platform/upstream/llvm.git
4 years agoGlobalISel: Start adding computeNumSignBits to GISelKnownBits
Matt Arsenault [Sat, 4 Jan 2020 19:13:06 +0000 (14:13 -0500)]
GlobalISel: Start adding computeNumSignBits to GISelKnownBits

4 years agoAMDGPU: Fix legalizing f16 fpow
Matt Arsenault [Sat, 4 Jan 2020 20:48:46 +0000 (15:48 -0500)]
AMDGPU: Fix legalizing f16 fpow

The existing test only covered one case for r600. The use of
mul_legacy also looks suspicious to me, but leave it for now. The
patterns are also not making use of source modifiers.

4 years agoAMDGPU: Use ImmLeaf
Matt Arsenault [Mon, 6 Jan 2020 20:39:05 +0000 (15:39 -0500)]
AMDGPU: Use ImmLeaf

This solves one GlobalISel importer error, but the pattern still fails
for another reason.

4 years agoAMDGPU: Use ImmLeaf for inline immediate predicates
Matt Arsenault [Mon, 6 Jan 2020 19:39:13 +0000 (14:39 -0500)]
AMDGPU: Use ImmLeaf for inline immediate predicates

4 years agollc/MIR: Fix setFunctionAttributes for MIR functions
Matt Arsenault [Fri, 6 Dec 2019 18:36:34 +0000 (00:06 +0530)]
llc/MIR: Fix setFunctionAttributes for MIR functions

A random set of attributes are implemented by llc/opt forcing the
string attributes on the IR functions before processing anything. This
would not happen for MIR functions, which have not yet been created at
this point.

Use a callback in the MIR parser, purely to avoid dealing with the
ugliness that the command line flags are in a .inc file, and would
require allowing access to these flags from multiple places (either
from the MIR parser directly, or a new utility pass to implement these
flags). It would probably be better to cleanup the flag handling into
a separate library.

This is in preparation for treating more command line flags with a
corresponding function attribute in a more uniform way. The fast math
flags in particular have a messy system where the command line flag
sets the behavior from a function attribute if present, and otherwise
the command line flag. This means if any other pass tries to inspect
the function attributes directly, it will be inconsistent with the
intended behavior. This is also inconsistent with the current behavior
of -mcpu and -mattr, which overwrites any pre-existing function
attributes. I would like to move this to consistenly have the command
line flags not overwrite any pre-existing attributes, and to always
ensure the command line flags are consistent with the function
attributes.

4 years ago[X86] Improve v4i32->v4f64 uint_to_fp for AVX1/AVX2 targets.
Craig Topper [Mon, 6 Jan 2020 22:06:20 +0000 (14:06 -0800)]
[X86] Improve v4i32->v4f64 uint_to_fp for AVX1/AVX2 targets.

Use zext+or+fsub to do the conversion. Similar to D71971.

Differential Revision: https://reviews.llvm.org/D71971

4 years ago[lldb/Docs] Describe optional dependencies on build page.
Jonas Devlieghere [Mon, 6 Jan 2020 21:57:55 +0000 (13:57 -0800)]
[lldb/Docs] Describe optional dependencies on build page.

List the different CMake flags controlling the optional dependencies as
per the discussion on the mailing list:

http://lists.llvm.org/pipermail/lldb-dev/2020-January/015867.html

4 years ago[LegalizeTypes] Add widening support for STRICT_FSETCC/FSETCCS
Craig Topper [Mon, 6 Jan 2020 21:37:47 +0000 (13:37 -0800)]
[LegalizeTypes] Add widening support for STRICT_FSETCC/FSETCCS

This patch adds widening which really just scalarizes because we don't have a strategy for the extra elements we would need to pad with.

Differential Revision: https://reviews.llvm.org/D72193

4 years ago[X86] Rename vec-strict-*-cmp.ll to vec-strict-cmp-*.ll to match other strict files...
Craig Topper [Mon, 6 Jan 2020 21:34:45 +0000 (13:34 -0800)]
[X86] Rename vec-strict-*-cmp.ll to vec-strict-cmp-*.ll to match other strict files wich have the size at the end. NFC

4 years ago[OPENMP50]Support lastprivate conditional updates in inc/dec unary ops.
Alexey Bataev [Mon, 6 Jan 2020 21:14:34 +0000 (16:14 -0500)]
[OPENMP50]Support lastprivate conditional updates in inc/dec unary ops.

Added support for checking of updates of variables used in unary
pre(pos) inc/dec expressions.

4 years ago [NFC] Test commit, revert whitespace change
stevewan [Mon, 6 Jan 2020 21:28:13 +0000 (16:28 -0500)]
 [NFC] Test commit, revert whitespace change

As per the Developer Policy, upon obtaining commit access.

4 years ago[NFC] Test commit, whitespace change
stevewan [Mon, 6 Jan 2020 21:24:27 +0000 (16:24 -0500)]
[NFC] Test commit, whitespace change

As per the Developer Policy, upon obtaining commit access.

4 years ago[x86] add tests for concat self + shuffle; NFC
Sanjay Patel [Mon, 6 Jan 2020 20:24:30 +0000 (15:24 -0500)]
[x86] add tests for concat self + shuffle; NFC

4 years ago[OpenMP] Fix incorrect property of __has_attribute() macro
Kelvin Li [Mon, 6 Jan 2020 20:00:10 +0000 (15:00 -0500)]
[OpenMP] Fix incorrect property of __has_attribute() macro

__has_attribute(fallthough) -> __has_attribute(fallthrough)

Submitted by: kiszk (Kazuaki Ishizaki <ishizaki@jp.ibm.com>)

Differential Revision: https://reviews.llvm.org/D72287

4 years agoLower TAGPstack with negative offset to SUBG.
Evgenii Stepanov [Tue, 10 Dec 2019 00:52:31 +0000 (16:52 -0800)]
Lower TAGPstack with negative offset to SUBG.

Summary:
This never really occurs in the current codegen, so only a MIR test is
possible.

Reviewers: ostannard, pcc

Subscribers: hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D72123

4 years ago[X86] Fix an 8 bit testb being selected when folding a volatile i32 load pattern.
Amara Emerson [Wed, 11 Dec 2019 22:42:16 +0000 (14:42 -0800)]
[X86] Fix an 8 bit testb being selected when folding a volatile i32 load pattern.

Differential Revision: https://reviews.llvm.org/D71581

4 years ago[PowerPC][LoopVectorize] Extend getRegisterClassForType to consider double and other...
Jinsong Ji [Fri, 3 Jan 2020 20:25:19 +0000 (20:25 +0000)]
[PowerPC][LoopVectorize] Extend getRegisterClassForType to consider double and other floating point type

In https://reviews.llvm.org/D67148, we use isFloatTy to test floating
point type, otherwise we return GPRRC.
So 'double' will be classified as GPRRC, which is not accurate.

This patch covers other floating point types.

Reviewed By: #powerpc, nemanjai

Differential Revision: https://reviews.llvm.org/D71946

4 years ago[lld] Fix trivial typos in comments
Kazuaki Ishizaki [Mon, 6 Jan 2020 18:21:05 +0000 (10:21 -0800)]
[lld] Fix trivial typos in comments

Reviewed By: ruiu, MaskRay

Differential Revision: https://reviews.llvm.org/D72196

4 years ago[libc] Move all tests to a top level `test` directory.
Siva Chandra Reddy [Fri, 3 Jan 2020 20:00:45 +0000 (12:00 -0800)]
[libc] Move all tests to a top level `test` directory.

A toplevel target, `check-libc` has also been added.

Reviewers: abrachet, phosek

Tags: #libc-project

Differential Revision: https://reviews.llvm.org/D72177

4 years ago[lldb/Docs] Fix capitalization typo.
Jonas Devlieghere [Mon, 6 Jan 2020 18:08:05 +0000 (10:08 -0800)]
[lldb/Docs] Fix capitalization typo.

This has been bothering me for way too long.

4 years ago[lldb/CMake] Only set PYTHON_HOME on Windows
Jonas Devlieghere [Mon, 6 Jan 2020 17:56:46 +0000 (09:56 -0800)]
[lldb/CMake] Only set PYTHON_HOME on Windows

My earlier change for Python auto-detection caused PYTHON_HOME to be set
unconditionally, while before the change this only happened for Windows.
This caused the PythonDataObjectsTest to fail with an import error.

4 years ago[CMake] Add $ORIGIN/../../../../lib to rpath if BUILD_SHARED_LIBS or LLVM_LINK_LLVM_D...
Fangrui Song [Sat, 21 Dec 2019 22:59:16 +0000 (14:59 -0800)]
[CMake] Add $ORIGIN/../../../../lib to rpath if BUILD_SHARED_LIBS or LLVM_LINK_LLVM_DYLIB on *nix

Summary:
lib/python2.7/dist-packages/lldb/_lldb.so is a symlink to lib/liblldb.so,
which depends on lib/libLLVM*.so (-DBUILD_SHARED_LIBS=ON) or lib/libLLVM-10git.so
(-DLLVM_LINK_LLVM_DYLIB=ON). Add an additional rpath `$ORIGIN/../../../../lib` so
that _lldb.so can be loaded from Python.

This fixes an import error from lib/python2.7/dist-packages/lldb/__init__.py

  from . import _lldb
  ImportError: libLLVMAArch64CodeGen.so.10git: cannot open shared object file: No such file or directory

 The following configurations will work:

* -DBUILD_SHARED_LIBS=ON
* -DBUILD_SHARED_LIBS=OFF -DLLVM_LINK_LLVM_DYLIB=ON
* -DBUILD_SHARED_LIBS=OFF -DLLVM_LINK_LLVM_DYLIB=ON -DCLANG_LINK_CLANG_DYLIB=ON
  (-DCLANG_LINK_CLANG_DYLIB=ON depends on -DLLVM_LINK_LLVM_DYLIB=ON)

Reviewed By: labath

Differential Revision: https://reviews.llvm.org/D71800

4 years agoMake check-llvm run 50% faster on macOS, 18% faster on Windows.
Nico Weber [Mon, 6 Jan 2020 15:54:13 +0000 (10:54 -0500)]
Make check-llvm run 50% faster on macOS, 18% faster on Windows.

While looking at cycle time graphs of some of my bots, I noticed
that 327894859cc made check-llvm noticeably slower on macOS and
Windows.

As it turns out, the 5 substitutions added in that change were
enough to cause lit to thrash the build-in cache in re.compile()
(re.sub() is implemented as re.compile().sub()), and apparently
applySubstitutions() is on the cricital path and slow when all
regexes need to compile all the time.

(See `_MAXCACHE = 512` in cpython/Lib/re.py)

Supporting full regexes for lit substitutions seems a bit like
overkill, but for now add a simple unbounded cache to recover
the lost performance.

No intended behavior change.

4 years ago[lldb/Test] Move @skipIfAsan from test class to test methods.
Jonas Devlieghere [Mon, 6 Jan 2020 17:52:09 +0000 (09:52 -0800)]
[lldb/Test] Move @skipIfAsan from test class to test methods.

skipTestIfFn can only be used to decorate a test method.

4 years ago[llvm-readelf] Print EI_ABIVERSION as decimal instead of hexadecimal
Fangrui Song [Sun, 5 Jan 2020 22:56:27 +0000 (14:56 -0800)]
[llvm-readelf] Print EI_ABIVERSION as decimal instead of hexadecimal

This matches GNU readelf and llvm-readobj.

Reviewed By: jhenderson

Differential Revision: https://reviews.llvm.org/D72234

4 years ago[gn build] Port 350da402ef6
LLVM GN Syncbot [Mon, 6 Jan 2020 17:22:13 +0000 (17:22 +0000)]
[gn build] Port 350da402ef6

4 years ago[clang-tidy] new check: bugprone-signed-char-misuse
Tamás Zolnai [Sat, 4 Jan 2020 13:05:09 +0000 (14:05 +0100)]
[clang-tidy] new check: bugprone-signed-char-misuse

Summary:
This check searches for signed char -> integer conversions which might
indicate programming error, because of the misinterpretation of char
values. A signed char might store the non-ASCII characters as negative
values. The human programmer probably expects that after an integer
conversion the converted value matches with the character code
(a value from [0..255]), however, the actual value is in
[-128..127] interval.

See also:
STR34-C. Cast characters to unsigned char before converting to larger integer sizes
<https://wiki.sei.cmu.edu/confluence/display/c/STR34-C.+Cast+characters+to+unsigned+char+before+converting+to+larger+integer+sizes>

By now this check is limited to assignment / variable declarations.
If we would catch all signed char -> integer conversion, then it would
produce a lot of findings and also false positives. So I added only
this use case now, but this check can be extended with additional
use cases later.
The CERT documentation mentions another use case when the char is
used for array subscript. Next to that a third use case can be
the signed char - unsigned char comparison, which also a use case
where things happen unexpectedly because of conversion to integer.

Reviewers: alexfh, hokein, aaron.ballman

Reviewed By: aaron.ballman

Subscribers: sylvestre.ledru, whisperity, Eugene.Zelenko, mgorny, xazax.hun, cfe-commits

Tags: #clang, #clang-tools-extra

Differential Revision: https://reviews.llvm.org/D71174

4 years ago[lldb/Test] Temporarily skip TestFoundationDisassembly on the ASan bot.
Jonas Devlieghere [Mon, 6 Jan 2020 17:06:09 +0000 (09:06 -0800)]
[lldb/Test] Temporarily skip TestFoundationDisassembly on the ASan bot.

This test is timing out on the sanitized bot on GreenDragon. Temporarily
disable it to increase the signal-to-noise ration.

4 years ago[lldb/CMake] Autodetect Python dependency
Jonas Devlieghere [Mon, 6 Jan 2020 16:49:36 +0000 (08:49 -0800)]
[lldb/CMake] Autodetect Python dependency

Python was the last remaining "optional" dependency for LLDB. This moves
the code to find Python into FindPythonInterpAndLibs using the same
principles as FindCursesAndPanel.

Differential revision: https://reviews.llvm.org/D72107

4 years ago[AIX] Use csect reference for function address constants
diggerlin [Mon, 6 Jan 2020 16:45:00 +0000 (11:45 -0500)]
[AIX] Use csect reference for function address constants

SUMMARY:
We currently emit a reference for function address constants as labels;
for example:

foo_ptr:
.long foo
however, there may be no such label in the case where the function is
undefined. Although the label exists when the function is defined, we
will (to be consistent) also use a csect reference in that case.

Address one comment
https://reviews.llvm.org/D71144#inline-653255

Reviewers: daltenty,hubert.reinterpretcast,jasonliu,Xiangling_L
Subscribers: cebowleratibm, wuzish, nemanjai

Differential Revision: https://reviews.llvm.org/D71144

4 years ago[llvm-libc] Fix missing virtual destructor
Guillaume Chatelet [Mon, 6 Jan 2020 12:31:45 +0000 (13:31 +0100)]
[llvm-libc] Fix missing virtual destructor

Summary: This patch adds a virtual destructor to the Command class.

Reviewers: sivachandra

Subscribers: mgorny, MaskRay, libc-commits

Tags: #libc-project

Differential Revision: https://reviews.llvm.org/D72253

4 years ago[ARM] Use the correct opcodes for Thumb2 segmented stack frame lowering
David Green [Mon, 6 Jan 2020 15:54:36 +0000 (15:54 +0000)]
[ARM] Use the correct opcodes for Thumb2 segmented stack frame lowering

The segmented stack lowering code appears to be using ARM opcodes under
Thumb2. The MRC opcode will be the same for Thumb and ARM, but t2LDR
seems wrong. Either way, using the correct thumb vs arm opcodes is more
correct.

Differential Revision: https://reviews.llvm.org/D72074

4 years ago[ARM] Use correct TRAP opcode for thumb in FastISel
David Green [Mon, 6 Jan 2020 10:58:14 +0000 (10:58 +0000)]
[ARM] Use correct TRAP opcode for thumb in FastISel

We were previously unconditionally using the ARM::TRAP opcode, even
under Thumb. My understanding is that these are essentially the same
thing (they both result in a trap under Thumb), but the ARM::TRAP opcode
is marked as requiring IsARM, so it is more correct to use ARM::tTRAP.

Differential Revision: https://reviews.llvm.org/D72075

4 years ago[AIX] Use csect reference for function address constants
diggerlin [Mon, 6 Jan 2020 16:38:22 +0000 (11:38 -0500)]
[AIX] Use csect reference for function address constants

SUMMARY:
We currently emit a reference for function address constants as labels;
for example:

foo_ptr:
.long foo
however, there may be no such label in the case where the function is
undefined. Although the label exists when the function is defined, we
will (to be consistent) also use a csect reference in that case.

Reviewers: daltenty,hubert.reinterpretcast,jasonliu,Xiangling_L
Subscribers: cebowleratibm, wuzish, nemanjai

Differential Revision: https://reviews.llvm.org/D71144

4 years agoAdds -Wrange-loop-analysis to -Wall
Mark de Wever [Wed, 1 Jan 2020 16:23:18 +0000 (17:23 +0100)]
Adds -Wrange-loop-analysis to -Wall

This makes the range loop warnings part of -Wall.

Fixes PR32823: Warn about accidental coping of data in range based for

Differential Revision: https://reviews.llvm.org/D68912

Recomitted after fixing the warnings it created.

4 years ago[NFC] Fixes -Wrange-loop-analysis warnings
Mark de Wever [Sun, 5 Jan 2020 13:26:39 +0000 (14:26 +0100)]
[NFC] Fixes -Wrange-loop-analysis warnings

This avoids new warnings due to D68912 adds -Wrange-loop-analysis to -Wall.

Differential Revision: https://reviews.llvm.org/D72210

4 years ago[AMDGPU] Fix "use of uninitialized variable" static analyzer warning. NFCI.
Simon Pilgrim [Mon, 6 Jan 2020 16:36:33 +0000 (16:36 +0000)]
[AMDGPU] Fix "use of uninitialized variable" static analyzer warning. NFCI.

Add "unreachable" default case to AMDGPUTargetStreamer::getArchNameFromElfMach

4 years agoFix "use of uninitialized variable" static analyzer warnings. NFCI.
Simon Pilgrim [Mon, 6 Jan 2020 16:17:05 +0000 (16:17 +0000)]
Fix "use of uninitialized variable" static analyzer warnings. NFCI.

Add "unreachable" default cases like we do for the other switch()s in X86MCInstLower::Lower

4 years agoFix "use of uninitialized variable" static analyzer warning. NFCI.
Simon Pilgrim [Mon, 6 Jan 2020 16:15:17 +0000 (16:15 +0000)]
Fix "use of uninitialized variable" static analyzer warning. NFCI.

4 years ago[ARM,MVE] Fix many signedness errors in MVE intrinsics.
Simon Tatham [Mon, 6 Jan 2020 16:33:14 +0000 (16:33 +0000)]
[ARM,MVE] Fix many signedness errors in MVE intrinsics.

Summary:
Running an end-to-end test last week I noticed that a lot of the ACLE
intrinsics that operate differently on vectors of signed and unsigned
integers were ending up generating the signed version of the
instruction unconditionally. This is because the IR intrinsics had no
way to distinguish signed from unsigned: the LLVM type system just
calls them both `v8i16` (or whatever), so you need either separate
intrinsics for signed and unsigned, or a flag parameter that tells
ISel which one to choose.

This patch fixes all the problems of that kind that I've noticed, by
adding an i32 flag parameter to many of the IR intrinsics which is set
to 1 for unsigned (matching the existing practice in cases where we
got it right), and conditioning all the isel patterns on that flag. So
the fundamental change is in `IntrinsicsARM.td`, changing the
low-level IR intrinsics API; there are knock-on changes in
`arm_mve.td` (adjusting code gen for the ACLE intrinsics to use the
modified API) and in `ARMInstrMVE.td` (adjusting isel to expect the
new unsigned flags). The rest of this patch is boringly updating tests.

Reviewers: dmgreen, miyuki, MarkMurrayARM

Reviewed By: dmgreen

Subscribers: kristof.beyls, hiraditya, cfe-commits, llvm-commits

Tags: #clang, #llvm

Differential Revision: https://reviews.llvm.org/D72270

4 years ago[ARM,MVE] Support -ve offsets in gather-load intrinsics.
Simon Tatham [Mon, 6 Jan 2020 16:33:05 +0000 (16:33 +0000)]
[ARM,MVE] Support -ve offsets in gather-load intrinsics.

Summary:
The ACLE intrinsics with `gather_base` or `scatter_base` in the name
are wrappers on the MVE load/store instructions that take a vector of
base addresses and an immediate offset. The immediate offset can be up
to 127 times the alignment unit, and it can be positive or negative.

At the MC layer, we got that right. But in the Sema error checking for
the wrapping intrinsics, the offset was erroneously constrained to be
positive.

To fix this I've adjusted the `imm_mem7bit` class in the Tablegen that
defines the intrinsics. But that causes integer literals like
`0xfffffffffffffe04` to appear in the autogenerated calls to
`SemaBuiltinConstantArgRange`, which provokes a compiler warning
because that's out of the non-overflowing range of an `int64_t`. So
I've also tweaked `MveEmitter` to emit that as `-0x1fc` instead.

Updated the tests of the Sema checks themselves, and also adjusted a
random sample of the CodeGen tests to actually use negative offsets
and prove they get all the way through code generation without causing
a crash.

Reviewers: dmgreen, miyuki, MarkMurrayARM

Reviewed By: dmgreen

Subscribers: kristof.beyls, cfe-commits, llvm-commits

Tags: #clang, #llvm

Differential Revision: https://reviews.llvm.org/D72268

4 years ago[ARM,MVE] Generate the right instruction for vmaxnmq_m_f16.
Simon Tatham [Mon, 6 Jan 2020 16:28:18 +0000 (16:28 +0000)]
[ARM,MVE] Generate the right instruction for vmaxnmq_m_f16.

Summary:
Due to a copy-paste error in the isel patterns, the predicated version
of this intrinsic was expanding to the `VMAXNMT.F32` instruction
instead of `VMAXNMT.F16`. Similarly for vminnm.

Reviewers: dmgreen, miyuki, MarkMurrayARM

Reviewed By: dmgreen

Subscribers: kristof.beyls, hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D72269

4 years agoAMDGPU/GlobalISel: Select scalar v2s16 G_BUILD_VECTOR
Matt Arsenault [Mon, 2 Sep 2019 07:27:20 +0000 (03:27 -0400)]
AMDGPU/GlobalISel: Select scalar v2s16 G_BUILD_VECTOR

4 years ago[lldb] [Process/NetBSD] Remove unused orig_*ax use
Michał Górny [Sat, 4 Jan 2020 04:34:41 +0000 (05:34 +0100)]
[lldb] [Process/NetBSD] Remove unused orig_*ax use

orig_*ax logic is Linux-specific, and was never used on NetBSD.
In fact, its support seems to be a dead code entirely.

Differential Revision: https://reviews.llvm.org/D72195

4 years agoAMDGPU/GlobalISel: Select more G_EXTRACTs correctly
Matt Arsenault [Wed, 1 Jan 2020 20:23:58 +0000 (15:23 -0500)]
AMDGPU/GlobalISel: Select more G_EXTRACTs correctly

This assumed a 32-bit extract size, which would produce invalid copies
with 64-bit extracts. Handle the easy case. Ideally we would have a
way to get the proper subreg index for any 32-bit offset, but there
should probably be a tablegenerated way of getting the subreg index
for any size and offset.

4 years ago[mlir][Linalg] Reimplement and extend getStridesAndOffset
Nicolas Vasilache [Thu, 2 Jan 2020 20:25:21 +0000 (15:25 -0500)]
[mlir][Linalg] Reimplement and extend getStridesAndOffset

Summary: This diff reimplements getStridesAndOffset in a significantly simpler way by operating on the AffineExpr and calling into simplifyAffineExpr instead of rolling its own saturating arithmetic.

As a consequence it becomes quite simple to extend the behavior of getStridesAndOffset to encompass more cases by manipulating the AffineExpr more directly.
The divisions are still filtered out and continue to yield fully dynamic strides.
Simplifying the divisions is left for a later time if compelling use cases arise.

Relevant tests are added.

Reviewers: ftynse

Subscribers: mehdi_amini, rriddle, jpienaar, burmako, shauheen, antiagainst, arpith-jacob, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D72098

4 years ago[clang-format] fix conflict between FormatStyle::BWACS_MultiLine and BeforeCatch...
Mitchell Balan [Mon, 6 Jan 2020 14:21:41 +0000 (09:21 -0500)]
[clang-format] fix conflict between FormatStyle::BWACS_MultiLine and BeforeCatch/BeforeElse

Summary:
Found a bug introduced with BraceWrappingFlags AfterControlStatement MultiLine. This feature conflicts with the existing BeforeCatch and BeforeElse flags.

For example, our team uses BeforeElse.

if (foo ||
    bar) {
  doSomething();
}
else {
  doSomethingElse();
}

If we enable MultiLine (which we'd really love to do) we expect it to work like this:

if (foo ||
    bar)
{
  doSomething();
}
else {
  doSomethingElse();
}

What we actually get is:

if (foo ||
    bar)
{
  doSomething();
}
else
{
  doSomethingElse();
}

Reviewers: MyDeveloperDay, Bouska, mitchell-stellar

Patch by: pastey

Subscribers: Bouska, cfe-commits

Tags: clang

Differential Revision: https://reviews.llvm.org/D71939

4 years ago[X86] Add extra PR43971 test case mentioned in D70267
Simon Pilgrim [Mon, 6 Jan 2020 13:44:55 +0000 (13:44 +0000)]
[X86] Add extra PR43971 test case mentioned in D70267

4 years ago[CostModel][X86] Add missing scalar i64->f32 uitofp costs
Simon Pilgrim [Mon, 6 Jan 2020 13:16:43 +0000 (13:16 +0000)]
[CostModel][X86] Add missing scalar i64->f32 uitofp costs

4 years ago[DAG] DAGCombiner::XformToShuffleWithZero - use APInt::extractBits helper. NFCI.
Simon Pilgrim [Mon, 6 Jan 2020 10:42:48 +0000 (10:42 +0000)]
[DAG] DAGCombiner::XformToShuffleWithZero - use APInt::extractBits helper. NFCI.

4 years ago[test][DebugInfo][NFC] Rename method for clarity
James Henderson [Fri, 3 Jan 2020 14:06:57 +0000 (14:06 +0000)]
[test][DebugInfo][NFC] Rename method for clarity

The checkGetOrParseLineTableEmitsError function could end up generating
both recoverable and unrecoverable errors, but it is only intended for
handling the latter.

Reviewed by: dblaikie

Differential Revision: https://reviews.llvm.org/D72156

4 years ago[NFC] Fix trivial typos in comments
James Henderson [Mon, 6 Jan 2020 10:15:44 +0000 (10:15 +0000)]
[NFC] Fix trivial typos in comments

Reviewed By: jhenderson

Differential Revision: https://reviews.llvm.org/D72143

Patch by Kazuaki Ishizaki.

4 years agoFix MSVC "not all control paths return a value" warning. NFCI.
Simon Pilgrim [Mon, 6 Jan 2020 10:20:20 +0000 (10:20 +0000)]
Fix MSVC "not all control paths return a value" warning. NFCI.

4 years ago[ARM][MVE] More MVETailPredication debug messages. NFC.
Sjoerd Meijer [Mon, 6 Jan 2020 09:56:02 +0000 (09:56 +0000)]
[ARM][MVE] More MVETailPredication debug messages. NFC.

I've added a few more debug messages to MVETailPredication because I wanted to
trace better which instructions are added/removed. And while I was at it, I
factored out one function which I thought was clearer, and have added some
comments to describe better the flow between MVETailPredication and
ARMLowOverheadLoops.

Differential Revision: https://reviews.llvm.org/D71549

4 years agoAdd interface emitPrefix for MCCodeEmitter
Shengchen Kan [Tue, 31 Dec 2019 16:19:56 +0000 (00:19 +0800)]
Add interface emitPrefix for MCCodeEmitter

Differential Revision: https://reviews.llvm.org/D72047

4 years ago[APFloat] Fix compilation warnings
Ehud Katz [Mon, 6 Jan 2020 08:51:55 +0000 (10:51 +0200)]
[APFloat] Fix compilation warnings

4 years ago[mlir] Update mlir/CMakeLists.txt to install *.def files
Kern Handa [Mon, 6 Jan 2020 09:01:59 +0000 (10:01 +0100)]
[mlir] Update mlir/CMakeLists.txt to install *.def files

This is needed to consume mlir after it has been installed of the source
tree. Without this, consuming mlir results a build error.

Differential Revision: https://reviews.llvm.org/D72232

4 years agoAdd ExternalAAWrapperPass to createLegacyPMAAResults.
Neil Henning [Wed, 11 Dec 2019 13:20:42 +0000 (13:20 +0000)]
Add ExternalAAWrapperPass to createLegacyPMAAResults.

Our out-of-tree custom aliasing solution for the HPC# Burst compiler
here at Unity makes use of the `ExternalAAwrapperPass` infrastructure to
insert our custom aliasing resolution into the core of LLVM. This is
great for all cases except for function inlining, where because
`createLegacyPMAAResults` does not make use of `ExternalAAWrapperPass`,
when we have a definite no-alias result within a function it won't be
propagated to the calling function during inlining.

This commit just rectifies this oversight by adding the missing
dependency.

Differential Revision: https://reviews.llvm.org/D71348

4 years ago[APFloat] Add recoverable string parsing errors to APFloat
Ehud Katz [Mon, 6 Jan 2020 08:05:00 +0000 (10:05 +0200)]
[APFloat] Add recoverable string parsing errors to APFloat

Implementing the APFloat part in PR4745.

Differential Revision: https://reviews.llvm.org/D69770

4 years ago[Metadata] Add TBAA struct metadata to `AAMDNode`
Anton Afanasyev [Fri, 29 Nov 2019 11:22:10 +0000 (14:22 +0300)]
[Metadata] Add TBAA struct metadata to `AAMDNode`

Summary:
Make `AAMDNodes`' `getAAMetadata()` and `setAAMetadata()` to take `!tbaa.struct`
into account as well as `!tbaa`. This impacts llvm.org/pr42022.
This is a temprorary fix needed to keep `!tbaa.struct` tag by SROA pass.
New field `TBAAStruct` should be deleted when `!tbaa` tag replaces `!tbaa.struct`.
Merging two `!tbaa.struct`'s to one is conservatively considered to be `nullptr`
(giving `MayAlias`) -- this could be enhanced, but relying on the said future
replacement.

Reviewers: RKSimon, spatel, vporpo

Subscribers: hiraditya, kosarev, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D70924

4 years ago[Clang] Force rtlib=platform in test to avoid fails with CLANG_DEFAULT_RTLIB
Kristina Brooks [Mon, 6 Jan 2020 07:21:13 +0000 (07:21 +0000)]
[Clang] Force rtlib=platform in test to avoid fails with CLANG_DEFAULT_RTLIB

Driver test `cross-linux.c` fails when CLANG_DEFAULT_RTLIB is "compiler-rt"
as the it expects a GCC-style `"crtbegin.o"` after `"crti.o"` but instead
receives something akin to this in the frontend invocation:

```
"crt1.o" "crti.o"
"/o/b/llvm/bin/../lib/clang/10.0.0/lib/linux/clang_rt.crtbegin-x86_64.o"
```

This patch adds an override to `cross-linux.c` tests so the expected result
is produced regardless of the compile-time default rtlib, as having tests
fail due to that is fairly confusing. After applying the patch, the test
passes regardless of the CLANG_DEFAULT_RTLIB setting.

Differential Revision: https://reviews.llvm.org/D72236

4 years ago[TargetLowering] Use SETCC input type to call getBooleanContents instead of the setcc...
Craig Topper [Mon, 6 Jan 2020 06:54:34 +0000 (22:54 -0800)]
[TargetLowering] Use SETCC input type to call getBooleanContents instead of the setcc result type.

This isn't a functonal change since we also check the bit width is the
same and the input type is integer. This guarantees the input and
output type are the same. But passing the input type makes the code
more readable.

4 years ago[mlir][spirv] Update SPIR-V documentation with information about
MaheshRavishankar [Thu, 2 Jan 2020 23:17:48 +0000 (15:17 -0800)]
[mlir][spirv] Update SPIR-V documentation with information about
lowering to SPIR-V dialect.

Add information about
- SPIRVTypeConverter
- SPIRVOpLowering
- Utility functions used in lowering to SPIR-V dialect.

4 years ago[MC] Reorder members of MCFragment's subclasses to decrease padding
Fangrui Song [Mon, 6 Jan 2020 04:00:07 +0000 (20:00 -0800)]
[MC] Reorder members of MCFragment's subclasses to decrease padding

On a 64-bit platform:

sizeof(MCBoundaryAlignFragment): 64 -> 56
sizeof(MCOrgFragment): 72 -> 64
sizeof(MCFillFragment): 80 -> 72
sizeof(MCLEBFragment): 88 -> 80

4 years ago[MC] Reorder MCFragment members to decrease padding
Fangrui Song [Mon, 6 Jan 2020 03:02:59 +0000 (19:02 -0800)]
[MC] Reorder MCFragment members to decrease padding

sizeof(MCFragment) does not change, but some if its subclasses do, e.g.
on a 64-bit platform,
sizeof(MCEncodedFragment) decreases from 64 to 56,
sizeof(MCDataFragment) decreases from 224 to 216.

4 years ago[DAGCombine] Don't check the legality of type when combine the SIGN_EXTEND_INREG
QingShan Zhang [Mon, 6 Jan 2020 03:00:58 +0000 (03:00 +0000)]
[DAGCombine] Don't check the legality of type when combine the SIGN_EXTEND_INREG

This is the DAG node for SIGN_EXTEND_INREG :

t21: v4i32 = sign_extend_inreg t18, ValueType:ch:v4i16

It has two operands. The first one is the value it want to extend, and the second
one is the type to specify how to extend the value. For this example, it means
that, it is signed extend the t18(v4i32) from v4i16 to v4i32. That is
the semantics of c code:

vector int foo(vector int m) {
   return m << 16 >> 16;
}

And it could be any vector type that hardware support the operation, though
the type 'v4i16' is NOT legal for the target. When we are trying to combine
the srl + sra, what we did now is calling the TLI.isOperationLegal(), which
will also check the legality of the type. That doesn't make sense.

Differential Revision: https://reviews.llvm.org/D70230

4 years ago[MC] Delete MCFragment::isDummy. NFC
Fangrui Song [Mon, 6 Jan 2020 02:44:22 +0000 (18:44 -0800)]
[MC] Delete MCFragment::isDummy. NFC

isa<...>, dyn_cast<...> and cast<...> are used by other fragments.
Don't make MCDummyFragment special.

4 years ago[X86] Improve v2i64->v2f32 and v4i64->v4f32 uint_to_fp on avx and avx2 targets.
Craig Topper [Mon, 6 Jan 2020 01:01:57 +0000 (17:01 -0800)]
[X86] Improve v2i64->v2f32 and v4i64->v4f32 uint_to_fp on avx and avx2 targets.

Summary:
Based on Simon's D52965, but improved to handle strict fp and improve some of the shuffling.

Rather than use v2i1/v4i1 and let type legalization continue, just generate all the code with legal types and use an explicit shuffle.

I also added an explicit setcc to the v4i64 code to match the semantics of vselect which doesn't just use the sign bit. I'm also using a v4i64->v4i32 truncate instead of the shuffle in Simon's original code. With the setcc this will become a pack.

Future work can look into using X86ISD::BLENDV and a different shuffle that only moves the sign bit.

Reviewers: RKSimon, spatel

Reviewed By: RKSimon

Subscribers: hiraditya, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D71956

4 years ago[NFC] Modify the format:
Liu, Chen3 [Mon, 6 Jan 2020 01:33:28 +0000 (09:33 +0800)]
[NFC] Modify the format:
Drop the else since we alerady returned in the if.

4 years ago[Coroutines] Remove corresponding phi values when apply simplifyTerminatorLeadingToRet
Brian Gesiak [Sun, 5 Jan 2020 22:07:57 +0000 (17:07 -0500)]
[Coroutines] Remove corresponding phi values when apply simplifyTerminatorLeadingToRet

Summary:
In addMustTailToCoroResumes, we set musttail on those resume instructions that are followed by a ret instruction. This is done by simplifyTerminatorLeadingToRet which replace a sequence of branches leading to a ret with a clone of the ret.

However it forgets to remove corresponding PHI values that come from basic block of replaced branch, and may cause jumpthreading pass hangs (https://bugs.llvm.org/show_bug.cgi?id=43720)

This patch fix this issue

Test Plan:
cppcoro library with O3+flto
check-llvm

Reviewers: modocache, GorNishanov, lewissbaker

Reviewed By: modocache

Subscribers: mehdi_amini, EricWF, hiraditya, dexonsmith, jfb, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D71826

Patch by junparser (JunMa)!

4 years agoClang-format previous commit
Stephen Kelly [Sun, 5 Jan 2020 22:58:03 +0000 (22:58 +0000)]
Clang-format previous commit

4 years ago[MC][ARM] Delete MCSection::HasData and move SHF_ARM_PURECODE logic to ARMELFObjectWr...
Fangrui Song [Sun, 5 Jan 2020 22:02:18 +0000 (14:02 -0800)]
[MC][ARM] Delete MCSection::HasData and move SHF_ARM_PURECODE logic to ARMELFObjectWriter::addTargetSectionFlags

This simplifies the generic interface and also makes SHF_ARM_PURECODE
more robust (fixes a TODO). Inspecting MCDataFragment contents covers
more cases than MCObjectStreamer::EmitBytes.

4 years agoAdd missing test
Stephen Kelly [Sun, 5 Jan 2020 21:55:13 +0000 (21:55 +0000)]
Add missing test

4 years ago[Gnu toolchain] Look at standard GCC paths for libstdcxx by default
Kristina Brooks [Sun, 5 Jan 2020 21:43:16 +0000 (21:43 +0000)]
[Gnu toolchain] Look at standard GCC paths for libstdcxx by default

Linux' current addLibCxxIncludePaths and addLibStdCxxIncludePaths
are actually almost non-Linux-specific at all, and can be reused
almost as such for all gcc toolchains. Only keep
Android/Freescale/Cray hacks in Linux's version.

Patch by sthibaul (Samuel Thibault)

Differential Revision: https://reviews.llvm.org/D69758

4 years ago[MC] Delete MCSection::{rbegin,rend}
Fangrui Song [Sun, 5 Jan 2020 20:48:48 +0000 (12:48 -0800)]
[MC] Delete MCSection::{rbegin,rend}

4 years agoAllow using traverse() with bindings
Stephen Kelly [Sun, 5 Jan 2020 20:48:20 +0000 (20:48 +0000)]
Allow using traverse() with bindings

4 years agoFix oversight in AST traversal helper
Stephen Kelly [Sun, 5 Jan 2020 20:27:37 +0000 (20:27 +0000)]
Fix oversight in AST traversal helper

4 years ago[MC] Merge MCSymbol::getSectionPtr into getSection and simplify
Fangrui Song [Sun, 5 Jan 2020 20:02:11 +0000 (12:02 -0800)]
[MC] Merge MCSymbol::getSectionPtr into getSection and simplify

4 years ago[MC] Drop an unused rule about absolute temporary symbols
Fangrui Song [Sun, 5 Jan 2020 19:39:46 +0000 (11:39 -0800)]
[MC] Drop an unused rule about absolute temporary symbols

4 years ago[X86][SSE] Combine combineLogicBlendIntoConditionalNegate for VSELECT nodes (PR43660)
Simon Pilgrim [Sun, 5 Jan 2020 18:50:44 +0000 (18:50 +0000)]
[X86][SSE] Combine combineLogicBlendIntoConditionalNegate for VSELECT nodes (PR43660)

Attempt to use combineLogicBlendIntoConditionalNegate for (select M, (sub 0, X), X) -> (sub (xor X, M), M)

We limit this to cases that can't easily replace the VSELECT with a shuffle (non-constant masks) or where a BLENDV is likely to occur (which tends to result in slower codegen).

4 years ago[X86] Move combineLogicBlendIntoConditionalNegate before combineSelect. NFCI.
Simon Pilgrim [Sun, 5 Jan 2020 17:17:41 +0000 (17:17 +0000)]
[X86] Move combineLogicBlendIntoConditionalNegate before combineSelect. NFCI.

Updates function order in preparation of future fix for PR43660

4 years ago[X86] Merge (identical) LowerGC_TRANSITION_START and LowerGC_TRANSITION_END (NFC)
Simon Pilgrim [Sun, 5 Jan 2020 15:24:23 +0000 (15:24 +0000)]
[X86] Merge (identical) LowerGC_TRANSITION_START and LowerGC_TRANSITION_END (NFC)

Silences a copy+paste analyzer warning - all they are doing are inserting NOOPs in exactly the same way.

4 years ago[ARM] Use isFMAFasterThanFMulAndFAdd for scalars as well as MVE vectors
David Green [Sun, 5 Jan 2020 10:59:21 +0000 (10:59 +0000)]
[ARM] Use isFMAFasterThanFMulAndFAdd for scalars as well as MVE vectors

This adds extra scalar handling to isFMAFasterThanFMulAndFAdd, allowing
the target independent code to handle more folds in more situations (for
example if the fast math flags are present, but the global
AllowFPOpFusion option isnt). It also splits apart the HasSlowFPVMLx
into HasSlowFPVFMx, to allow VFMA and VMLA to be controlled separately
if needed.

Differential Revision: https://reviews.llvm.org/D72139

4 years ago[ARM] Fill in FP16 FMA patterns
David Green [Sun, 5 Jan 2020 10:54:49 +0000 (10:54 +0000)]
[ARM] Fill in FP16 FMA patterns

This adds fp16 variants of all the fma patterns in the ARM backend.

Differential Revision: https://reviews.llvm.org/D72138

4 years ago[ARM] Add and update FMA tests. NFC
David Green [Sat, 4 Jan 2020 20:16:32 +0000 (20:16 +0000)]
[ARM] Add and update FMA tests. NFC

4 years ago[ParserTest] Move raw string literal out of macro
David Green [Sun, 5 Jan 2020 11:23:35 +0000 (11:23 +0000)]
[ParserTest] Move raw string literal out of macro

Some combinations of gcc and ccache do not deal well with raw strings in
macros. Moving the string out to attempt to fix the bots.

4 years ago[LegalizeVectorOps][X86] Enable expansion of vector fp_to_uint in LegalizeVectorOps...
Craig Topper [Sun, 5 Jan 2020 03:18:50 +0000 (19:18 -0800)]
[LegalizeVectorOps][X86] Enable expansion of vector fp_to_uint in LegalizeVectorOps to avoid scalarization.

The code here isn't great in all caess. Particularly v4f64->v4i32
on 64-bit AVX targets. But there is some improvement in some
configurations.

There's definitely some issues with computeNumSignBits with
X86ISD::STRICT_FCMP. As well as not being able to propagate sign
bits through merge_values nodes that get created during custom
legalization.

4 years ago[TargetLowering] In expandFP_TO_UINT, add proper extend or truncate for the condition...
Craig Topper [Sun, 5 Jan 2020 02:14:33 +0000 (18:14 -0800)]
[TargetLowering] In expandFP_TO_UINT, add proper extend or truncate for the condition to feed the DstVT select.

Previously, for vectors we created a vselect with a condition that
didn't match what the target wanted according to getSetCCResultType.

To make up for this, X86 had a special DAG combine to detect if
the condition was all sign bits and then insert its own truncate
or extend. By adding the extend/truncate here explicitly we can
avoid that.

4 years ago[LegalizeVectorOps] Split most of ExpandStrictFPOp into a separate UnrollStrictFPOp...
Craig Topper [Sun, 5 Jan 2020 01:03:33 +0000 (17:03 -0800)]
[LegalizeVectorOps] Split most of ExpandStrictFPOp into a separate UnrollStrictFPOp method. Call that method from ExpandUINT_TO_FLOAT.

ExpandStrictFPOp calls ExpandUINT_TO_FLOAT. Previously, ExpandUINT_TO_FLOAT
returned SDValue() if it wasn't able to handle and needed to unroll.
Then ExpandStrictFPOp would detect his SDValue() and do the unroll.

After this change, ExpandUINT_TO_FLOAT will directly call
UnrollStrictFPOp and return the unrolled result.

4 years ago[ELF] Drop const qualifier to fix -Wrange-loop-analysis. NFC
Fangrui Song [Sat, 4 Jan 2020 19:53:17 +0000 (11:53 -0800)]
[ELF] Drop const qualifier to fix -Wrange-loop-analysis. NFC

```
lld/ELF/Relocations.cpp:1622:56: warning: loop variable 'ts' of type 'const std::pair<ThunkSection *, uint32_t>' (aka 'const pair<lld::elf::ThunkSection *, unsigned int>') creates a copy from type 'const std::pair<ThunkSection *, uint32_t>' [-Wrange-loop-analysis]
        for (const std::pair<ThunkSection *, uint32_t> ts : isd->thunkSections)
```

Drop const qualifier to fix -Wrange-loop-analysis.
We can make -Wrange-loop-analysis warnings (DiagnoseForRangeConstVariableCopies) on `const A` more
permissive on more types (e.g. POD -> trivially copyable), unfortunately it will not make std::pair
good, because `constexpr pair& operator=(const pair& p);` is unfortunately user-defined.

Reviewed By: Mordante

Differential Revision: https://reviews.llvm.org/D72211

4 years agoGlobalISel: Scalarize all division operations
Matt Arsenault [Sat, 4 Jan 2020 18:24:09 +0000 (13:24 -0500)]
GlobalISel: Scalarize all division operations

This only handled G_SDIV, but they all are trivially scalarizable.

Also define placeholder AMDGPU division legalizer rules.

4 years agoRevert "[SCEV] Move ScalarEvolutionExpander.cpp to Transforms/Utils (NFC)."
Florian Hahn [Sat, 4 Jan 2020 18:44:38 +0000 (18:44 +0000)]
Revert "[SCEV] Move ScalarEvolutionExpander.cpp to Transforms/Utils (NFC)."

This reverts commit 51ef53f3bd23559203fe9af82ff2facbfedc1db3, as it
breaks some bots.

4 years ago[SCEV] Move ScalarEvolutionExpander.cpp to Transforms/Utils (NFC).
Florian Hahn [Sat, 4 Jan 2020 18:15:02 +0000 (18:15 +0000)]
[SCEV] Move ScalarEvolutionExpander.cpp to Transforms/Utils (NFC).

SCEVExpander modifies the underlying function so it is more suitable in
Transforms/Utils, rather than Analysis. This allows using other
transform utils in SCEVExpander.

Reviewers: sanjoy.google, efriedma, reames

Reviewed By: sanjoy.google

Differential Revision: https://reviews.llvm.org/D71537

4 years ago[SCEV] Remove unused ScalarEvolutionExpander.h includes (NFC).
Florian Hahn [Sat, 4 Jan 2020 18:20:55 +0000 (18:20 +0000)]
[SCEV] Remove unused ScalarEvolutionExpander.h includes (NFC).

4 years agoGlobalISel: Define G_READCYCLECOUNTER
Matt Arsenault [Sat, 4 Jan 2020 17:46:58 +0000 (12:46 -0500)]
GlobalISel: Define G_READCYCLECOUNTER

4 years agoAMDGPU/GlobalISel: Refine SMRD selection rules
Matt Arsenault [Fri, 27 Dec 2019 15:17:45 +0000 (10:17 -0500)]
AMDGPU/GlobalISel: Refine SMRD selection rules

Fix selecting these for volatile global loads, and ensure the loads
are constant enough.

4 years agoAMDGPU/GlobalISel: Legalize more odd sized loads
Matt Arsenault [Thu, 2 Jan 2020 19:45:11 +0000 (14:45 -0500)]
AMDGPU/GlobalISel: Legalize more odd sized loads

The attempts to widen sufficently aligned, odd sized loads wasn't
consistently applied.

4 years agoAMDGPU/GlobalISel: Assume vcc phis for any vcc input
Matt Arsenault [Mon, 23 Dec 2019 23:34:59 +0000 (18:34 -0500)]
AMDGPU/GlobalISel: Assume vcc phis for any vcc input

This produces more intelligible looking results, more comparabble to
the DAG output in the simplest cases. This is probably wrong in
complex control flow, but RegBankSelect doesn't attempt analyzing if
this is on a masked path for selecting the bank yet.

4 years ago[Pass Registration] XFAIL load_extension.ll test on macOS.
Florian Hahn [Sat, 4 Jan 2020 17:32:15 +0000 (17:32 +0000)]
[Pass Registration] XFAIL load_extension.ll test on macOS.

This test fails  on macOS, causing the following bots to fail

http://green.lab.llvm.org/green/job/clang-stage1-cmake-RA-incremental/7438/
http://green.lab.llvm.org/green/job/clang-stage1-RA/5034/

Error:
Error opening 'build/./lib/libBye.dylib': dlopen(build/./lib/libBye.dylib, 9): image not found
  -load request ignored.

4 years agoAMDGPU/GlobalISel: Implement applyMappingImpl less incorrectly
Matt Arsenault [Mon, 23 Dec 2019 19:39:01 +0000 (14:39 -0500)]
AMDGPU/GlobalISel: Implement applyMappingImpl less incorrectly

We're checking the current register bank of the registers in the
instruction, but the mapping may have inserted cross bank copies and
is expecting to replace the registers.

We mostly get away with this currently, because VGPR->SGPR copies are
illegal, and we assume this won't happen. In a future change, we'll
start relying on more cross register bank copies being inserted, and
this starts to break down.