Balaram Makam [Wed, 4 Oct 2017 16:45:24 +0000 (16:45 +0000)]
"[ARM] Mark flaky test MachineBranchProb.ll unsupported again for ARM/AArch64"
r314857 changed the CFG that resulted in the flaky test MachineBranchProb.ll to
fail the bots again. Marking it as unsupported for ARM/AArch64 again until we
find the cause.
Adam Nemet [Wed, 4 Oct 2017 15:18:11 +0000 (15:18 +0000)]
[OptRemark] Move YAML writing to IR
Before the patch this was in Analysis. Moving it to IR and making it implicit
part of LLVMContext::diagnose allows the full opt-remark facility to be used
outside passes e.g. the pass manager. Jessica is planning to use this to
report function size after each pass. The same could be used for time
reports.
Dylan McKay [Wed, 4 Oct 2017 09:51:28 +0000 (09:51 +0000)]
[AVR] Insert JMP for long branches
Previously, on long branches (relative jumps of >4 kB), an assertion
failure was hit, as AVRInstrInfo::insertIndirectBranch was not
implemented. Despite its name, it is called by the branch relaxator
for *all* unconditional jumps.
Dylan McKay [Wed, 4 Oct 2017 09:51:21 +0000 (09:51 +0000)]
[AVR] Fix displacement overflow for LDDW/STDW
In some cases, the code generator attempts to generate instructions such as:
lddw r24, Y+63
which expands to:
ldd r24, Y+63
ldd r25, Y+64 # Oops! This is actually ld r25, Y in the binary
This commit limits the first offset to 62, and thus the second to 63.
It also updates some asserts in AVRExpandPseudoInsts.cpp, including for
INW and OUTW, which appear to be unused.
Oliver Stannard [Wed, 4 Oct 2017 09:18:07 +0000 (09:18 +0000)]
[ARM, Asm] Change grammar of immediate operand diagnostics
Currently, our diagnostics for assembly operands are not consistent.
Some start with (for example) "immediate operand must be ...",
and some with "operand must be an immediate ...". I think the latter
form is preferable for a few reasons:
* It's unambiguous that it is referring to the expected type of operand, not
the type the user provided. For example, the user could provide an register
operand, and get a message taking about an operand is if it is already an
immediate, just not in the accepted range.
* It allows us to have a consistent style once we add diagnostics for operands
that could take two forms, for example a label or pc-relative memory operand.
Jatin Bhateja [Wed, 4 Oct 2017 09:02:10 +0000 (09:02 +0000)]
[X86] Improvement in CodeGen instruction selection for LEAs (re-applying post required revision changes.)
Summary:
1/ Operand folding during complex pattern matching for LEAs has been
extended, such that it promotes Scale to accommodate similar operand
appearing in the DAG.
e.g.
T1 = A + B
T2 = T1 + 10
T3 = T2 + A
For above DAG rooted at T3, X86AddressMode will no look like
Base = B , Index = A , Scale = 2 , Disp = 10
2/ During OptimizeLEAPass down the pipeline factorization is now performed over LEAs
so that if there is an opportunity then complex LEAs (having 3 operands)
could be factored out.
e.g.
leal 1(%rax,%rcx,1), %rdx
leal 1(%rax,%rcx,2), %rcx
will be factored as following
leal 1(%rax,%rcx,1), %rdx
leal (%rdx,%rcx) , %edx
3/ Aggressive operand folding for AM based selection for LEAs is sensitive to loops,
thus avoiding creation of any complex LEAs within a loop.
George Rimar [Wed, 4 Oct 2017 08:50:08 +0000 (08:50 +0000)]
[MC] - Don't assert when non-english characters are used.
I found that llvm-mc does not like non-english characters even in comments,
which it tries to tokenize.
Problem happens because of functions like isdigit(), isalnum() which takes
int argument and expects it is not negative.
But at the same time MCParser uses char* to store input buffer poiner, char has signed value,
so it is possible to pass negative value to one of functions from above and
that triggers an assert.
Testcase for demonstration is provided.
To fix the issue helper functions were introduced in StringExtras.h
then we can't just change %vreg0 into %vreg3, since %vreg2 is actually
undef. We would have to also copy the undef flag to be able to change the
register.
Instead we deal with this case like other cases where we can't just
replace the register: we insert a COPY. The code creating the COPY already
copied all flags from the PHI input, so the undef flag will be transferred
as it should.
Max Kazantsev [Wed, 4 Oct 2017 06:53:22 +0000 (06:53 +0000)]
[IRCE] Temporarily disable unsigned latch conditions by default
We have found some corner cases connected to range intersection where IRCE makes
a bad thing when the latch condition is unsigned. The fix for that will go as a follow up.
This patch temporarily disables IRCE for unsigned latch conditions until the issue is fixed.
The unsigned latch conditions were introduced to IRCE by rL310027.
then we can't just change %vreg0 into %vreg3, since %vreg2 is actually
undef. We would have to also copy the undef flag to be able to change the
register.
Instead we deal with this case like other cases where we can't just
replace the register: we insert a COPY. The code creating the COPY already
copied all flags from the PHI input, so the undef flag will be transferred
as it should.
Martin Storsjo [Wed, 4 Oct 2017 05:12:10 +0000 (05:12 +0000)]
[X86] Fix using the SJLJ jump table on x86_64
The previous version didn't work if the jump table base address didn't
fit in 32 bit, since it was encoded as an immediate offset. And in case
the jump table is encoded as 32 bit label differences, we need to
load and add them to the table base first.
This solves the first half of the issues mentioned in PR34720.
Also fix some of the errors pointed out by -verify-machineinstrs, by
using GR32_NOSPRegClass.
Adam Nemet [Wed, 4 Oct 2017 04:26:23 +0000 (04:26 +0000)]
Move verbosity check for remarks to the diag handler
Test needs some slight adjustment because we no longer check the existence of
BFI but rather that the actual hotness is set on the remark. If entry_count
is not set getBlockProfileCount returns None.
Balaram Makam [Tue, 3 Oct 2017 22:39:24 +0000 (22:39 +0000)]
[AArch64] Use LateSimplifyCFG after expanding atomic operations.
Summary:
After r308422 we defer optimizations that can destroy loop canonical forms to
LateSimplifyCFG. Running LateSimplifyCFG after expanding atomic operations
can exploit more control-flow opportunities.
Jakub Kuderski [Tue, 3 Oct 2017 21:17:48 +0000 (21:17 +0000)]
[Dominators] Make eraseNode invalidate DFS numbers
This patch makes DT::eraseNode mark DFSInfo as invalid.
Not marking it as invalid leads to DFS numbers getting corrupted
and failing VerifyDFSNumbers check.
This patch also makes children iterator const (NFC).
Reid Kleckner [Tue, 3 Oct 2017 21:12:18 +0000 (21:12 +0000)]
[X86] Remove dead declaration convertArgMovsToPushes, NFC
This was dead when it landed in r252578. We have this functionality, if
not for stack probe calls, but for regular calls in
X86CallFrameOptimization.cpp.
The symbol table refers to offsets in the tail. A complication is that
we would like to support symbol tables that use 64 bit offsets if it
turns out that any of the offsets is too big.
This patch changes the archive writer to first compute the tail. We
cannot just compute one big StringRef since that would require reading
every member upfront, but we can represent it as a series of
StringRefs.
Having done that it is much easier to compute the symbol table and all
offsets are computed before it is written. With this if there is an
accounting problem it will show up with a regular symbol table, not
just when a 64 bit one is needed.
This commit does two things. Firstly, it cleans up some of the benefit
calculation wrt outlined functions and candidates. Secondly, it fixes an
off-by-one bug in the cost model which was caused by the benefit value of
an OutlinedFunction and Candidate differing by 1. It updates the remarks test
to reflect this change.
Tim Renouf [Tue, 3 Oct 2017 19:03:52 +0000 (19:03 +0000)]
[AMDGPU] implemented pal metadata
Summary:
For the amdpal OS type:
We write an AMDGPU_PAL_METADATA record in the .note section in the ELF
(or as an assembler directive). It contains key=value pairs of 32 bit
ints. It is a merge of metadata from codegen of the shaders, and
metadata provided by the frontend as _amdgpu_pal_metadata IR metadata.
Where both sources have a key=value with the same key, the two values
are ORed together.
This .note record is part of the amdpal ABI and will be documented in
docs/AMDGPUUsage.rst in a future commit.
Eventually the amdpal OS type will stop generating the .AMDGPU.config
section once the frontend has safely moved over to using the .note
records above instead of .AMDGPU.config.
Hans Wennborg [Tue, 3 Oct 2017 18:39:13 +0000 (18:39 +0000)]
Revert r314817 "[dwarfdump] Add -lookup option"
The test fails on Linux; see follow-up email on the llvm-commits list.
> Add the option to lookup an address in the debug information and print
> out the file, function, block and line table details.
>
> Differential revision: https://reviews.llvm.org/D38409
All the buildbots are red, e.g.
http://lab.llvm.org:8011/builders/clang-cmake-aarch64-lld/builds/2436/
> Summary:
> This patch tries to vectorize loads of consecutive memory accesses, accessed
> in non-consecutive or jumbled way. An earlier attempt was made with patch D26905
> which was reverted back due to some basic issue with representing the 'use mask' of
> jumbled accesses.
>
> This patch fixes the mask representation by recording the 'use mask' in the usertree entry.
>
> Change-Id: I9fe7f5045f065d84c126fa307ef6ebe0787296df
>
> Reviewers: mkuper, loladiro, Ayal, zvi, danielcdh
>
> Reviewed By: Ayal
>
> Subscribers: hans, mzolotukhin
>
> Differential Revision: https://reviews.llvm.org/D36130
Hans Wennborg [Tue, 3 Oct 2017 18:27:22 +0000 (18:27 +0000)]
CodeView: Provide a .def file with the register ids
The list of register ids was previously written out in a couple of dirrent
places. This puts it in a .def file and also adds a few more registers (e.g.
the x87 regs) which should lead to more readable dumps, but I didn't include
the whole list since that seems unnecessary.
X86_MC::initLLVMToSEHAndCVRegMapping is pretty ugly, but at least it's not
relying on magic constants anymore. The TODO of using tablegen still stands.
Reid Kleckner [Tue, 3 Oct 2017 17:59:02 +0000 (17:59 +0000)]
[DebugInfo] Correctly coalesce DBG_VALUEs that mix direct and indirect values
Summary:
This should fix a regression introduced by r313786, which switched from
MachineInstr::isIndirectDebugValue() to checking if operand 1 is an
immediate. I didn't have a test case for it until now.
A single UserValue, which approximates a user variable, may have many
DBG_VALUE instructions that disagree about whether the variable is in
memory or in a virtual register. This will become much more common once
we have llvm.dbg.addr, but you can construct such a test case manually
today with llvm.dbg.value.
Before this change, we would get two UserValues: one for direct and one
for indirect DBG_VALUE instructions describing the same variable. If we
build separate interval maps for direct and indirect locations, we will
end up accidentally coalescing identical DBG_VALUE intervals that need
to remain separate because they are broken up by intervals of the
opposite direct-ness.
Use sched_getaffinity instead of std::thread::hardware_concurrency.
The issue with std::thread::hardware_concurrency is that it forwards
to libc and some implementations (like glibc) don't take thread
affinity into consideration.
With this change a llvm program that can execute in only 2 cores will
use 2 threads, even if the machine has 32 cores.
This makes benchmarking a lot easier, but should also help if someone
doesn't want to use all cores for compilation for example.
Mohammad Shahid [Tue, 3 Oct 2017 15:28:48 +0000 (15:28 +0000)]
[SLP] Vectorize jumbled memory loads.
Summary:
This patch tries to vectorize loads of consecutive memory accesses, accessed
in non-consecutive or jumbled way. An earlier attempt was made with patch D26905
which was reverted back due to some basic issue with representing the 'use mask' of
jumbled accesses.
This patch fixes the mask representation by recording the 'use mask' in the usertree entry.
Oliver Stannard [Tue, 3 Oct 2017 14:38:52 +0000 (14:38 +0000)]
[ARM] Use table-gen'd assembly operand diags in ARM asm parser
This switches the ARM AsmParser to use assembly operand diagnostics from
tablegen, rather than a switch statement on the ARMMatchResultTy. It
moves the existing diagnostic strings to tablegen, but adds no new ones,
so this is NFC except for one diagnostic string that had an off-by-1 error
in the hand-written switch statement.
Oliver Stannard [Tue, 3 Oct 2017 14:34:57 +0000 (14:34 +0000)]
[AsmParser] Add DiagnosticString to AsmOperands in tablegen
This adds a DiagnosticString member to the AsmOperand tablegen class, so
that the diagnostic text to be used when an assembly operand is
incorrect can be stored in the tablegen description of the operand,
rather than in a separate switch statement in the AsmParser.
If DiagnosticString is used for any operands, tablegen will emit a
getMatchKindDiag function, to map from diagnostic enums to strings.
Jakub Kuderski [Tue, 3 Oct 2017 14:33:41 +0000 (14:33 +0000)]
[Dominators] Add DFS number verification
Summary:
This patch teaches the DominatorTree verifier to check DFS In/Out numbers which are used to answer dominance queries.
DFS number verification is done in O(nlogn), so it shouldn't add much overhead on top of the O(n^3) sibling property verification.
This check should detect errors like the one spotted in PR34466 and related bug reports.
The patch also cleans up the DFS calculation a bit, as all constructed trees should have a single root now.
I see 2 new test failures when running check-all after this change:
John Brawn [Tue, 3 Oct 2017 13:08:22 +0000 (13:08 +0000)]
[CGP] Make optimizeMemoryInst capable of handling multiple AddrModes
Currently optimizeMemoryInst requires that all of the AddrModes it sees are
identical. This patch makes it capable of tracking multiple AddrModes, so long
as they differ in at most one field.
This patch does nothing by itself, but later patches will make use of it to
insert or reuse phi or select instructions for the differing fields.
John Brawn [Tue, 3 Oct 2017 13:04:15 +0000 (13:04 +0000)]
[CGP] In optimizeMemoryInst handle select similarly to phi
This lets us optimize away selects that perform the same address computation in
two different ways and is also the first step towards being able to handle
selects between two different, but compatible, address computations.
Oliver Stannard [Tue, 3 Oct 2017 12:45:18 +0000 (12:45 +0000)]
[ARM, Asm] Fix ubsan failure caused by out-of-range enum value
In this code, we use ~0U as a sentinel value for any operand class that doesn't
have a user-friendly error message, but this value isn't in range of the
MatchClassKind enum, so we need to ensure it does not get passed to isSubclass.
Simon Pilgrim [Tue, 3 Oct 2017 12:01:31 +0000 (12:01 +0000)]
[X86][SSE] Add support for lowering shuffles to PACKSS/PACKUS
If the upper bits of a truncation shuffle patterns have at least the minimum number of sign/zero bits on their inputs then we can safely use PACKSS/PACKUS as shuffles.
Partial fix for https://bugs.llvm.org/show_bug.cgi?id=34773
[InlineCost, NFC] Extract code dealing with inbounds GEPs from visitGetElementPtr into a function
The code responsible for analysis of inbounds GEPs is extracted into a separate
function: CallAnalyzer::canFoldInboundsGEP. With the patch SROA
enabling/disabling code is localized at one place instead of spreading across
the code of CallAnalyzer::visitGetElementPtr.
[DebugInfo] Handle endianness when moving debug info for split integer values (reapplied)
Summary:
Take the target's endianness into account when splitting the
debug information in DAGTypeLegalizer::SetExpandedInteger.
This patch fixes so that, for big-endian targets, the fragment
expression corresponding to the high part of a split integer
value is placed at offset 0, in order to correctly represent
the memory address order.
I have attached a PPC32 reproducer where the resulting DWARF
pieces for a 64-bit integer were incorrectly reversed.
Original patch was reverted due to using -stop-after=isel in
the test case (but that is only working when AMDGPU target
is included in the llc build). The test case has now been
updated to use -stop-before=expand-isel-pseudos instead.
Oliver Stannard [Tue, 3 Oct 2017 10:26:11 +0000 (10:26 +0000)]
[ARM] Use new assembler diags for ARM
This converts the ARM AsmParser to use the new assembly matcher error
reporting mechanism, which allows errors to be reported for multiple
instruction encodings when it is ambiguous which one the user intended
to use.
By itself this doesn't improve many error messages, because we don't have
diagnostic text for most operand types, but as we add that then this will allow
more of those diagnostic strings to be used when they are relevant.
Oliver Stannard [Tue, 3 Oct 2017 09:33:12 +0000 (09:33 +0000)]
[Assembler] Report multiple near misses for invalid instructions
The current table-generated assembly instruction matcher returns a
64-bit error code when matching fails. Since multiple instruction
encodings with the same mnemonic can fail for different reasons, it uses
some heuristics to decide which message is important.
This heuristic does not work well for targets that have many encodings
with the same mnemonic but different operands, or which have different
versions of instructions controlled by subtarget features, as it is hard
to know which encoding the user was intending to use.
Instead of trying to improve the heuristic in the table-generated
matcher, this patch changes it to report a list of near-miss encodings.
This list contains an entry for each encoding with the correct mnemonic,
but with exactly one thing preventing it from being valid. This thing
could be a single invalid operand, a missing target feature or a failed
target-specific validation function.
The target-specific assembly parser can then report an error message
giving multiple options for instruction variants that the user may have
been trying to use. For example, I am working on a patch to use this for
ARM, which can give this error for an invalid instruction for ARMv6-M:
<stdin>:8:3: error: invalid instruction, multiple near-miss encodings found
adds r0, r1, #0x8
^
<stdin>:8:3: note: for one encoding: instruction requires: thumb2
adds r0, r1, #0x8
^
<stdin>:8:16: note: for one encoding: expected an integer in range [0, 7]
adds r0, r1, #0x8
^
<stdin>:8:16: note: for one encoding: expected a register in range [r0, r7]
adds r0, r1, #0x8
^
This also allows the target-specific assembly parser to apply its own
heuristics to suppress some errors. For example, the error "instruction
requires: arm-mode" is never going to be useful when targeting an
M-profile architecture (which does not have ARM mode).
This patch just adds the target-independent mechanism for doing this,
all targets still use the old mechanism. I've added a bit in the
AsmParser tablegen class to allow targets to switch to this new
mechanism. To use this, the target-specific assembly parser will have to
be modified for the change in signature of MatchInstructionImpl, and to
report errors based on the list of near-misses.
Sjoerd Meijer [Tue, 3 Oct 2017 08:54:15 +0000 (08:54 +0000)]
ISel type legalization: add debug messages. NFCI.
This adds some more debug messages to the type legalizer and functions
like PromoteNode, ExpandNode, ExpandLibCall in an attempt to make
the debug messages a little bit more informative and useful.
Mikael Holmen [Tue, 3 Oct 2017 06:03:49 +0000 (06:03 +0000)]
[Lint] Avoid failed assertion by fetching the proper pointer type
Summary:
When checking if a constant expression is a noop cast we fetched the
IntPtrType by doing DL->getIntPtrType(V->getType())). However, there can
be cases where V doesn't return a pointer, and then getIntPtrType()
triggers an assertion.
Now we pass DataLayout to isNoopCast so the method itself can determine
what the IntPtrType is.
Craig Topper [Tue, 3 Oct 2017 04:55:04 +0000 (04:55 +0000)]
[InstCombine] Replace an equality compare of two APInt pointers with a compare of the APInts themselves.
Apparently this works by virtue of the fact that the pointers are pointers to the APInts stored inside of the ConstantInt objects. But I really don't think we should be relying on that.
Rui Ueyama [Tue, 3 Oct 2017 03:09:05 +0000 (03:09 +0000)]
Rewrite a function so that it doesn't use pointers to pointers. NFC.
Previous code was a bit puzzling because of its use of pointers.
In this patch, we pass a vector and its offsets, instead of pointers to
vector elements.
LTO: Improve error reporting when adding a cache entry.
Move error handling code next to the code that returns the error,
and change the error message in order to distinguish it from a similar
error message elsewhere in this file.
Matt Arsenault [Tue, 3 Oct 2017 00:06:41 +0000 (00:06 +0000)]
AMDGPU: Remove global isGCN predicates
These are problematic because they apply to everything,
and can easily clobber whatever more specific predicate
you are trying to add to a function.
Currently instructions use SubtargetPredicate/PredicateControl
to apply this to patterns applied to an instruction definition,
but not to free standing Pats. Add a wrapper around Pat
so the special PredicateControls requirements can be appended
to the final predicate list like how Mips does it.