Chandler Carruth [Wed, 10 May 2017 12:30:07 +0000 (12:30 +0000)]
Revert r301950: SpeculativeExecution: Stop using whitelist for costs
This pass doesn't correctly handle testing for when it is legal to hoist
arbitrary instructions. The whitelist happens to make it safe, so before
it is removed the pass's legality checks will need to be enhanced.
Details have been added to the code review thread for the patch.
Martin Storsjo [Wed, 10 May 2017 10:51:32 +0000 (10:51 +0000)]
[AArch64] Fix a comment to match the code. NFC.
For the ELF case, the default/preferred form is the generic one, not
the short one as used for Apple - fix the comment to say so. Currently
it is a copy-paste typo.
Make the comments on the darwin default a bit more verbose.
Use enum names instead of literal 0/1 to further increase readability
and reduce fragility.
Mikael Holmen [Wed, 10 May 2017 06:33:43 +0000 (06:33 +0000)]
[UnreachableBlockElim] Check return value of constrainRegClass().
Summary:
MachineRegisterInfo::constrainRegClass() can fail if two register classes
don't have a common subclass or if the register class doesn't contain
enough registers. Check the return value before trying to remove Phi nodes,
and if we can't constrain, we output a COPY instead of simply replacing
registers.
Ahmed Bougacha [Wed, 10 May 2017 00:39:30 +0000 (00:39 +0000)]
[CodeGen] Don't require AA in SDAGISel at -O0.
Before r247167, the pass manager builder controlled which AA
implementations were used, exporting them all in the AliasAnalysis
analysis group.
Now, AAResultsWrapperPass always uses BasicAA, but still uses other AA
implementations if made available in the pass pipeline.
But regardless, SDAGISel is required at O0, and really doesn't need to
be doing fancy optimizations based on useful AA results.
Don't require AA at CodeGenOpt::None, and only use it otherwise.
This does have a functional impact (and one testcase is pessimized
because we can't reuse a load). But I think that's desirable no matter
what.
Note that this alone doesn't result in less DT computations: TwoAddress
was previously able to reuse the DT we computed for SDAG. That will be
fixed separately.
Ahmed Bougacha [Wed, 10 May 2017 00:39:25 +0000 (00:39 +0000)]
[CodeGen] Compute DT/LI lazily in SafeStackLegacyPass. NFC.
We currently require SCEV, which requires DT/LI. Those are expensive to
compute, but the pass only runs for functions that have the safestack
attribute.
Compute DT/LI to build SCEV lazily, only when the pass is actually going
to transform the function.
Ahmed Bougacha [Wed, 10 May 2017 00:39:17 +0000 (00:39 +0000)]
[CodeGen] Add an -O0 backend pipeline test. NFC.
This should hopefully makes changes to the O0 pipeline obvious; it's
easy to require expensive passes, and this helps make informed
decisions.
Case in point: in the few weeks separating the time when I initially
wrote this patch to the time when I committed, the test regressed as
r302103 added another use of DT!
Sam Clegg [Wed, 10 May 2017 00:14:04 +0000 (00:14 +0000)]
[WebAssembly] Fix build error in wasm YAML code
This warning didn't show up on my local build
but is causing the bots to fail. Seems like a
bad idea to have types and variables with the
same name anyhow.
Easwaran Raman [Tue, 9 May 2017 23:21:10 +0000 (23:21 +0000)]
[ProfileSummary] Make getProfileCount a non-static member function.
This change is required because the notion of count is different for
sample profiling and getProfileCount will need to determine the
underlying profile type.
Lang Hames [Tue, 9 May 2017 21:32:18 +0000 (21:32 +0000)]
[ExecutionEngine] Make RuntimeDyld::MemoryManager responsible for tracking EH
frames.
RuntimeDyld was previously responsible for tracking allocated EH frames, but it
makes more sense to have the RuntimeDyld::MemoryManager track them (since the
frames are allocated through the memory manager, and written to memory owned by
the memory manager). This patch moves the frame tracking into
RTDyldMemoryManager, and changes the deregisterFrames method on
RuntimeDyld::MemoryManager from:
Separating this responsibility will allow ORC to continue to throw the
RuntimeDyld instances away post-link (saving a few dozen bytes per lazy
function) while properly deregistering frames when modules are unloaded.
This patch also updates ORC to call deregisterEHFrames when modules are
unloaded. This fixes a bug where an exception that tears down the JIT can then
unwind through dangling EH frames that have been deallocated but not
deregistered, resulting in UB.
For people using SectionMemoryManager this should be pretty much a no-op. For
people with custom allocators that override registerEHFrames/deregisterEHFrames,
you will now be responsible for tracking allocated EH frames.
Keno Fischer [Tue, 9 May 2017 21:07:20 +0000 (21:07 +0000)]
[GVN] Fix a crash on encountering non-integral pointers
Summary:
This fixes the immediate crash caused by introducing an incorrect inttoptr
before attempting the conversion. There may still be a legality
check missing somewhere earlier for non-integral pointers, but this change
seems necessary in any case.
Matthew Simpson [Tue, 9 May 2017 20:18:12 +0000 (20:18 +0000)]
[AArch64] Consider widening instructions in cost calculations
The AArch64 instruction set has a few "widening" instructions (e.g., uaddl,
saddl, uaddw, etc.) that take one or more doubleword operands and produce
quadword results. The operands are automatically sign- or zero-extended as
appropriate. However, in LLVM IR, these extends are explicit. This patch
updates TTI to consider these widening instructions as single operations whose
cost is attached to the arithmetic instruction. It marks extends that are part
of a widening operation "free" and applies a sub-target specified overhead
(zero by default) to the arithmetic instructions.
Sanjay Patel [Tue, 9 May 2017 20:05:05 +0000 (20:05 +0000)]
[InstCombine] clean up matchDeMorgansLaws(); NFCI
The motivation for getting rid of dyn_castNotVal is to allow fixing:
https://bugs.llvm.org/show_bug.cgi?id=32706
So this was supposed to be functional-change-intended for the case
of inverting constants and applying DeMorgan. However, I can't find
any cases where that pattern will actually get to matchDeMorgansLaws()
because we have other folds in visitAnd/visitOr that do the same
thing. So this ends up just being a clean-up patch with slight efficiency
improvement, but no-functional-change-intended.
Reid Kleckner [Tue, 9 May 2017 19:59:29 +0000 (19:59 +0000)]
[codeview] Check for a DIExpression offset for local variables
Fixes inalloca parameters, which previously all pointed to the same
offset. Extend the test to use llvm-readobj so that we can test the
offset in a readable way.
Adrian Prantl [Tue, 9 May 2017 19:47:37 +0000 (19:47 +0000)]
Make it illegal for two Functions to point to the same DISubprogram
As recently discussed on llvm-dev [1], this patch makes it illegal for
two Functions to point to the same DISubprogram and updates
FunctionCloner to also clone the debug info of a function to conform
to the new requirement. To simplify the implementation it also factors
out the creation of inlineAt locations from the Inliner into a
general-purpose utility in DILocation.
Puyan Lotfi [Tue, 9 May 2017 17:13:37 +0000 (17:13 +0000)]
Adding VSCode syntax colorizer to utils (generated from textmate colorizer).
--This line, and those below, will be igored--
A utils/vscode
A utils/vscode/README
A utils/vscode/tablegen
A utils/vscode/tablegen/.vscode
A utils/vscode/tablegen/.vscode/launch.json
A utils/vscode/tablegen/CHANGELOG.md
A utils/vscode/tablegen/README.md
A utils/vscode/tablegen/language-configuration.json
A utils/vscode/tablegen/package.json
A utils/vscode/tablegen/syntaxes
A utils/vscode/tablegen/syntaxes/TableGen.tmLanguage
A utils/vscode/tablegen/vsc-extension-quickstart.md
and the in-memory representations of the two SSA registers have an
inconsistent order. This violation of NewGVN assumptions results into
two PHIs found congruent when they're not. While we think it's useful
to have always a consistent order enforced, let's fix this in NewGVN
sorting uses in predecessor order before creating a PHI expression.
Craig Topper [Tue, 9 May 2017 16:47:33 +0000 (16:47 +0000)]
[APInt] Remove return value from tcFullMultiply.
The description says it returns the number of words needed to represent the results. But the way it was coded it always returns (lhsWords + rhsWords) or (lhsWords + rhsWords - 1). But the result could be even smaller than that and it wouldn't tell you.
No one uses the result today so rather than try to fix it, just remove it.
Craig Topper [Tue, 9 May 2017 16:32:11 +0000 (16:32 +0000)]
[X86] Add more patterns for BZHI isel
This patch adds more patterns that a reasonable person might write that can be compiled to BZHI.
This adds support for
(~0U >> (32 - b)) & a;
and
a << (32 - b) >> (32 - b);
This was inspired by the code in APInt::clearUnusedBits.
This can pass an index of 32 to the bzhi instruction which a quick test of Haswell hardware shows will not mask any bits. Though the description text in the Intel manual says the "index is saturated to OperandSize-1". The pseudocode in the same manual indicates no bits will be zeroed for this case.
I think this is still missing cases where the subtract portion is an 8-bit operation.
Sanjay Patel [Tue, 9 May 2017 16:24:59 +0000 (16:24 +0000)]
[InstCombineCasts] Fix checks in sext->lshr->trunc pattern.
The comment says to avoid the case where zero bits are shifted into the truncated value,
but the code checks that the shift is smaller than the truncated value instead of the
number of bits added by the sign extension. Fixing this allows a shift by more than the
value size to be introduced, which is undefined behavior, so the shift is capped at the
value size minus one, which has the expected behavior of filling the value with the sign
bit.
Guy Blank [Tue, 9 May 2017 16:16:48 +0000 (16:16 +0000)]
VX512] Only look at lower bit in constant scalar masks
for scalar masked instructions only the lower bit of the mask is relevant. so for constant masks we should either do an unmasked operation or no operation, depending on the value of the lower bit.
This patch handles cases where the lower bit is '1'.
Tim Shen [Tue, 9 May 2017 15:27:17 +0000 (15:27 +0000)]
[Atomic] Remove IsStore/IsLoad in the interface, and pass the instruction instead. NFC.
Now both emitLeadingFence and emitTrailingFence take the instruction
itself, instead of taking IsLoad/IsStore pairs.
Instruction::mayReadFromMemory and Instrucion::mayWriteToMemory are used
for determining those two booleans.
The instruction argument is also useful for later D32763, in
emitTrailingFence. For emitLeadingFence, it seems to have cleaner
interface with the proposed change.
Aaron Ballman [Tue, 9 May 2017 14:59:48 +0000 (14:59 +0000)]
ARMRegisterBankInfo.h requires LLVM_BUILD_GLOBAL_ISEL to be defined. If it is not defined, then ARMGenRegisterBank.inc is not table generated and the inclusion of this header causes the build to fail.
Hans Wennborg [Tue, 9 May 2017 14:44:15 +0000 (14:44 +0000)]
Revert r302469 "Make it illegal for two Functions to point to the same DISubprogram"
This caused PR32977.
Original commit message:
> Make it illegal for two Functions to point to the same DISubprogram
>
> As recently discussed on llvm-dev [1], this patch makes it illegal for
> two Functions to point to the same DISubprogram and updates
> FunctionCloner to also clone the debug info of a function to conform
> to the new requirement. To simplify the implementation it also factors
> out the creation of inlineAt locations from the Inliner into a
> general-purpose utility in DILocation.
>
> [1] http://lists.llvm.org/pipermail/llvm-dev/2017-May/112661.html
> <rdar://problem/31926379>
>
> Differential Revision: https://reviews.llvm.org/D32975
Anna Thomas [Tue, 9 May 2017 14:29:33 +0000 (14:29 +0000)]
[LV] Fix insertion point for shuffle vectors in first order recurrence
Summary:
In first order recurrence vectorization, when the previous value is a phi node, we need to
set the insertion point to the first non-phi node.
We can have the previous value being a phi node, due to the generation of new
IVs as part of trunc optimization [1].
Guy Blank [Tue, 9 May 2017 14:03:51 +0000 (14:03 +0000)]
[X86][AVX512] Refine some avx512er intrinsics tests. NFC.
The modified tests should test the masked intrinsics.
Currently the mask is constant, which with a future patch (https://reviews.llvm.org/D32805) will cause the intrinsics to be replaced with an unmasked version.
This patch changes the constant mask to be a variable one.
Serge Pavlov [Tue, 9 May 2017 13:35:13 +0000 (13:35 +0000)]
Add extra operand to CALLSEQ_START to keep frame part set up previously
Using arguments with attribute inalloca creates problems for verification
of machine representation. This attribute instructs the backend that the
argument is prepared in stack prior to CALLSEQ_START..CALLSEQ_END
sequence (see http://llvm.org/docs/InAlloca.htm for details). Frame size
stored in CALLSEQ_START in this case does not count the size of this
argument. However CALLSEQ_END still keeps total frame size, as caller can
be responsible for cleanup of entire frame. So CALLSEQ_START and
CALLSEQ_END keep different frame size and the difference is treated by
MachineVerifier as stack error. Currently there is no way to distinguish
this case from actual errors.
This patch adds additional argument to CALLSEQ_START and its
target-specific counterparts to keep size of stack that is set up prior to
the call frame sequence. This argument allows MachineVerifier to calculate
actual frame size associated with frame setup instruction and correctly
process the case of inalloca arguments.
The changes made by the patch are:
- Frame setup instructions get the second mandatory argument. It
affects all targets that use frame pseudo instructions and touched many
files although the changes are uniform.
- Access to frame properties are implemented using special instructions
rather than calls getOperand(N).getImm(). For X86 and ARM such
replacement was made previously.
- Changes that reflect appearance of additional argument of frame setup
instruction. These involve proper instruction initialization and
methods that access instruction arguments.
- MachineVerifier retrieves frame size using method, which reports sum of
frame parts initialized inside frame instruction pair and outside it.
The patch implements approach proposed by Quentin Colombet in
https://bugs.llvm.org/show_bug.cgi?id=27481#c1.
It fixes 9 tests failed with machine verifier enabled and listed
in PR27481.
Simon Pilgrim [Tue, 9 May 2017 13:14:40 +0000 (13:14 +0000)]
[X86][SSE42] Lower v2i64/v4i64 ASHR(X, 63) as PCMPGTQ(0, X)
Similar to what we do for vXi8 ASHR(X, 7), use SSE42's PCMPGTQ to splat the sign instead of using the PSRAD+PSHUFD.
Avoiding bitcasts this improves combines that utilize computeNumSignBits, permits memory folding and reduces pipe pressure. Although it does require a second register, given that this is a (cheap) zero register the impact is minimal.
Renato Golin [Tue, 9 May 2017 12:36:50 +0000 (12:36 +0000)]
[Dwarf] Disable reference verification for now (PR32972)
There is no other explanation about why this only started happening
now, even though it crashes on old code (supposedly reachable from
here).
The only common factor between the failing bots is that they use GCC
(4.9 and 5.3) to compile Clang, while the others use Clang 3.8, but the
failure is while building the tests, as an assertion, on Clang.
Commenting it out for now in hope the bots will go back green, but we
should keep looking for the real cause, and update bugzilla.
Amara Emerson [Tue, 9 May 2017 10:43:25 +0000 (10:43 +0000)]
Introduce experimental generic intrinsics for horizontal vector reductions.
- This change allows targets to opt-in to using them instead of the log2
shufflevector algorithm.
- The SLP and Loop vectorizers have the common code to do shuffle reductions
factored out into LoopUtils, and now have a unified interface for generating
reductions regardless of the preference of the target. LoopUtils now uses TTI
to determine what kind of reductions the target wants to handle.
- For CodeGen, basic legalization support is added.
Craig Topper [Tue, 9 May 2017 07:04:04 +0000 (07:04 +0000)]
[ConstantRange] Rewrite shl to avoid repeated calls to getUnsignedMax and avoid creating the min APInt until we're sure we need it. Use inplace shift operations.
Teresa Johnson [Tue, 9 May 2017 01:43:24 +0000 (01:43 +0000)]
Fix code section prefix for proper layout
Summary:
r284533 added hot and cold section prefixes based on profile
information, to enable grouping of hot/cold functions at link time.
However, it used "cold" as the prefix for cold sections, but gold only
recognizes "unlikely" (which is used by gcc for cold sections).
Therefore, cold sections were not properly being grouped. Switch to
using "unlikely"
Even though these GEPs can be simplified to a constant offset from EBP
or RSP, we don't do that at -O0, and each GEP is computed into a
register. Registers used to compute argument addresses are typically
spilled and clobbered very quickly after the initial computation, so
live debug variable tracking loses information very quickly if we use
DBG_VALUE instructions.
This change moves processing of dbg.declare between argument lowering
and basic block isel, so that we can ask if an argument has a frame
index or not. If the argument lives in a register as is the case for
byval arguments on some targets, then we don't put it in the side table
and during ISel we emit DBG_VALUE instructions.
Adrian Prantl [Mon, 8 May 2017 21:17:08 +0000 (21:17 +0000)]
Make it illegal for two Functions to point to the same DISubprogram
As recently discussed on llvm-dev [1], this patch makes it illegal for
two Functions to point to the same DISubprogram and updates
FunctionCloner to also clone the debug info of a function to conform
to the new requirement. To simplify the implementation it also factors
out the creation of inlineAt locations from the Inliner into a
general-purpose utility in DILocation.
Sanjay Patel [Mon, 8 May 2017 20:49:59 +0000 (20:49 +0000)]
[InstCombine] add folds for not-of-shift-right
This is another step towards getting rid of dyn_castNotVal,
so we can recommit:
https://reviews.llvm.org/rL300977
As the tests show, we were missing the lshr case for constants
and both ashr/lshr vector splat folds. The ashr case with constant
was being performed inefficiently in 2 steps. It's also possible
there was a latent bug in that case because we can't do that fold
if the constant is positive:
http://rise4fun.com/Alive/Bge
Reid Kleckner [Mon, 8 May 2017 19:58:15 +0000 (19:58 +0000)]
Don't add DBG_VALUE instructions for static allocas in dbg.declare
Summary:
An llvm.dbg.declare of a static alloca is always added to the
MachineFunction dbg variable map, so these values are entirely
redundant. They survive all the way through codegen to be ignored by
DWARF emission.
Effectively revert r113967
Two bugpoint-reduced test cases from 2012 broke as a result of this
change. Despite my best efforts, I haven't been able to rewrite the test
case using dbg.value. I'm not too concerned about the lost coverage
because these were reduced from the test-suite, which we still run.
Zachary Turner [Mon, 8 May 2017 18:38:43 +0000 (18:38 +0000)]
[CodeView] Add support for random access type visitors.
Previously type visitation was done strictly sequentially, and
TypeIndexes were computed by incrementing the TypeIndex of the
last visited record. This works fine for situations like dumping,
but not when you want to visit types in random order. For example,
in a debug session someone might lookup a symbol by name, find that
it has TypeIndex 10,000 and then want to go straight to TypeIndex
10,000.
In order to make this work, the visitation framework needs a mode
where it can plumb TypeIndices through the callback pipeline. This
patch adds such a mode. In doing so, it is necessary to provide
an alternative implementation of TypeDatabase that supports random
access, so that is done as well.
Nothing actually uses these random access capabilities yet, but
this will be done in subsequent patches.
[AArch64][RegisterBankInfo] Fix mapping cost for GPR.
In r292478, we changed the order of the enum that is referenced by
PMI_FirstXXX. This had the side effect of changing the cost of the
mapping of all the loads, instead of just the FPRs ones.
Reinstate the higher cost for all but GPR loads.
Note: This did not have any external visible effects:
- For Fast mode, the cost would have been higher, but we don't care
because we don't try to use alternative mappings.
- For Greedy mode, the higher cost of the GPR loads, would have
triggered the use of the supposedly alternative mapping, that
would be in fact the same GPR mapping but with a lower cost.
projects / llvm / log