of {}.
There is no case where Optional<nullptr> has a special meaning.

Differential Revision: https://reviews.llvm.org/D103489

3 years ago[CSE] Ask DominanceInfo about "hasSSADominance" instead of reconstructing it.
Chris Lattner [Tue, 1 Jun 2021 22:15:28 +0000 (15:15 -0700)]
[CSE] Ask DominanceInfo about "hasSSADominance" instead of reconstructing it.

I backed this off to make the previous patch easier to wrangle, but now
this is an efficient query and it is better to not replace it in CSE.

Differential Revision: https://reviews.llvm.org/D103494

3 years ago[lld/mac] Emit only one LC_LOAD_DYLIB per dylib
Nico Weber [Tue, 1 Jun 2021 19:13:46 +0000 (15:13 -0400)]
[lld/mac] Emit only one LC_LOAD_DYLIB per dylib

In some cases, we end up with several distinct DylibFiles that
have the same install name. Only emit a single LC_LOAD_DYLIB in
those cases.

This happens in 3 cases I know of:

1. Some tbd files are symlinks. libpthread.tbd is a symlink against
   libSystem.tbd for example, so `-lSystem -lpthread` loads
   libSystem.tbd twice. We could (and maybe should) cache loaded
   dylibs by realpath() to catch this.

2. Some tbd files are copies of each other. For example,
   CFNetwork.framework/CFNetwork.tbd and
   CFNetwork.framework/Versions/A/CFNetwork.tbd are two distinct
   copies of the same file. The former is found by
   `-framework CFNetwork` and the latter by the reexport in
   CoreServices.tbd. We could conceivably catch this by
   making `-framework` search look in `Versions/Current` instead
   of in the root, and/or by using a content hash to cache
   tbd files, but that's starting to sound complicated.

3. Magic $ld$ symbol processing can change the install name of
   a dylib based on the target platform_version. Here, two
   truly distinct dylibs can have the same install name.

So we need this code to deal with (3) anyways. Might as well use
it for 1 and 2, at least for now :)

With this (and D103430), clang-format links in the same dylibs
when linked with lld and ld64.

Differential Revision: https://reviews.llvm.org/D103488

3 years ago[Fuchsia] Add compat multilibs to cache file
Leonard Chan [Tue, 1 Jun 2021 22:14:03 +0000 (15:14 -0700)]
[Fuchsia] Add compat multilibs to cache file

Differential Revision: https://reviews.llvm.org/D103477

3 years ago[Dominators] Rewrite the dominator implementation for efficiency. NFC.
Chris Lattner [Mon, 31 May 2021 01:02:51 +0000 (18:02 -0700)]
[Dominators] Rewrite the dominator implementation for efficiency. NFC.

The previous impl densely scanned the entire region starting with an op
when dominators were created, creating a DominatorTree for every region.

This is extremely expensive up front -- particularly for clients like
Linalg/Transforms/Fusion.cpp that construct DominanceInfo for a single
query.  It is also extremely memory wasteful for IRs that use single
block regions commonly (e.g. affine.for) because it's making a
dominator tree for a region that has trivial dominance.  The
implementation also had numerous unnecessary minor efficiencies, e.g.
doing multiple walks of the region tree or tryGetBlocksInSameRegion
building a DenseMap that it didn't need.

This patch switches to an approach where [Post]DominanceInfo is free
to construct, and which lazily constructs DominatorTree's for any
multiblock regions that it needs.  This avoids the up-front cost
entirely, making its runtime proportional to the complexity of the
region tree instead of # ops in a region.  This also avoids the memory
and time cost of creating DominatorTree's for single block regions.

Finally this rewrites the implementation for simplicity and to avoids
the constant factor problems the old implementation had.

Differential Revision: https://reviews.llvm.org/D103384

3 years ago[lld][WebAssemlby] Fix for string merging of -dwarf-5 sections
Sam Clegg [Tue, 1 Jun 2021 19:48:21 +0000 (12:48 -0700)]
[lld][WebAssemlby] Fix for string merging of -dwarf-5 sections

We were mistakenly treating `.debug_str_offsets` as a string mergable
section when it is not (it contains integers not strings).  This is an
indication that we really should find a way to store flags for custom
sections.

Fixes: https://bugs.llvm.org/show_bug.cgi?id=48828
Fixes: https://bugs.chromium.org/p/chromium/issues/detail?id=1172217

Differential Revision: https://reviews.llvm.org/D103486

3 years ago[gn build] Port 065cf3f9d703
LLVM GN Syncbot [Tue, 1 Jun 2021 21:08:31 +0000 (21:08 +0000)]
[gn build] Port 065cf3f9d703

3 years ago[libcxx][ranges] Add `default_sentinel` and `default_sentinel_t`.
zoecarver [Tue, 1 Jun 2021 19:55:33 +0000 (12:55 -0700)]
[libcxx][ranges] Add `default_sentinel` and `default_sentinel_t`.

Refs https://eel.is/c++draft/default.sentinel and https://eel.is/c++draft/iterator.synopsis

Differential Revision: https://reviews.llvm.org/D103487

3 years agoFix a Clang diagnostic to start with a lowercase letter; NFC
Aaron Ballman [Tue, 1 Jun 2021 21:03:00 +0000 (17:03 -0400)]
Fix a Clang diagnostic to start with a lowercase letter; NFC

3 years agofixup: Missing operator in [globalisel][legalizer] Separate the deprecated LegalizerI...
Daniel Sanders [Tue, 1 Jun 2021 20:57:46 +0000 (13:57 -0700)]
fixup: Missing operator in [globalisel][legalizer] Separate the deprecated LegalizerInfo from the current one

My local compiler was fine with it but the bots complain about ambiguous types.

3 years ago[lld/mac] fix test failure after 24979e111
Nico Weber [Tue, 1 Jun 2021 20:34:10 +0000 (16:34 -0400)]
[lld/mac] fix test failure after 24979e111

If there is an error reading the dylib, we shouldn't try
to load its reexports.

Caught e.g. by https://lab.llvm.org/buildbot/#/builders/36/builds/8946

3 years ago[libc++] Fix the std_include.sh.cpp test that has been failing since c9385297ce
Louis Dionne [Tue, 1 Jun 2021 17:58:17 +0000 (13:58 -0400)]
[libc++] Fix the std_include.sh.cpp test that has been failing since c9385297ce

The compiler used on Apple bots doesn't know about -std=c++20 yet, so
we can't use that just yet.

Differential Revision: https://reviews.llvm.org/D103475

3 years ago[gn build] Port aaac268285ff
LLVM GN Syncbot [Tue, 1 Jun 2021 20:28:25 +0000 (20:28 +0000)]
[gn build] Port aaac268285ff

3 years ago[mlir] Updated depthwise conv to support kernel dilation
Rob Suderman [Thu, 27 May 2021 01:02:30 +0000 (18:02 -0700)]
[mlir] Updated depthwise conv to support kernel dilation

Depthwise convolution should support kernel dilation and non-dilation should
not be a special case. Updated op definition to include a dilation attribute.

This also adds a tosa.depthwise_conv2d lowering to linalg to support the new
linalg behavior.

Differential Revision: https://reviews.llvm.org/D103219

3 years ago[globalisel][legalizer] Separate the deprecated LegalizerInfo from the current one
Daniel Sanders [Sat, 29 May 2021 02:36:56 +0000 (19:36 -0700)]
[globalisel][legalizer] Separate the deprecated LegalizerInfo from the current one

It's still in use in a few places so we can't delete it yet but there's not
many at this point.

Differential Revision: https://reviews.llvm.org/D103352

3 years agoConvert TableGen assert to error
Stephen Neuendorffer [Wed, 28 Apr 2021 04:34:35 +0000 (21:34 -0700)]
Convert TableGen assert to error

This gives a nice message about the location of errors in a large
tablegen file, which is much more useful for users

Differential Revision: https://reviews.llvm.org/D102740

3 years ago[NFC][OpaquePtr] Explicitly pass GEP source type to IRBuilder in more places
Arthur Eubanks [Tue, 18 May 2021 22:09:06 +0000 (15:09 -0700)]
[NFC][OpaquePtr] Explicitly pass GEP source type to IRBuilder in more places

3 years ago[lld/mac] Implement removal of unused dylibs
Nico Weber [Tue, 1 Jun 2021 02:12:35 +0000 (22:12 -0400)]
[lld/mac] Implement removal of unused dylibs

This omits load commands for unreferenced dylibs if:
- the dylib was loaded implicitly,
- it is marked MH_DEAD_STRIPPABLE_DYLIB
- or -dead_strip_dylibs is passed

This matches ld64.

Currently, the "is dylib referenced" state is computed before dead code
stripping and is not updated after dead code stripping. This too matches ld64.
We should do better here.

With this, clang-format linked with lld (like with ld64) no longer has
libobjc.A.dylib in `otool -L` output. (It was implicitly loaded as a reexport
of CoreFoundation.framework, but it's not needed.)

Differential Revision: https://reviews.llvm.org/D103430

3 years agoWindowsSupport.h: do not depend on private config header
Andrew Kelley [Tue, 1 Jun 2021 19:58:05 +0000 (22:58 +0300)]
WindowsSupport.h: do not depend on private config header

WindowsSupport.h is a public header, however if it gets included, will cause a compile error indicating that llvm/Config/config.h cannot be found, because config.h is a private header. However there is no actual dependency on the private things in this header, so it can be changed to the public config header.

Reviewed By: amccarth

Differential Revision: https://reviews.llvm.org/D103370

3 years ago[InstCombine] add tests for cast folding; NFC
Sanjay Patel [Tue, 1 Jun 2021 16:05:36 +0000 (12:05 -0400)]
[InstCombine] add tests for cast folding; NFC

https://llvm.org/PR49543

3 years ago[SystemZ][z/OS] Stricter condition for HLASM class instantiation
Anirudh Prasad [Tue, 1 Jun 2021 19:55:34 +0000 (15:55 -0400)]
[SystemZ][z/OS] Stricter condition for HLASM class instantiation

- A lot of lit tests simply specify the arch minus the triple. On z/OS, this could result in a scenario of some-other-triple-unknown-ibm-zos. This points to an incorrect triple + arch combo.
- To prevent this, isOSzOS change is switched in favour of isOSBinFormatGOFF.
- This is because, the GOFF format is set only if the triple is systemz and if the operating system is GOFF. And currently, there are no other architectures/os's using the GOFF file format.
- An argument could be made that the problematic tests be fixed to explicitly specify the arch-vendor-triple string, but there's a large number of these tests, and adding this stricter scope ensures that we aren't instantiating the incorrect instance of the AsmParser for other platforms when run on z/OS.

Reviewed By: uweigand

Differential Revision: https://reviews.llvm.org/D103343

3 years ago[clang] Fix ICE with typeid & polymorphic class (pr50497)
Nathan Sidwell [Thu, 27 May 2021 14:34:39 +0000 (07:34 -0700)]
[clang] Fix ICE with typeid & polymorphic class (pr50497)

This addresses pr50497. The argument of a typeid expression is
unevaluated, *except* when it's a polymorphic type. We handle this by
parsing as unevaluated and then transforming to evaluated if we
discover it should have been an evaluated context.

We do the same in TreeTransform<Derived>::TransformCXXTypeidExpr,
entering unevaluated context before transforming and rebuilding the
typeid. But that's incorrect and can lead us to converting to
evaluated context twice -- and hitting an assert.

During normal template instantiation we're always cloning the
expression, but during generic lambda processing we do not necessarily
AlwaysRebuild, and end up with TransformDeclRefExpr unconditionally
calling MarkDeclRefReferenced around line 10226. That triggers the
assert.

// Mark it referenced in the new context regardless.
// FIXME: this is a bit instantiation-specific.
SemaRef.MarkDeclRefReferenced(E);

This patch makes 2 changes.

a) TreeTransform<Derived>::TransformCXXTypeidExpr only enters
unevaluated context if the typeid's operand is not a polymorphic
glvalue. If it is, it keeps the same evaluation context.

b) Sema::BuildCXXTypeId is altered to only transform to evaluated, if
the current context is unevaluated.

Differential Revision: https://reviews.llvm.org/D103258

3 years ago[gn build] Port 5671ff20d92b
LLVM GN Syncbot [Tue, 1 Jun 2021 19:37:29 +0000 (19:37 +0000)]
[gn build] Port 5671ff20d92b

3 years ago[libcxx] Implement view.interface.
zoecarver [Fri, 30 Apr 2021 22:24:28 +0000 (15:24 -0700)]
[libcxx] Implement view.interface.

This will unblock work on ranges::view. Based on D101396.

Refs http://eel.is/c++draft/view.interface.

Differential Revision: https://reviews.llvm.org/D101737

3 years ago[lld/mac] Don't load DylibFiles from the DylibFile constructor
Nico Weber [Mon, 31 May 2021 18:59:48 +0000 (14:59 -0400)]
[lld/mac] Don't load DylibFiles from the DylibFile constructor

loadDylib() keeps a name->DylibFile cache, but it only writes
to the cache once the DylibFile constructor has completed.
So dylib loads done recursively from the DylibFile constructor
wouldn't use the cache.

Now, we load additional dylibs after writing to the cache,
which means the cache now gets used for dylibs loaded because
they're referenced from other dylibs.

Related to PR49514 and PR50101, but no dramatic behavior change in itself.
(Technically we no longer crash when a tbd file reexports itself,
but that doesn't happen in practice. We now accept it silently instead
of crashing; ld64 has a diag for the reexport cycle.)

Differential Revision: https://reviews.llvm.org/D103423

3 years ago[lld/mac] Don't write mtimes to N_OSO entries if ZERO_AR_DATE is set.
Nico Weber [Tue, 1 Jun 2021 10:55:36 +0000 (06:55 -0400)]
[lld/mac] Don't write mtimes to N_OSO entries if ZERO_AR_DATE is set.

This is important for build determinism. This matches ld64.

Differential Revision: https://reviews.llvm.org/D103446

3 years ago[AMDGPU][NFC] Remove author's name from codebase
madhur13490 [Tue, 1 Jun 2021 19:18:26 +0000 (00:48 +0530)]
[AMDGPU][NFC] Remove author's name from codebase

This must have made to code by accident.

Differential Revision: https://reviews.llvm.org/D103484

3 years ago[SLPVectorizer] Ignore unreachable blocks
Harald van Dijk [Tue, 1 Jun 2021 19:21:04 +0000 (20:21 +0100)]
[SLPVectorizer] Ignore unreachable blocks

As the existing test unreachable.ll shows, we should be doing more
work to avoid entering unreachable blocks: we should not stop
vectorization just because a PHI incoming value from an unreachable
block cannot be vectorized. We know that particular value will never
be used so we can just replace it with poison.

3 years ago[OpenMP] Add L2-Tile equivalence for KNL
Peyton, Jonathan L [Thu, 20 May 2021 18:50:03 +0000 (13:50 -0500)]
[OpenMP] Add L2-Tile equivalence for KNL

When on KNL and L2 or Tile layer is detected, manually add
the corresponding layer which is equivalent.

Differential Revision: https://reviews.llvm.org/D102865

3 years ago[clang] RecursiveASTVisitor visits ObjCPropertyRefExpr's class receiver
David Goldman [Fri, 30 Apr 2021 16:14:32 +0000 (12:14 -0400)]
[clang] RecursiveASTVisitor visits ObjCPropertyRefExpr's class receiver

We now make up a TypeLoc for the class receiver to simplify visiting,
notably for indexing, availability, and clangd.

Differential Revision: https://reviews.llvm.org/D101645

3 years ago[flang][openacc] Enforce restriction on attach and detach variables
Valentin Clement [Tue, 1 Jun 2021 18:36:58 +0000 (14:36 -0400)]
[flang][openacc] Enforce restriction on attach and detach variables

Each var argument to an attach or detach clause must be a
Fortran variable or array with the pointer or allocatable attribute.
This patch enforce this restruction.

Reviewed By: kiranchandramohan

Differential Revision: https://reviews.llvm.org/D103279

3 years ago[ORC-RT] Add WrapperFunctionResult, Simple Packed Serialization (SPS) system.
Lang Hames [Mon, 31 May 2021 22:50:26 +0000 (15:50 -0700)]
[ORC-RT] Add WrapperFunctionResult, Simple Packed Serialization (SPS) system.

WrapperFunctionResult is a C++ wrapper for __orc_rt_CWrapperFunctionResult
that automatically manages the underlying struct.

The Simple Packed Serialization (SPS) utilities support a simple serialization
scheme for wrapper function argument and result buffers:

Primitive typess (bool, char, int8_t, and uint8_t, int16_t, uint16_t, int32_t,
uint32_t, int64_t, uint64_t) are serialized in little-endian form.

SPSTuples are serialized by serializing each of the tuple members in order
without padding.

SPSSequences are serialized by serializing a sequence length (as a uint64_t)
followed by each of the elements of the sequence in order without padding.

Serialization/deserialization always involves a pair of SPS type tag (a tag
representing the serialized format to use, e.g. uint32_t, or
SPSTuple<bool, SPSString>) and a concrete type to be serialized from or
deserialized to (uint32_t, std::pair<bool, std::string>). Serialization for new
types can be implemented by specializing the SPSSerializationTraits type.

3 years ago[ORC-RT] Rename C-API functions to use __orc_rt_ prefix (instead of OrcRT).
Lang Hames [Mon, 31 May 2021 22:37:00 +0000 (15:37 -0700)]
[ORC-RT] Rename C-API functions to use __orc_rt_ prefix (instead of OrcRT).

This matches the C++ namespace name, and is consistent with other C linkage
functions (e.g. __orc_rt_jit_dispatch).

3 years ago[ORC-RT] Add equality/inequality comparison to string_view.
Lang Hames [Mon, 31 May 2021 22:25:13 +0000 (15:25 -0700)]
[ORC-RT] Add equality/inequality comparison to string_view.

3 years ago[OpenMP] Define named constants for interop's foreign runtime ID
Hansang Bae [Tue, 1 Jun 2021 14:31:27 +0000 (09:31 -0500)]
[OpenMP] Define named constants for interop's foreign runtime ID

Also added missing Fortran definitions for interop support.

Differential Revision: https://reviews.llvm.org/D102883

3 years agoFix a git apply that went bad somehow.
Aaron Ballman [Tue, 1 Jun 2021 18:05:26 +0000 (14:05 -0400)]
Fix a git apply that went bad somehow.

When applying the changes in 8edd3464afbff65d7d5945b3a8b20009d6ff5deb,
it seems that this bit got merged incorrectly and no test coverage
caught the issue. This fixes the diagnostic and adds a test.

3 years ago[GlobalISel][AArch64] Combine and (lshr x, cst), mask -> ubfx x, cst, width
Jessica Paquette [Wed, 24 Mar 2021 17:52:21 +0000 (10:52 -0700)]
[GlobalISel][AArch64] Combine and (lshr x, cst), mask -> ubfx x, cst, width

Also add a target hook which allows us to get around custom legalization on
AArch64.

Differential Revision: https://reviews.llvm.org/D99283

3 years ago[libc++] Temporary hack to unbreak CI
Louis Dionne [Tue, 1 Jun 2021 17:49:53 +0000 (13:49 -0400)]
[libc++] Temporary hack to unbreak CI

Until I figure out what the issue is with this test on AppleClang (and
in particular which change caused it), I want to avoid all CI being broken.

3 years ago[clangd][ObjC] Fix issue completing a method decl by name
David Goldman [Mon, 19 Apr 2021 20:38:37 +0000 (16:38 -0400)]
[clangd][ObjC] Fix issue completing a method decl by name

When completing an Objective-C method declaration by name, we need to
preserve the leading text as a `qualifier` so we insert it properly
before the first typed text chunk.

Differential Revision: https://reviews.llvm.org/D100798

3 years ago[X86FixupLEAs] Transform the sequence LEA/SUB to SUB/SUB
Guozhi Wei [Tue, 1 Jun 2021 17:31:30 +0000 (10:31 -0700)]
[X86FixupLEAs] Transform the sequence LEA/SUB to SUB/SUB

This patch transforms the sequence

    lea (reg1, reg2), reg3
    sub reg3, reg4

to two sub instructions

    sub reg1, reg4
    sub reg2, reg4

Similar optimization can also be applied to LEA/ADD sequence.
The modifications to TwoAddressInstructionPass is to ensure the operands of ADD
instruction has expected order (the dest register of LEA should be src register of ADD).

Differential Revision: https://reviews.llvm.org/D101970

3 years ago[MLIR] Add missing APSInt.h include
Nikita Popov [Tue, 1 Jun 2021 17:18:19 +0000 (19:18 +0200)]
[MLIR] Add missing APSInt.h include

Since fd7e309e02fd226b0390888388ed732608e52c73 this is no longer
pulled in indirectly through DenseMapInfo.h.

3 years ago[SystemZ] Return true from hasBitPreservingFPLogic().
Jonas Paulsson [Tue, 1 Jun 2021 16:24:06 +0000 (18:24 +0200)]
[SystemZ] Return true from hasBitPreservingFPLogic().

This is currently NFC on benchmarks and tests.

Review: Ulrich Weigand

3 years ago[polly] Fix SCEVLoopAddRecRewriter to avoid invalid AddRecs.
Eli Friedman [Fri, 21 May 2021 23:58:36 +0000 (16:58 -0700)]
[polly] Fix SCEVLoopAddRecRewriter to avoid invalid AddRecs.

When we're remapping an AddRec, the AddRec constructed by a partial
rewrite might not make sense.  This triggers an assertion complaining
it's not loop-invariant.

Instead of constructing the partially rewritten AddRec, just skip
straight to calling evaluateAtIteration.

Testcase was automatically reduced using llvm-reduce, so it's a little
messy, but hopefully makes sense.

Differential Revision: https://reviews.llvm.org/D102959