Chandler Carruth [Tue, 20 Dec 2016 03:32:17 +0000 (03:32 +0000)]
[PM] Rework a loop in the CGSCC update logic to be more conservative and
clear. The current RefSCC can occur in exactly one position so we should
just enforce that and leverage the property rather than checking for it
anywhere.
This addresses review comments made on another patch.
Chandler Carruth [Tue, 20 Dec 2016 03:15:32 +0000 (03:15 +0000)]
[PM] Provide an initial, minimal port of the inliner to the new pass manager.
This doesn't implement *every* feature of the existing inliner, but
tries to implement the most important ones for building a functional
optimization pipeline and beginning to sort out bugs, regressions, and
other problems.
Notable, but intentional omissions:
- No alloca merging support. Why? Because it isn't clear we want to do
this at all. Active discussion and investigation is going on to remove
it, so for simplicity I omitted it.
- No support for trying to iterate on "internally" devirtualized calls.
Why? Because it adds what I suspect is inappropriate coupling for
little or no benefit. We will have an outer iteration system that
tracks devirtualization including that from function passes and
iterates already. We should improve that rather than approximate it
here.
- Optimization remarks. Why? Purely to make the patch smaller, no other
reason at all.
The last one I'll probably work on almost immediately. But I wanted to
skip it in the initial patch to try to focus the change as much as
possible as there is already a lot of code moving around and both of
these *could* be skipped without really disrupting the core logic.
A summary of the different things happening here:
1) Adding the usual new PM class and rigging.
2) Fixing minor underlying assumptions in the inline cost analysis or
inline logic that don't generally hold in the new PM world.
3) Adding the core pass logic which is in essence a loop over the calls
in the nodes in the call graph. This is a bit duplicated from the old
inliner, but only a handful of lines could realistically be shared.
(I tried at first, and it really didn't help anything.) All told,
this is only about 100 lines of code, and most of that is the
mechanics of wiring up analyses from the new PM world.
4) Updating the LazyCallGraph (in the new PM) based on the *newly
inlined* calls and references. This is very minimal because we cannot
form cycles.
5) When inlining removes the last use of a function, eagerly nuking the
body of the function so that any "one use remaining" inline cost
heuristics are immediately refined, and queuing these functions to be
completely deleted once inlining is complete and the call graph
updated to reflect that they have become dead.
6) After all the inlining for a particular function, updating the
LazyCallGraph and the CGSCC pass manager to reflect the
function-local simplifications that are done immediately and
internally by the inline utilties. These are the exact same
fundamental set of CG updates done by arbitrary function passes.
7) Adding a bunch of test cases to specifically target CGSCC and other
subtle aspects in the new PM world.
Many thanks to the careful review from Easwaran and Sanjoy and others!
Adrian Prantl [Tue, 20 Dec 2016 02:09:43 +0000 (02:09 +0000)]
[IR] Remove the DIExpression field from DIGlobalVariable.
This patch implements PR31013 by introducing a
DIGlobalVariableExpression that holds a pair of DIGlobalVariable and
DIExpression.
Currently, DIGlobalVariables holds a DIExpression. This is not the
best way to model this:
(1) The DIGlobalVariable should describe the source level variable,
not how to get to its location.
(2) It makes it unsafe/hard to update the expressions when we call
replaceExpression on the DIGLobalVariable.
(3) It makes it impossible to represent a global variable that is in
more than one location (e.g., a variable with multiple
DW_OP_LLVM_fragment-s). We also moved away from attaching the
DIExpression to DILocalVariable for the same reasons.
This reapplies r289902 with additional testcase upgrades and a change
to the Bitcode record for DIGlobalVariable, that makes upgrading the
old format unambiguous also for variables without DIExpressions.
Chris Bieneman [Tue, 20 Dec 2016 00:42:06 +0000 (00:42 +0000)]
Revert "[ObjectYAML] Support for DWARF debug_info section"
This reverts commit r290147.
This commit is breaking a bot (http://lab.llvm.org:8011/builders/clang-atom-d525-fedora-rel/builds/621). I don't have time to investigate at the moment, so I'll revert for now.
Michael LeMay [Mon, 19 Dec 2016 21:02:41 +0000 (21:02 +0000)]
[TargetInstrInfo] replace redundant expression in getMemOpBaseRegImmOfs
Summary:
The expression for computing the return value of getMemOpBaseRegImmOfs has only
one possible value. The other value would result in a return earlier in the
function. This patch replaces the expression with its only possible value.
Greg Clayton [Mon, 19 Dec 2016 20:36:41 +0000 (20:36 +0000)]
Make a function to correctly extract the DW_AT_high_pc given the low pc value.
DWARF 4 and later supports encoding the PC as an address or as as offset from the low PC. Clients using DWARFDie should be insulated from how to extract the high PC value. This function takes care of extracting the form value and looking for the correct form.
Diana Picus [Mon, 19 Dec 2016 14:08:02 +0000 (14:08 +0000)]
[ARM] GlobalISel: Lower i8 and i16 register args
This allows lowering i8 and i16 arguments if they can fit in the registers. Note
that the lowering is incomplete - ABI extensions are handled in a subsequent
patch.
(Last part of)
Differential Revision: https://reviews.llvm.org/D27704
Diana Picus [Mon, 19 Dec 2016 11:55:41 +0000 (11:55 +0000)]
[ARM] GlobalISel: Lower more than 4 arguments
This adds support for lowering more than 4 arguments (although still i32 only).
It uses the handleAssignments / ValueHandler infrastructure extracted from
the AArch64 backend in r288658.
Sam Kolton [Mon, 19 Dec 2016 11:43:15 +0000 (11:43 +0000)]
AMDGPU: [AMDGPU] Assembler: add .hsa_code_object_metadata directive for functime metadata V2.0
Summary:
Added pair of directives .hsa_code_object_metadata/.end_hsa_code_object_metadata.
Between them user can put YAML string that would be directly put to the generated note. E.g.:
'''
.hsa_code_object_metadata
{
amd.MDVersion: [ 2, 0 ]
}
.end_hsa_code_object_metadata
'''
Based on D25046
Diana Picus [Mon, 19 Dec 2016 11:26:31 +0000 (11:26 +0000)]
[ARM] GlobalISel: Support loading from the stack
Add support for selecting simple G_LOAD and G_FRAME_INDEX instructions (32-bit
scalars only). This will be useful for functions that need to pass arguments on
the stack.
Bjorn Pettersson [Mon, 19 Dec 2016 11:20:57 +0000 (11:20 +0000)]
[CodeGen] Make MachineInstr::isIdenticalTo() symmetric.
Summary:
MachineInstr::isIdenticalTo() is for some reason not
symmetric when comparing bundles, which gives us the
property:
I1->isIdenticalTo(*I2) != I2->isIdenticalTo(*I1)
when comparing bundles where one bundle is longer than
the other.
This patch makes sure that bundles of different length
always are considered as not being identical. Thus, the
result of the comparison will be the same regardless of
which side that happens to be to the left.
The original version of the code in XRayInstrumentation.cpp assumed that
functions may not have empty machine basic blocks (or that the first one
couldn't be). This change addresses that by special-casing that specific
situation.
We provide two .mir test-cases to make sure we're handling this
appropriately.
Craig Topper [Mon, 19 Dec 2016 08:35:56 +0000 (08:35 +0000)]
[X86] When recognizing vector loads or VZEXT_LOAD in selectScalarSSELoad make sure we pass the load's user rather than load itself to the second operand of IsLegalToFold.
Craig Topper [Mon, 19 Dec 2016 00:42:28 +0000 (00:42 +0000)]
[X86] Remove all of the patterns that use X86ISD:FAND/FXOR/FOR/FANDN except for the ones needed for SSE1. Anything SSE2 or above uses the integer ISD opcode.
This removes 11721 bytes from the DAG isel table or 2.2%
David Majnemer [Sun, 18 Dec 2016 20:10:50 +0000 (20:10 +0000)]
[PDB] Don't use the long type
Long is not the same size across a number of the platforms we support.
Use unsigned int here instead, it is more appropriate because
overflow/wrap-around is possible and, in this case, expected.
Sanjay Patel [Sun, 18 Dec 2016 18:49:48 +0000 (18:49 +0000)]
[InstCombine] use commutative matchers for patterns with commutative operators
Background/motivation - I was circling back around to:
https://llvm.org/bugs/show_bug.cgi?id=28296
I made a simple patch for that and noticed some regressions, so added test cases for
those with rL281055, and this is hopefully the minimal fix for just those cases.
But as you can see from the surrounding untouched folds, we are missing commuted patterns
all over the place, and of course there are no regression tests to cover any of those cases.
We could sprinkle "m_c_" dust all over this file and catch most of the missing folds, but
then we still wouldn't have test coverage, and we'd still miss some fraction of commuted
patterns because they require adjustments to the match order.
I'm aware of the concern about the potential compile-time performance impact of adding
matches like this (currently being discussed on llvm-dev), but I don't think there's any
evidence yet to suggest that handling commutative pattern matching more thoroughly is not
a worthwhile goal of InstCombine.
Daniel Jasper [Sun, 18 Dec 2016 14:36:38 +0000 (14:36 +0000)]
Revert r289955 and r289962. This is causing lots of ASAN failures for us.
Not sure whether it causes and ASAN false positive or whether it
actually leads to incorrect code or whether it even exposes bad code.
Hans, I'll get you instructions to reproduce this.
[X86] [AVX512] Minor fix in encoding of scalar EVEX instructions. NFC.
Commit on behalf of Gadi Haber
Removed EVEX_V512 prefix from scalar EVEX instructions since HW ignores L'L bits anyway (LIG). 4 instructions are modified.
The changed encodings are validated with XED.
Rviewers: delena, igorb
Tom de Vries [Sun, 18 Dec 2016 09:41:20 +0000 (09:41 +0000)]
[FileCheck] Fix comment in ReadCheckFile
The comment in ReadCheckFile claims that both leading and trailing whitespace
are removed, but the associated statement only removes leading whitespace.
Craig Topper [Sun, 18 Dec 2016 07:54:23 +0000 (07:54 +0000)]
[X86][SSE][AVX-512] Convert FAND/FOR/FXOR/FANDN nodes to integer operations if they are available. This will allow a bunch of patterns to be removed.
These nodes are only emitted for lowering FABS/FNEG/FNABS/FCOPYSIGN. Ideally we just wouldn't create these nodes if SSE2 or higher is available, but it was simple to just convert them in DAG combine.
For SSE2, AVX, and AVX512 with DQI this is no functional change as the execution domain fixing pass ensures the right domain is selected regardless of the ISD opcode.
For AVX-512 without DQI we end up using integer instructions since the floating point versions aren't available. But we were already doing that for any logical operations in code that didn't come from FABS/FNEG/FNABS/FCOPYSIGN so this seems no worse. And we get the benefit of being able to fold broadcasts now.
Daniel Jasper [Sat, 17 Dec 2016 12:27:49 +0000 (12:27 +0000)]
Revert "[libFuzzer] add an experimental flag -experimental_len_control=1 that sets max_len to 1M and tries to increases the actual max sizes of mutations very gradually. Also remove a bit of dead code"
George Rimar [Sat, 17 Dec 2016 09:10:32 +0000 (09:10 +0000)]
[DWARF] - Introduce DWARFDebugPubTable class for dumping pub* sections.
Patch implements parser of pubnames/pubtypes tables instead of static
function used before. It is now should be possible to reuse it
in LLD or other projects and clean up the duplication code.
Tom Stellard [Sat, 17 Dec 2016 04:41:53 +0000 (04:41 +0000)]
Add custom type for PseudoSourceValue
Summary:
PseudoSourceValue can be used to attach a target specific value for "well behaved" side-effects lowered from target specific intrinsics.
This is useful whenever there is not an LLVM IR Value around when representing such "well behaved" side-effected operations in backends by attaching a MachineMemOperand with a custom PseudoSourceValue as this makes the scheduler not treating them as "GlobalMemoryObjects" which triggers a logic that makes the operation act like a barrier in the Schedule DAG.
This patch adds another Kind to the PseudoSourceValue object which is "TargetCustom". It indicates a type of PseudoSourceValue that has a target specific meaning (aka. LLVM shouldn't assume any specific usage for such a PSV).
It supports the possibility of having many different kinds of "TargetCustom" PseudoSourceValues.
We had a discussion about if this was valuable or not (in particular because there was a believe that PSV were going away sooner or later) but seems like they are not going anywhere and I think they are useful backend side.
It is not clear the interaction of this with MIRParser (do we need a target hook to parse these?) and I would like a comment from Alex about that :)
[libFuzzer] when tracing switch statements, handle only one case at a time (to make things faster). Also ensure that the signals from value profile do not intersect with the regular coverage
Evgeniy Stepanov [Sat, 17 Dec 2016 01:31:46 +0000 (01:31 +0000)]
Fix compilation.
unittests/ADT/TwineTest.cpp:106:38: error: field 'Count' will be initialized after base 'llvm::FormatAdapter<int>' [-Werror,-Wreorder]
explicit formatter(int &Count) : Count(Count), FormatAdapter(0) {}
Matthias Braun [Sat, 17 Dec 2016 01:15:59 +0000 (01:15 +0000)]
AArch64CollectLOH: Rewrite as block-local analysis.
Re-apply r288561: Liveness tracking should be correct now after r290014.
Previously this pass was using up to 5% compile time in some cases which
is a bit much for what it is doing. The pass featured a full blown
data-flow analysis which in the default configuration was restricted to a
single block.
This rewrites the pass under the assumption that we only ever work on a
single block. This is done in a single pass maintaining a state machine
per general purpose register to catch LOH patterns.
Vedant Kumar [Sat, 17 Dec 2016 01:02:08 +0000 (01:02 +0000)]
Retry: [BPI] Use a safer constructor to calculate branch probabilities
BPI may trigger signed overflow UB while computing branch probabilities for
cold calls or to unreachables. For example, with our current choice of weights,
we'll crash if there are >= 2^12 branches to an unreachable.
Use a safer BranchProbability constructor which is better at handling fractions
with large denominators.
Changes since the initial commit:
- Use explicit casts to ensure that multiplication operands are 64-bit
ints.
Vedant Kumar [Sat, 17 Dec 2016 00:09:51 +0000 (00:09 +0000)]
[BPI] Use a safer constructor to calculate branch probabilities
BPI may trigger signed overflow UB while computing branch probabilities
for cold calls or to unreachables. For example, with our current choice
of weights, we'll crash if there are >= 2^12 branches to an unreachable.
Use a safer BranchProbability constructor which is better at handling
fractions with large denominators.
David Blaikie [Fri, 16 Dec 2016 23:53:14 +0000 (23:53 +0000)]
GDB pretty printers: Basic DenseMap support
Still prints the empty/tombstone keys (which some people would prefer,
but I find pretty noisy) because I haven't yet found a reliable way to
skip them (it requires calling into the running process to do so, which
isn't ideal for a pretty printer (doesn't work on a core file, for
example) - and gdb's ability to do so (or my ability to figure out how
to get gdb to do so) is limited) left some breadcrumbs for the next
person who might try to address that.
Kevin Enderby [Fri, 16 Dec 2016 22:54:02 +0000 (22:54 +0000)]
Fix a bugs with using some Mach-O command line flags like "-arch armv7m".
The Mach-O command line flag like "-arch armv7m" does not match the
arch name part of its llvm Triple which is "thumbv7m-apple-darwin”.
I think the best way to fix this is to have
llvm::object::MachOObjectFile::getArchTriple() optionally return the
name of the Mach-O arch flag that would be used with -arch that
matches the CPUType and CPUSubType. Then change
llvm::object::MachOUniversalBinary::ObjectForArch::getArchTypeName()
to use that and change it to getArchFlagName() as the type name is
really part of the Triple and the -arch flag name is a Mach-O thing
for a specific Triple with a specific Mcpu value.
[libFuzzer] add an experimental flag -experimental_len_control=1 that sets max_len to 1M and tries to increases the actual max sizes of mutations very gradually. Also remove a bit of dead code
Teresa Johnson [Fri, 16 Dec 2016 21:25:01 +0000 (21:25 +0000)]
[ThinLTO] Import composite types as declarations
Summary:
When reading the metadata bitcode, create a type declaration when
possible for composite types when we are importing. Doing this in
the bitcode reader saves memory. Also it works naturally in the case
when the type ODR map contains a definition for the same composite type
because it was used in the importing module (buildODRType will
automatically use the existing definition and not create a type
declaration).
For Chromium built with -g2, this reduces the aggregate size of the
generated native object files by 66% (from 31G to 10G). It reduced
the time through the ThinLTO link and backend phases by about 20% on
my machine.
Jun Bum Lim [Fri, 16 Dec 2016 20:38:39 +0000 (20:38 +0000)]
[CodeGenPrep] Skip merging empty case blocks
This is recommit of r287553 after fixing the invalid loop info after eliminating an empty block and unit test failures in AVR and WebAssembly :
Summary: Merging an empty case block into the header block of switch could cause ISel to add COPY instructions in the header of switch, instead of the case block, if the case block is used as an incoming block of a PHI. This could potentially increase dynamic instructions, especially when the switch is in a loop. I added a test case which was reduced from the benchmark I was targetting.
Adrian Prantl [Fri, 16 Dec 2016 19:39:01 +0000 (19:39 +0000)]
Revert "[IR] Remove the DIExpression field from DIGlobalVariable."
This reverts commit 289920 (again).
I forgot to implement a Bitcode upgrade for the case where a DIGlobalVariable
has not DIExpression. Unfortunately it is not possible to safely upgrade
these variables without adding a flag to the bitcode record indicating which
version they are.
My plan of record is to roll the planned follow-up patch that adds a
unit: field to DIGlobalVariable into this patch before recomitting.
This way we only need one Bitcode upgrade for both changes (with a
version flag in the bitcode record to safely distinguish the record
formats).
David Blaikie [Fri, 16 Dec 2016 19:16:22 +0000 (19:16 +0000)]
GDB pretty printer for llvm::Optional
(some other implementations of an optional pretty printer print the full
name of the optional type (including template parameter) - but seems if
the template parameter isn't printed for std::vector, not sure why it
would be printed for optional, so erring on the side of consistency in
that direction here - compact, etc, as well)
Matthew Simpson [Fri, 16 Dec 2016 19:12:02 +0000 (19:12 +0000)]
Reapply "[LV] Enable vectorization of loops with conditional stores by default"
This patch reapplies r289863. The original patch was reverted because it
exposed a bug causing the loop vectorizer to crash in the Python runtime on
PPC. The underlying issue was fixed with r289958.
Sanjoy Das [Fri, 16 Dec 2016 18:52:33 +0000 (18:52 +0000)]
Fix CodeGenPrepare::stripInvariantGroupMetadata
`dropUnknownNonDebugMetadata` takes a list of "known" metadata IDs. The
only reason it worked at all is that `getMetadataID` returns something
unrelated -- it returns the subclass ID of the receiver (which is used
in `dyn_cast` etc.). That does not numerically match
`LLVMContext::MD_invariant_group` and ends up dropping `invariant_group`
along with every other metadata that does not numerically match
`LLVMContext::MD_invariant_group`.
Eli Friedman [Fri, 16 Dec 2016 18:44:08 +0000 (18:44 +0000)]
[ARM] Add ARMISD::VLD1DUP to match vld1_dup more consistently.
Currently, there are substantial problems forming vld1_dup even if the
VDUP survives legalization. The lack of an actual node
leads to terrible results: not only can we not form post-increment vld1_dup
instructions, but we form scalar pre-increment and post-increment
loads which force the loaded value into a GPR. This patch fixes that
by combining the vdup+load into an ARMISD node before DAGCombine
messes it up.
Also includes a crash fix for vld2_dup (see testcase @vld2dupi8_postinc_variable).
Recommiting with fix to avoid forming vld1dup if the type of the load
doesn't match the type of the vdup (see
https://llvm.org/bugs/show_bug.cgi?id=31404).
projects / llvm / log