ExpandPostRAPseudos should transfer implicit uses, not only implicit defs
Previously, we would expand:
%BL<def> = COPY %DL<kill>, %EBX<imp-use,kill>, %EBX<imp-def>
Into:
%BL<def> = MOV8rr %DL<kill>, %EBX<imp-def>
Dropping the imp-use on the floor.
That confused CriticalAntiDepBreaker, which (correctly) assumes that if an
instruction defs but doesn't use a register, that register is dead immediately
before the instruction - while in this case, the high lanes of EBX can be very
much alive.
[pdb] Teach MsfBuilder and other classes about the Free Page Map.
Block 1 and 2 of an MSF file are bit vectors that represent the
list of blocks allocated and free in the file. We had been using
these blocks to write stream data and other data, so we mark them
as the free page map now. We don't yet serialize these pages to
the disk, but at least we make a note of what it is, and avoid
writing random data to them.
Doing this also necessitated cleaning up some of the tests to be
more general and hardcode fewer values, which is nice.
Previously we would read a PDB, then write some of it back out,
but write the directory, super block, and other pertinent metadata
back out unchanged. This generates incorrect PDBs since the amount
of data written was not always the same as the amount of data read.
This patch changes things to use the newly introduced `MsfBuilder`
class to write out a correct and accurate set of Msf metadata for
the data *actually* written, which opens up the door for adding and
removing type records, symbol records, and other types of data to
an existing PDB.
[Hexagon] Handle instruction latency for 0 or 2 cycles
The Hexagon schedulers need to handle instructions with a latency
of 0 or 2 more accurately. The problem, in v60, is that a dependence
between two instructions with a 2 cycle latency can use a .cur version
of the source to achieve a 0 cycle latency when the use is in the
same packet. Any othe use, must be at least 2 packets later, or a
stall occurs. In other words, the compiler does not want to schedule
the dependent instructions 1 cycle later.
To achieve this, the latency adjustment code allows only a single
dependence to have a zero latency. All other instructions have the
other value, which is typically 2 cycles. We use a heuristic to
determine which instruction gets the 0 latency.
The Hexagon machine scheduler was also changed to increase the cost
associated with 0 latency dependences than can be scheduled in the
same packet.
Remove unnecessary clutter in assembly output. When using SjLj EH, the CFI is
not actually used for anything. Do not emit the CFI needlessly. The minor test
adjustments are interesting. The prologue test was just overzealous matcching.
The interesting case is the LSDA change. It was originally added to ensure that
various compilations did not mangle the name (it explicitly checked the name!).
However, subsequent cleanups made it more reliant on the CFI to find the name.
Parse the generated code flow to generically find the label still.
Make processInstruction from LCSSA.cpp externally available.
Summary:
When a pass tries to keep LCSSA form it's often convenient to be able to update
LCSSA for a set of instructions rather than for the entire loop. This patch makes the
processInstruction from LCSSA externally available under a name
formLCSSAForInstruction.
Teach fast isel about the win64 calling convention.
This mostly just works.
Vectorcall rets are still not supported.
The win64_eh test change is because fast isel doesn't use rsi for temporary
computations, so it doesn't need to be pushed. The test case I'm changing was
originally added to test pushes, but by now there are other test cases in that
file exercising that code path.
This patch adds proper handling of stratified attributes into our
anders-style CFLAA implementation. It also comes bundled with more
CFLAnders tests. :)
This adds an incomplete anders-style implementation for CFLAA. It's
incomplete in that it's missing interprocedural analysis, attrs
handling, etc. and that it needs more tests. More tests and features
will be added in future commits.
[ReassociateGEP] Update tests to allow missing "inbounds" on certain GEPs.
With r275532 fixing miscompilation of GVN, "inbounds" on certain GEPs in these
tests cannot be preserved any more. Left a TODO in the tests for future
reference.
[SelectionDAG] Get rid of bool parameters in SelectionDAG::getLoad, getStore, and friends.
Summary:
Instead, we take a single flags arg (a bitset).
Also add a default 0 alignment, and change the order of arguments so the
alignment comes before the flags.
This greatly simplifies many callsites, and fixes a bug in
AMDGPUISelLowering, wherein the order of the args to getLoad was
inverted. It also greatly simplifies the process of adding another flag
to getLoad.
Rong Xu [Fri, 15 Jul 2016 18:10:49 +0000 (18:10 +0000)]
[PGO] IRPGO pre-cleanup pass changes
This patch adds a selected set of cleanup passes including a pre-inline pass
before LLVM IR PGO instrumentation. The inline is only intended to apply those
obvious/trivial ones before instrumentation so that much less instrumentation
is needed to get better profiling information. This will drastically improve
the instrumented code performance for large C++ applications. Another benefit
is the context sensitive counts that can potentially improve the PGO
optimization.
Adam Nemet [Fri, 15 Jul 2016 17:23:20 +0000 (17:23 +0000)]
[OptRemark,LDist] RFC: Add hotness attribute
Summary:
This is the first set of changes implementing the RFC from
http://thread.gmane.org/gmane.comp.compilers.llvm.devel/98334
This is a cross-sectional patch; rather than implementing the hotness
attribute for all optimization remarks and all passes in a patch set, it
implements it for the 'missed-optimization' remark for Loop
Distribution. My goal is to shake out the design issues before scaling
it up to other types and passes.
Hotness is computed as an integer as the multiplication of the block
frequency with the function entry count. It's only printed in opt
currently since clang prints the diagnostic fields directly. E.g.:
remark: /tmp/t.c:3:3: loop not distributed: use -Rpass-analysis=loop-distribute for more info (hotness: 300)
A new API added is similar to emitOptimizationRemarkMissed. The
difference is that it additionally takes a code region that the
diagnostic corresponds to. From this, hotness is computed using BFI.
The new API is exposed via an analysis pass so that it can be made
dependent on LazyBFI. (Thanks to Hal for the analysis pass idea.)
This feature can all be enabled by setDiagnosticHotnessRequested in the
LLVM context. If this is off, LazyBFI is not calculated (D22141) so
there should be no overhead.
A new command-line option is added to turn this on in opt.
My plan is to switch all user of emitOptimizationRemark* to use this
module instead.
David Majnemer [Fri, 15 Jul 2016 17:19:24 +0000 (17:19 +0000)]
[AliasAnalysis] Give back AA results for fence instructions
Calling getModRefInfo with a fence resulted in crashes because fences
don't have a memory location. Add a new predicate to Instruction
called isFenceLike which indicates that the instruction mutates memory
but not any single memory location in particular. In practice, it is a
proxy for the set of instructions which "mayWriteToMemory" but cannot be
used with MemoryLocation::get.
IR: Sort generic intrinsics before target specific ones
This splits out the intrinsic table such that generic intrinsics come
first and target specific intrinsics are grouped by target. From here
we can find out which target an intrinsic is for or differentiate
between generic and target intrinsics.
The motivation here is to make it easier to move target specific
intrinsic handling out of generic code.
[DSE]Enhance shorthening MemIntrinsic based on OverlapIntervals
Summary:
This change use the overlap interval map built from partial overwrite tracking to perform shortening MemIntrinsics.
Add test cases which was missing opportunities before.
[Hexagon] Improve patterns with stack-based addressing
- Treat bitwise OR with a frame index as an ADD wherever possible, fold it
into addressing mode.
- Extend patterns for memops to allow memops with frame indexes as address
operands.
In dag-optnone.ll, use varargs instead of win64 to fast SDIsel.
The test used to rely on targeting win64 to disable fast isel,
but I'd like to teach fast isel about win64 rets. Change the
test to use varargs to disable fast isel.
Matthew Simpson [Fri, 15 Jul 2016 15:22:43 +0000 (15:22 +0000)]
[LV] Swap A and B in interleaved access analysis (NFC)
This patch swaps A and B in the interleaved access analysis and clarifies
related comments. The algorithm is more intuitive if we let access A precede
access B in program order rather than the reverse. This change was requested in
the review of D19984.
Yaxun Liu [Fri, 15 Jul 2016 14:58:21 +0000 (14:58 +0000)]
[AMDGPU] Add metadata for runtime
Added emitting metadata to elf for runtime.
Runtime requires certain information (metadata) about kernels to be able to execute and query them. Such information is emitted to an elf section as a key-value pair stream.
Jacques Pienaar [Fri, 15 Jul 2016 14:41:04 +0000 (14:41 +0000)]
Rename AnalyzeBranch* to analyzeBranch*.
Summary: NFC. Rename AnalyzeBranch/AnalyzeBranchPredicate to analyzeBranch/analyzeBranchPredicate to follow LLVM coding style and be consistent with TargetInstrInfo's analyzeCompare and analyzeSelect.
Sebastian Pop [Fri, 15 Jul 2016 13:45:20 +0000 (13:45 +0000)]
code hoisting pass based on GVN
This pass hoists duplicated computations in the program. The primary goal of
gvn-hoist is to reduce the size of functions before inline heuristics to reduce
the total cost of function inlining.
Pass written by Sebastian Pop, Aditya Kumar, Xiaoyu Hu, and Brian Rzycki.
Important algorithmic contributions by Daniel Berlin under the form of reviews.
Simon Pilgrim [Fri, 15 Jul 2016 09:49:12 +0000 (09:49 +0000)]
[X86][AVX2] Improve lowerShuffleAsRepeatedMaskAndLanePermute permutation of 64-bit sub-lanes
As discussed on PR28136, lowerShuffleAsRepeatedMaskAndLanePermute was attempting to match repeated masks at the 128-bit level and then permute the resultant lanes at the 128-bit (AVX1) or 64-bit (AVX2) sub-lane level.
This change allows us to create the repeated masks at the sub-lane level (and then concat them together to create a 128-bit repeated mask) and then select which sub-lane to permute. This has no effect on the AVX1 codegen.
James Molloy [Fri, 15 Jul 2016 08:03:56 +0000 (08:03 +0000)]
[Thumb-1] Select post-increment load and store where possible
Thumb-1 doesn't have post-inc or pre-inc load or store instructions. However the LDM/STM instructions with writeback can function as post-inc load/store:
ldm r0!, {r1} @ load from r0 into r1 and increment r0 by 4
Obviously, this only works if the post increment is 4.
James Molloy [Fri, 15 Jul 2016 07:55:21 +0000 (07:55 +0000)]
[ARM] Prefer indirect calls in minsize mode
... When we emit several calls to the same function in the same basic block.
An indirect call uses a "BLX r0" instruction which has a 16-bit encoding. If many calls are made to the same target, this can enable significant code size reductions.
Taking a lock before appending to a vector does no good unless threads
reading from the vector also take the lock, because the vector could be
re-sized.
I don't have a good isolated test for this. I found the issue with ASan
while testing a large project. I'm working on a bot that does this.
[llvm-cov] Clean up an awkward capture-by-reference (NFC)
Writing `for (StringRef &SourceFile : ...)` is strange to begin with.
Subsequently capturing "SourceFile" by reference is even stranger. Just
copy the StringRef, since that's cheap to do.
Matt Arsenault [Fri, 15 Jul 2016 00:58:15 +0000 (00:58 +0000)]
AMDGPU: Fix not expanding control flow after some kill blocks
Also stop trying to insert skip blocks at end_cf. This
was inserting them at the end of the block which doesn't make
sense. The skip should be inserted at the beginning of the block
right after the end cf. Just remove this for now since no tests
seem to stress this and I think this can be handled more generally
later.
[codeview] Shrink inlined call site line info tables
For a fully inlined call chain like a -> b -> c -> d, we were emitting
line info for 'd' 3 separate times: once for d's actual InlineSite line
table, and twice for 'b' and 'c'. This is particularly inefficient when
all these functions are in different headers, because now we need to
encode the file change. Windbg was coping with our suboptimal output, so
this should not be noticeable from the debugger.
Tim Northover [Thu, 14 Jul 2016 23:13:03 +0000 (23:13 +0000)]
llvm-objdump: extend __mh_execute_header handling to other special syms
We don't need to print any of the special __mh_*_header symbols when
disassembling. Since they point at the beginning of the segment (not where the
actual code is) they're pretty misleading.
Tim Northover [Thu, 14 Jul 2016 22:13:32 +0000 (22:13 +0000)]
llvm-objdump: handle stubbed and malformed dylibs better
We were quite happy to read past the end of the valid section data when
disassembling. Instead we entirely skip stub dylibs, and tell the user what's
happened if their section only has partial data.
Matthew Simpson [Thu, 14 Jul 2016 21:05:08 +0000 (21:05 +0000)]
[LV] Rename StrideAccesses to AccessStrideInfo (NFC)
We now collect all accesses with a constant stride, not just the ones with a
stride greater than one. This change was requested in the review of D19984.
Matthew Simpson [Thu, 14 Jul 2016 20:59:47 +0000 (20:59 +0000)]
[LV] Allow interleaved accesses in loops with predicated blocks
This patch allows the formation of interleaved access groups in loops
containing predicated blocks. However, the predicated accesses are prevented
from forming groups.
[Hexagon] Packetize function call arguments with tail call instructions
On Hexagon is it legal to packetize the instructions setting up call
arguments with the call instruction itself. This was already done,
except for tail calls. Make sure tail calls are handled as well.
Summary:
Make the target-specific flags in MachineMemOperand::Flags real, bona
fide enum values. This simplifies users, prevents various constants
from going out of sync, and avoids the false sense of security provided
by declaring static members in classes and then forgetting to define
them inside of cpp files.
projects / llvm / log