Craig Topper [Tue, 5 Dec 2017 17:37:19 +0000 (17:37 +0000)]
[SelectionDAG] Remove the code that handles SETCC with a scalar result type from vector widening.
There's no such thing as a setcc with vector operands and scalar result. And if we're trying to widen the result we would have to already be looking at a vector result type.
So this patch renames the VSETCC function as the SETCC function and delete the original SETCC function.
Without this when lld failed to replace the output file it would leave
the temporary behind. The problem is that the existing logic is
- cancel the delete flag
- rename
We have to cancel first to avoid renaming and then crashing and
deleting the old version. What is missing then is deleting the
temporary file if the rename fails.
This can be an issue on both unix and windows, but I am not sure how
to cause the rename to fail reliably on unix. I think it can be done
on ZFS since it has an ACL system similar to what windows uses, but
adding support for checking that in llvm-lit is probably not worth it.
Sam Parker [Tue, 5 Dec 2017 15:13:47 +0000 (15:13 +0000)]
[DAGCombine] Move AND nodes to multiple load leaves
Search from AND nodes to find whether they can be propagated back to
loads, so that the AND and load can be combined into a narrow load.
We search through OR, XOR and other AND nodes and all bar one of the
leaves are required to be loads or constants. The exception node then
needs to be masked off meaning that the 'and' isn't removed, but the
loads(s) are narrowed still.
[DAGCombine] Handle big endian correctly in CombineConsecutiveLoads
Summary:
Found out, at code inspection, that there was a fault in
DAGCombiner::CombineConsecutiveLoads for big-endian targets.
A BUILD_PAIR is always having the least significant bits of
the composite value in element 0. So when we are doing the checks
for consecutive loads, for big endian targets, we should check
if the load to elt 1 is at the lower address and the load
to elt 0 is at the higher address.
Normally this bug only resulted in missed oppurtunities for
doing the load combine. I guess that in some rare situation it
could lead to faulty combines, but I've not seen that happen.
Note that this patch actually will trigger load combine for
some big endian regression tests.
One example is test/CodeGen/PowerPC/anon_aggr.ll where we now get
t76: i64,ch = load<LD8[FixedStack-9]
instead of
t37: i32,ch = load<LD4[FixedStack-10]>
t35: i32,ch = load<LD4[FixedStack-9]>
t41: i64 = build_pair t37, t35
before legalization. Then the legalization will split the LD8
into two loads, so the end result is the same. That should
verify that the transfomation is correct now.
Mikael Holmen [Tue, 5 Dec 2017 14:14:00 +0000 (14:14 +0000)]
Bail out of a SimplifyCFG switch table opt at undef values.
Summary:
A true or false result is expected from a comparison, but it seems the possibility of undef was overlooked, which could lead to a failed assert. This is fixed by this patch by bailing out if we encounter undef.
The bug is old and the assert has been there since the end of 2014, so it seems this is unusual enough to forego optimization.
[XRay][docs] Document xray_mode and log registration API.
This marks certain flags in XRay as deprecated (in particular,
`xray_naive_log=` and `xray_fdr_log=`), and recommends the use of the
`xray_mode=` flag.
Simon Pilgrim [Tue, 5 Dec 2017 12:14:36 +0000 (12:14 +0000)]
[X86][AVX512] Tag VPCMP/VPCMPU instruction scheduler classes
Move hardcoded itinerary out to the instruction declarations. Not sure that IIC_SSE_ALU_F32P is the best schedule for integer comparisons, but I'm not going to change it right now.
Simon Pilgrim [Tue, 5 Dec 2017 12:02:22 +0000 (12:02 +0000)]
[X86][AVX512] Cleanup VPCMP scheduler classes
Move hardcoded itinerary out to the instruction declarations. Not sure that IIC_SSE_ALU_F32P is the best schedule for integer comparisons, but I'm not going to change it right now.
Jonas Paulsson [Tue, 5 Dec 2017 11:24:39 +0000 (11:24 +0000)]
[SystemZ] set 'guessInstructionProperties = 0' and set flags as needed.
This has proven a healthy exercise, as many cases of incorrect instruction
flags were corrected in the process. As part of this, IntrWriteMem was added
to several SystemZ instrinsics.
Furthermore, a bug was exposed in TwoAddress with this change (as incorrect
hasSideEffects flags were removed and instructions could now be sunk), and
the test case for that bugfix (r319646) is included here as
test/CodeGen/SystemZ/twoaddr-sink.ll.
One temporary test regression (one extra copy) which will hopefully go away
in upcoming patches for similar cases:
test/CodeGen/SystemZ/vec-trunc-to-i1.ll
Jonas Paulsson [Tue, 5 Dec 2017 10:52:24 +0000 (10:52 +0000)]
[Regalloc] Generate and store multiple regalloc hints.
MachineRegisterInfo used to allow just one regalloc hint per virtual
register. This patch extends this to a vector of regalloc hints, which is
filled in by common code with sorted copy hints. Such hints will make for
more ID copies that can be removed.
NB! This improvement is currently (and hopefully temporarily) *disabled* by
default, except for SystemZ. The only reason for this is the big impact this
has on tests, which has unfortunately proven unmanageable. It was a long
while since all the tests were updated and just waiting for review (which
didn't happen), but now targets have to enable this themselves
instead. Several targets could get a head-start by downloading the tests
updates from the Phabricator review. Thanks to those who helped, and sorry
you now have to do this step yourselves.
This should be an improvement generally for any target!
The target may still create its own hint, in which case this has highest
priority and is stored first in the vector. If it has target-type, it will
not be recomputed, as per the previous behaviour.
The temporary hook enableMultipleCopyHints() will be removed as soon as all
targets return true.
Pavel Labath [Tue, 5 Dec 2017 10:24:15 +0000 (10:24 +0000)]
Re-commit "[cmake] Enable zlib support on windows"
This recommits r319533 which was broken llvm-config --system-libs
output. The reason was that I used find_libraries for searching for the
z library. This returns absolute paths, and when these paths made it
into llvm-config, it made it produce nonsensical flags. To fix this, I
hand-roll a search for the library in the same way that we search for
the terminfo library a couple of lines below.
This is a bit less flexible than the find_library option, as it does not
allow the user to specify the path to the library at configure time
(which is important on windows, as zlib is unlikely to be found in any
of the standard places cmake searches), but I was able to guide the
build to find it with appropriate values of LIB and INCLUDE environment
variables.
George Rimar [Tue, 5 Dec 2017 10:09:59 +0000 (10:09 +0000)]
[Support/TarWriter] - Don't allow TarWriter to add the same file more than once.
This is for PR35460.
Currently when LLD adds files to TarWriter it may pass the same file
multiple times. For example it happens for clang reproduce file which specifies
archive (.a) files more than once in command line.
Patch makes TarWriter to ignore files with the same path, so it will
add only the first one to archive.
Guy Blank [Tue, 5 Dec 2017 09:08:24 +0000 (09:08 +0000)]
[X86] Fix a bug in handling GRXX subclasses in Domain Reassignment pass
When trying to determine the correct Mask register class corresponding
to a GPR register class, not all register classes were handled.
This caused an assertion to be raised on some scenarios.
Craig Topper [Tue, 5 Dec 2017 08:15:03 +0000 (08:15 +0000)]
[SelectionDAG] Use WidenTargetBoolean in WidenVecRes_MLOAD and WidenVecOp_MSTORE instead of implementing it manually and incorrectly.
The CONCAT_VECTORS operand get its type from getSetCCResultType, but if the mask type and the setcc have different scalar sizes this creates an illegal CONCAT_VECTORS operation. The concat type should be 2x the mask type, and then an extend should be added if needed.
Matt Arsenault [Tue, 5 Dec 2017 03:09:23 +0000 (03:09 +0000)]
AMDGPU: Fix crash when scheduling DBG_VALUE
This calls handleMove with a DBG_VALUE instruction,
which isn't tracked by LiveIntervals. I'm not sure
this is the correct place to fix this. The generic
scheduler seems to have more deliberate region
selection that skips dbg_value.
The test is also really hard to reduce. I haven't been able
to figure out what exactly causes this particular case to
try moving the dbg_value.
Craig Topper [Tue, 5 Dec 2017 01:28:04 +0000 (01:28 +0000)]
[X86] Use getZeroVector and remove an unnecessary creation of an APInt before calling getConstant. NFCI
The getConstant function can take care of creating the APInt internally.
getZeroVector will take care of using the correct type for the build vector to avoid re-lowering.
The test change here is because execution domain constraints apparently pass through undef inputs of a zeroing xor. So the different ordering of register allocation here caused the dependency to change.
Craig Topper [Tue, 5 Dec 2017 01:28:00 +0000 (01:28 +0000)]
[X86] Rearrange some of the code around AVX512 sign/zero extends. NFCI
Move the AVX512 code out of LowerAVXExtend. LowerAVXExtend has two callers but one of them pre-checks for AVX-512 so the code is only live from the other caller. So move the AVX-512 checks up to that caller for symmetry.
Move all of the i1 input type code in Lower_AVX512ZeroExend together.
Shoaib Meenai [Tue, 5 Dec 2017 01:19:48 +0000 (01:19 +0000)]
[cmake] Modernize some conditionals. NFC
The "x${...}" form was a workaround for CMake versions prior to 3.1,
where the if command would interpret arguments as variables even when
quoted [1]. We can drop the workaround now that our minimum CMake
version is 3.4.
Matthias Braun [Tue, 5 Dec 2017 00:56:14 +0000 (00:56 +0000)]
TwoAddressInstructionPass: Trigger -O0 behavior on optnone
While we cannot skip the whole TwoAddressInstructionPass even for -O0
there are some parts of the pass that are currently skipped at -O0 but
not for optnone. Changing this as there is no reason to have those two
hit different code paths here.
Matt Arsenault [Mon, 4 Dec 2017 22:18:27 +0000 (22:18 +0000)]
AMDGPU: Fix creating invalid copy when adjusting dmask
Move the entire optimization to one place. Before it was possible
to adjust dmask without changing the register class of the output
instruction, since they were done in separate places. Fix all
lane sizes and move all of the optimization into the DAG folding.
Daniel Sanders [Mon, 4 Dec 2017 21:40:57 +0000 (21:40 +0000)]
Allow similar TargetOpcodes to use inheritance to factor out commonality. NFC.
Summary:
While implementing atomicrmw in https://reviews.llvm.org/D40092 I found that
inheritance is unusable for all the Generic Opcodes in GlobalISel. This is
because the whole header is included inside a 'let mayLoad = 0, mayStore = 0 ... in'
block. In TableGen, the order of precedence for field assignments is:
1. Values from classes the record inherits from.
2. Values from 'let Name=Value in { ... }'
3. Values from 'let Name=Value;'
As such the 'let mayLoad = 0, mayStore = 0, ... in' surrounding the
'include "GenericOpcodes.td"' was overriding any values provided via inheritance.
We hadn't noticed this before because we were only using 'let Name=Value;' to
specialize opcodes.
Fix this by moving the default values to the lowest precedence. This is
accomplished by moving the values to a common base class
(StandardPseudoInstruction for most TargetOpcodes, and GenericOpcode for
GlobalISel specific TargetOpcodes)
Hans Wennborg [Mon, 4 Dec 2017 20:48:50 +0000 (20:48 +0000)]
DAG: Follow-up to r319692 check the truncates inputs have the same type
MatchRotate assumes the types of the types of LHS and RHS are equal,
which is always the case then they come from an OR node, but here
we're getting them from two different TRUNC nodes, so we have to check
the types.
Daniel Sanders [Mon, 4 Dec 2017 20:39:32 +0000 (20:39 +0000)]
[globalisel][tablegen] Split atomic load/store into separate opcode and enable for AArch64.
This patch splits atomics out of the generic G_LOAD/G_STORE and into their own
G_ATOMIC_LOAD/G_ATOMIC_STORE. This is a pragmatic decision rather than a
necessary one. Atomic load/store has little in implementation in common with
non-atomic load/store. They tend to be handled very differently throughout the
backend. It also has the nice side-effect of slightly improving the common-case
performance at ISel since there's no longer a need for an atomicity check in the
matcher table.
All targets have been updated to remove the atomic load/store check from the
G_LOAD/G_STORE path. AArch64 has also been updated to mark
G_ATOMIC_LOAD/G_ATOMIC_STORE legal.
There is one issue with this patch though which also affects the extending loads
and truncating stores. The rules only match when an appropriate G_ANYEXT is
present in the MIR. For example,
(G_ATOMIC_STORE (G_TRUNC:s16 (G_ANYEXT:s32 (G_ATOMIC_LOAD:s16 X))))
will match but:
(G_ATOMIC_STORE (G_ATOMIC_LOAD:s16 X))
will not. This shouldn't be a problem at the moment, but as we get better at
eliminating extends/truncates we'll likely start failing to match in some
cases. The current plan is to fix this in a patch that changes the
representation of extending-load/truncating-store to allow the MMO to describe
a different type to the operation.
Haicheng Wu [Mon, 4 Dec 2017 19:56:33 +0000 (19:56 +0000)]
[ConstantFold] Support vector index when factoring out GEP index into preceding dimensions
Follow-up of r316824. This patch supports the vector type for both current and
previous index when factoring out the current one into the previous one.
Sanjoy Das [Mon, 4 Dec 2017 19:22:00 +0000 (19:22 +0000)]
[SCEV] A different fix for PR33494
Summary:
I don't think rL309080 is the right fix for PR33494 -- caching ExitLimit only
hides the problem[0]. The real issue is that because of how we forget SCEV
expressions ScalarEvolution::getBackedgeTakenInfo, in the test case for PR33494
computing the backedge for any loop invalidates the trip count for every other
loop. This effectively makes the SCEV cache useless.
I've instead made the SCEV expression invalidation in
ScalarEvolution::getBackedgeTakenInfo less aggressive to fix this issue.
[0]: One way to think about this is that rL309080 essentially augmented the
backedge-taken-count cache with another equivalent exit-limit cache. The bug
went away because we were explicitly not clearing the exit-limit cache in
getBackedgeTakenInfo. But instead of doing all of that, we can just avoid
clearing the backedge-taken-count cache.
Sanjoy Das [Mon, 4 Dec 2017 19:21:58 +0000 (19:21 +0000)]
[BypassSlowDivision] Improve our handling of divisions by constants
(This reapplies r314253. r314253 was reverted on r314482 because of a
correctness regression on P100, but that regression was identified to be
something else.)
Summary:
Don't bail out on constant divisors for divisions that can be narrowed without
introducing control flow . This gives us a 32 bit multiply instead of an
emulated 64 bit multiply in the generated PTX assembly.
Pablo Barrio [Mon, 4 Dec 2017 16:55:49 +0000 (16:55 +0000)]
Fix function pointer tail calls in armv8-M.base
Summary:
The compiler fails with the following error message:
fatal error: error in backend: ran out of registers during
register allocation
Tail call optimization for Armv8-M.base fails to meet all the required
constraints when handling calls to function pointers where the
arguments take up r0-r3. This is because the pointer to the
function to be called can only be stored in r0-r3, but these are
all occupied by arguments. This patch makes sure that tail call
optimization does not try to handle this type of calls.
Pavel Labath [Mon, 4 Dec 2017 16:46:20 +0000 (16:46 +0000)]
Revert "[cmake] Enable zlib support on windows"
This reverts commit r319533 as it broke llvm-config --system-libs output
and everything that depends on it (which is mostly out of tree or
downstream folks, but includes a couple of llvm buildbots as well).
I think I have a fix for this in D40779, but I want someone to look
review it first. In the mean time, I am reverting this change, as it
seems to break a lot of people.
Anna Thomas [Mon, 4 Dec 2017 15:11:48 +0000 (15:11 +0000)]
[Loop Predication] Teach LP about reverse loops
Summary:
Currently, we only support predication for forward loops with step
of 1. This patch enables loop predication for reverse or
countdownLoops, which satisfy the following conditions:
1. The step of the IV is -1.
2. The loop has a singe latch as B(X) = X <pred>
latchLimit with pred as s> or u>
3. The IV of the guard is the decrement
IV of the latch condition (Guard is: G(X) = X-1 u< guardLimit).
This patch was downstream for a while and is the last series of patches
that's from our LP implementation downstream.
Jonas Hahnfeld [Mon, 4 Dec 2017 14:19:33 +0000 (14:19 +0000)]
[NVPTX] Assign valid global names
PTX requires that identifiers consist only of [a-zA-Z0-9_$]. The
existing pass already ensured this for globals and this patch adds
the cleanup for functions with local linkage.
However, there was a different problem in the case of collisions
of the adjusted name: The ValueSymbolTable then automatically
appended ".N" with increasing Ns to get a unique name while helping
the ABI demangling. Special case this behavior to omit the dots and
append N directly. This will always give us legal names according
to the PTX requirements.
[NFC][lit] Use proper semantic versioning names for variables
The variable named `minor` was actually pointing to the patch part of
the version. While I was changing this I also made the check for Apple
clang more robust by checking both patch and minor rather than just
minor.
This is causing a failure in the llvm-clang-x86_64-expensive-checks-win
buildbot, and I can't reproduce it locally, so reverting until I can work out
what is wrong.
Tim Corringham [Mon, 4 Dec 2017 12:30:49 +0000 (12:30 +0000)]
AMDGPU: fix missing s_waitcnt
Summary:
The pass that inserts s_waitcnt instructions where needed propagated
info used to track dependencies for each block by iterating over the
predecessor blocks. The iteration was terminated when a predecessor
that had not yet been processed was encountered. Any info in blocks
later in the list was therefore not processed, leading to the
possiblility of a required s_waitcnt not being inserted.
The fix is simply to change the "break" to "continue" for the
relevant loops, so that all visited blocks are processed. This
is likely what was intended when the code was written.
There is no test case provided for this fix because:
1) the only example that reproduces this is large and resistant to
being reduced
2) the change is trivial
Oliver Stannard [Mon, 4 Dec 2017 12:02:32 +0000 (12:02 +0000)]
[Asm, ARM] Add fallback diag for multiple invalid operands
This adds a "invalid operands for instruction" diagnostic for
instructions where there is an instruction encoding with the correct
mnemonic and which is available for this target, but where multiple
operands do not match those which were provided. This makes it clear
that there is some combination of operands that is valid for the current
target, which the default diagnostic of "invalid instruction" does not.
Since this is a very general error, we only emit it if we don't have a
more specific error.
Jonas Paulsson [Mon, 4 Dec 2017 10:03:14 +0000 (10:03 +0000)]
[TwoAddressInstructionPass] Bugfix in handling of sunk instructions.
An instruction returned by TII->convertToThreeAddress() may contain a %noreg
(undef) operand, which is not expected by tryInstructionTransform(). So if
this MI is sunk to a lower point in MBB, it must be skipped when later
encountered.
A new set SunkInstrs is used for this purpose.
Note: there is no test supplied here, as this was triggered on SystemZ while
working on a review of instruction flags. A test case for this bugfix will be
included in the upcoming SystemZ commit.
Martin Storsjo [Mon, 4 Dec 2017 09:09:04 +0000 (09:09 +0000)]
[AArch64] Allow using emulated tls on platforms other than ELF
This matches how it is done on X86.
This allows using emulated tls on windows; in MinGW environments,
native tls isn't supported at the moment.
Set the right Data*bitsDirective for windows to match the existing
tests for other platforms. Make parts of the existing tests a regex,
to allow matching .section .rdata for windows, to avoid having to
duplicate the rest of the tests for windows.
Craig Topper [Mon, 4 Dec 2017 05:38:44 +0000 (05:38 +0000)]
[X86] Don't turn UINT_TO_FP into SINT_TO_FP during lowering.
We already do this as a DAG combine. The version during lowering can only trigger if known bits changes something that improves known bits analysis. But this means we should be improving known bits analysis to work on the unlowered form instead.
Craig Topper [Mon, 4 Dec 2017 05:38:42 +0000 (05:38 +0000)]
[SelectionDAG] Teach computeKnownBits some improvements to ISD::SRL with a non-splat constant shift amount.
If we have a non-splat constant shift amount, the minimum shift amount can be used to infer the number of zero upper bits of the result. There's probably a lot more that we can do here, but this
fixes a case where I wanted to infer the sign bit as zero when all the shift amounts are non-zero.
Yaxun Liu [Sat, 2 Dec 2017 22:13:22 +0000 (22:13 +0000)]
CodeGen: Fix pointer info in SplitVecOp_EXTRACT_VECTOR_ELT/SplitVecRes_INSERT_VECTOR_ELT
Two issues found when doing codegen for splitting vector with non-zero alloca addr space:
DAGTypeLegalizer::SplitVecRes_INSERT_VECTOR_ELT/SplitVecOp_EXTRACT_VECTOR_ELT uses dummy pointer info for creating
SDStore. Since one pointer operand contains multiply and add, InferPointerInfo is unable to
infer the correct pointer info, which ends up with a dummy pointer info for the target to lower
store and results in isel failure. The fix is to introduce MachinePointerInfo::getUnknownStack to
represent MachinePointerInfo which is known in alloca address space but without other information.
TargetLowering::getVectorElementPointer uses value type of pointer in addr space 0 for
multiplication of index and then add it to the pointer. However the pointer may be in an addr
space which has different size than addr space 0. The fix is to use the pointer value type for
index multiplication.
Don Hinton [Sat, 2 Dec 2017 19:12:38 +0000 (19:12 +0000)]
[cmake] Re-commit: Remove redundant call to cmake when building host tools.
Also pass CMAKE_(C|CXX)_COMPILER to add_custom_command.
Summary:
Remove the redundant, config-time call to cmake when
building host tools for cross compiles or optimized tablegen..
The config-time call to cmake is redundant because it will always get
called again when the CONFIGURE_LLVM_${target_name} target fires at
build-time. This speeds up initial configuration, but has no affect
on build behavior.
projects / llvm / log