Pavel Labath [Thu, 21 Mar 2019 10:21:55 +0000 (10:21 +0000)]
Fix two more issues with r356652
The first problem was a use-after-free in the tests (detected by asan
bots). The temporary array created for the "create" call is guaranteed
to live only until the end of the statement. The fix there is to store
the test data in a local variable to ensure it has the right lifetime
The second issue is broken BUILD_SHARED_LIBS build, which I fix by
adding the appropriate BinaryFormat dependency to the Object unit tests.
Pavel Labath [Thu, 21 Mar 2019 09:18:59 +0000 (09:18 +0000)]
[Object] Add basic minidump support
Summary:
This patch adds basic support for reading minidump files. It contains
the definitions of various important minidump data structures (header,
stream directory), and of one minidump stream (SystemInfo). The ability
to read other streams will be added in follow-up patches. However, all
streams can be read even now as raw data, which means lldb's minidump
support (where this code is taken from) can be immediately rebased on
top of this patch as soon as it lands.
As we don't have any support for generating minidump files (yet), this
tests the code via unit tests with some small handcrafted binaries in
the form of c char arrays.
Alina Sbirlea [Thu, 21 Mar 2019 05:02:05 +0000 (05:02 +0000)]
[BasicAA] Reduce no of map seaches [NFCI].
Summary:
This is a refactoring patch.
- Reduce the number of map searches by reusing the iterator.
- Add asserts to check that the entry is in the cache, as this is something BasicAA relies on to avoid infinite recursion.
Code archaeology in D59315 revealed that MSSA should never be moved.
Rather than trying to check dynamically that this hasn't happened in the
verify() functions of Walkers, it's likely best to just delete its move
constructor.
Since all these verify() functions did is check that MSSA hasn't moved,
this allows us to remove these verify functions.
I can readd the verification checks if someone's super concerned about
us trying to `memcpy` MemorySSA or something somewhere, but I imagine we
have other problems if we're trying anything like that...
Craig Topper [Wed, 20 Mar 2019 23:35:49 +0000 (23:35 +0000)]
[X86] Add CMPXCHG8B feature flag. Set it for all CPUs except i386/i486 including 'generic'. Disable use of CMPXCHG8B when this flag isn't set.
CMPXCHG8B was introduced on i586/pentium generation.
If its not enabled, limit the atomic width to 32 bits so the AtomicExpandPass will expand to lib calls. Unclear if we should be using a different limit for other configs. The default is 1024 and experimentation shows that using an i256 atomic will cause a crash in SelectionDAG.
Michael Trent [Wed, 20 Mar 2019 23:21:16 +0000 (23:21 +0000)]
Fix Mach-O bind and rebase validation errors in libObject
Summary:
llvm-objdump (via libObject) validates DYLD_INFO rebase and bind
entries against the basic structure found in the Mach-O file before
evaluating the contents of those entries. Certain malformed Mach-Os can
defeat the validation check and force llvm-objdump (libObject) to crash.
The previous logic verified a rebase or bind started in a valid Mach-O
section, but did not verify that the section wholely contained the
fixup. It also generally allows rebases or binds to start immediately
after a valid section even if that range is not itself part of a valid
section. Finally, bind and rebase opcodes that indicate more than one
fixup (apply N times...) are not completely validated: only the first
and final fixups are checked.
The previous logic also rejected certain binaries as false positives.
Some bind and rebase opcodes can modify the state machine such that the
next bind or rebase will fail. libObject will reject these opcodes as
invalid in order to be helpful and print an error message associated
with the instruction that caused the problem, even though the binary is
not actually illegal until it consumes the invalid state in the state
machine. In other words, libObject may reject a Mach-O binary that
Apple's dynamic linker may consider legal. The original version of
macho-rebase-add-addr-uleb-too-big is an example of such a binary.
I have replaced the existing checkSegAndOffset and checkCountAndSkip
functions with a single function, checkSegAndOffsets, which validates
all of the fixups realized by a DYLD_INFO opcode. checkSegAndOffsets
verifies that a Mach-O section fully contains each fixup. Every fixup
realized by an opcode is validated, and some (but not all!)
inconsistencies in the state machine are allowed until a fixup is
realized. This means that libObject may fail on an opcode that realizes
a fixup, not on the opcode that introduced the arithmetic error.
Existing test cases have been modified to reflect the changes in error
messages returned by libObject. What's more, the test case for
macho-rebase-add-addr-uleb-too-big has been modified so that it actually
triggers the error condition; the new code in libObject considers the
original test binary "legal".
Tim Renouf [Wed, 20 Mar 2019 22:02:09 +0000 (22:02 +0000)]
[AMDGPU] Do not generate spurious PAL metadata
My previous fix rL356591 "[AMDGPU] Added MsgPack format PAL metadata"
accidentally caused a spurious PAL metadata .note record to be emitted
for any AMDGPU output. That caused failures in the lld test
amdgpu-relocs.s. Fixed.
Craig Topper [Wed, 20 Mar 2019 21:30:20 +0000 (21:30 +0000)]
[X86] Call lowerShuffleAsBitMask for 512-bit vectors in lowerShuffleAsBlend.
This patch enables the use of lowerShuffleAsBitMask for 512-bit blends before
falling back to move immedate, GPR to k-register, and masked op.
I had to make some changes to support v8i64 when i64 is not a legal type. And to
support floating point types.
This trades a load for the move immediate and GPR move which is higher latency.
But its probably better for register pressure not having to hop through other
register classes. The load+and should play better with LICM and
rematerialization I think.
Matt Arsenault [Wed, 20 Mar 2019 20:41:34 +0000 (20:41 +0000)]
AMDGPU: Don't look for constant in insert/extract_vector_elt regbankselect
The constantness shouldn't change the register bank choice. We also
don't need to restrict this to only indexing VGPRs, since it's
possible to index SGPRs (but SelectionDAG made using this
difficult). Allow directly indexing SGPRs when appropriate.
Thomas Lively [Wed, 20 Mar 2019 20:26:45 +0000 (20:26 +0000)]
[WebAssembly] Target features section
Summary:
Implements a new target features section in assembly and object files
that records what features are used, required, and disallowed in
WebAssembly objects. The linker uses this information to ensure that
all objects participating in a link are feature-compatible and records
the set of used features in the output binary for use by optimizers
and other tools later in the toolchain.
The "atomics" feature is always required or disallowed to prevent
linking code with stripped atomics into multithreaded binaries. Other
features are marked used if they are enabled globally or on any
function in a module.
Future CLs will add linker flags for ignoring feature compatibility
checks and for specifying the set of allowed features, implement using
the presence of the "atomics" feature to control the type of memory
and segments in the linked binary, and add front-end flags for
relaxing the linkage policy for atomics.
Michael Liao [Wed, 20 Mar 2019 20:18:56 +0000 (20:18 +0000)]
[AMDGPU] Fix clamp bit DAG operand
Summary:
- Should use `targetconstant` instead of `constant` operand for clamp
bit, which is expected as an immediate operand. Under certain
conditions, such as a common `i1 false` constant is used in other
place and selected before the instruction with clamp bit, register
operand may be added instead of immediate one. Use `targetcosntant` to
enforce that.
Eli Friedman [Wed, 20 Mar 2019 19:40:45 +0000 (19:40 +0000)]
[ARM] Eliminate redundant "mov rN, sp" instructions in Thumb1.
This takes sequences like "mov r4, sp; str r0, [r4]", and optimizes them
to something like "str r0, [sp]".
For regular stack variables, this optimization was already implemented:
we lower loads and stores using frame indexes, which are expanded later.
However, when constructing a call frame for a call with more than four
arguments, the existing optimization doesn't apply. We need to use
stores which are actually relative to the current value of sp, and don't
have an associated frame index.
This patch adds a special case to handle that construct. At the DAG
level, this is an ISD::STORE where the address is a CopyFromReg from SP
(plus a small constant offset).
This applies only to Thumb1: in Thumb2 or ARM mode, a regular store
instruction can access SP directly, so the COPY gets eliminated by
existing code.
The change to ARMDAGToDAGISel::SelectThumbAddrModeSP is a related
cleanup: we shouldn't pretend that it can select anything other than
frame indexes.
Rafael Auler [Wed, 20 Mar 2019 19:20:07 +0000 (19:20 +0000)]
[Linker] Fix crash handling appending linkage
Summary:
When linking two llvm.used arrays, if the resulting merged
array ends up with duplicated elements (with the same name) but with
different types, the IRLinker was crashing. This was supposed to be
legal, as the IRLinker bitcasts elements to match types in these
situations.
This bug was exposed by D56928 in clang to support attribute used
in member functions of class templates. Crash happened when self-hosting
with LTO. Since LLVM depends on attribute used to generate code
for the dump() method, ubiquitous in the code base, many input bc
had a definition of this method referenced in their llvm.used array.
Some of these classes got optimized, changing the type of the first
parameter (this) in the dump method, leading to a scenario with a
pool of valid definitions but some with a different type, triggering
this bug.
This is a memory bug: ValueMapper depends on (calls) the materializer
provided by IRLinker, and this materializer was freely calling RAUW
methods whenever a global definition was updated in the temporary merged
output file. However, replaceAllUsesWith may or may not destroy
constants that use this global. If the linked definition has a type
mismatch regarding the new def and the old def, the materializer would
bitcast the old type to the new type and the elements of the llvm.used
array, which already uses bitcast to i8*, would end up with elements
cascading two bitcasts. RAUW would then indirectly call the
constantfolder to update the constant to the new ref, which would,
instead of updating the constant, destroy it to be able to create
a new constant that folds the two bitcasts into one. The problem is that
ValueMapper works with pointers to the same constants that may be
getting destroyed by RAUW. Obviously, RAUW can update references in the
Module to do not use the old destroyed constant, but it can't update
ValueMapper's internal pointers to these constants, which are now
invalid.
The approach here is to move the task of RAUWing old definitions
outside of the materializer.
Test Plan:
Added LIT test case, tested clang self-hosting with D56928 and
verified it works
Philip Reames [Wed, 20 Mar 2019 18:44:58 +0000 (18:44 +0000)]
Simplify operands of masked stores and scatters based on demanded elements
If we know we're not storing a lane, we don't need to compute the lane. This could be improved by using the undef element result to further prune the mask, but I want to separate that into its own change since it's relatively likely to expose other problems.
Alina Sbirlea [Wed, 20 Mar 2019 18:33:37 +0000 (18:33 +0000)]
[LICM & MemorySSA] Don't sink/hoist stores in the presence of ordered loads.
Summary:
Before this patch, if any Use existed in the loop, with a defining
access in the loop, we conservatively decide to not move the store.
What this approach was missing, is that ordered loads are not Uses, they're Defs
in MemorySSA. So, even when the clobbering walker does not find that
volatile load to interfere, we still cannot hoist a store past a
volatile load.
Resolves PR41140.
Nikita Popov [Wed, 20 Mar 2019 18:16:02 +0000 (18:16 +0000)]
[ValueTracking] Compute range for abs without nsw
This is a small followup to D59511. The code that was moved into
computeConstantRange() there is a bit overly conversative: If the
abs is not nsw, it does not compute any range. However, abs without
nsw still has a well-defined contiguous unsigned range from 0 to
SIGNED_MIN. This is a lot less useful than the usual 0 to SIGNED_MAX
range, but if we're already here we might as well specify it...
Tim Renouf [Wed, 20 Mar 2019 17:42:00 +0000 (17:42 +0000)]
[AMDGPU] Factored PAL metadata handling out into its own class
Summary:
This commit introduces a new AMDGPUPALMetadata class that:
* is inside the AMDGPU target;
* keeps an in-memory representation of PAL metadata;
* provides a method to read the frontend-supplied metadata from LLVM IR;
* provides methods for the asm printer to set metadata items;
* provides methods to write the metadata as a binary blob to put in a
.note record or as an asm directive;
* provides a method to read the metadata as a binary blob from a .note
record.
Because llvm-readobj cannot call directly into a target, I had to remove
llvm-readobj's ability to dump PAL metadata, pending a resolution to
https://reviews.llvm.org/D52821
Sjoerd Meijer [Wed, 20 Mar 2019 14:15:46 +0000 (14:15 +0000)]
[TTI] getMemcpyCost
This adds new function getMemcpyCost to TTI so that the cost of a memcpy can be
modeled and queried. The default implementation returns Expensive, but targets
can override this function to model the cost more accurately.
Andrea Di Biagio [Wed, 20 Mar 2019 11:21:15 +0000 (11:21 +0000)]
[X86] Remove X86 specific dag nodes for RDTSC/RDTSCP/RDPMC. NFCI
This patch removes the following dag node opcodes from namespace X86ISD:
RDTSC_DAG,
RDTSCP_DAG,
RDPMC_DAG
The logic that expands RDTSC/RDPMC/XGETBV intrinsics is basically the same. The
only differences are:
RDTSC/RDTSCP don't implicitly read ECX.
RDTSCP also implicitly writes ECX.
I moved the common expansion logic into a helper function with the goal to get
rid of code repetition. That helper is now used for the expansion of
RDTSC/RDTSCP/RDPMC/XGETBV intrinsics.
David Stuttard [Wed, 20 Mar 2019 09:29:55 +0000 (09:29 +0000)]
[AMDGPU] Allow MIMG with no uses in adjustWritemask in isel
Summary:
If an MIMG instruction has managed to get through to adjustWritemask in isel but
has no uses (and doesn't enable TFC) then prevent an assertion by not attempting
to adjust the writemask.
Eli Friedman [Tue, 19 Mar 2019 21:48:08 +0000 (21:48 +0000)]
[ARM] Make sure to save/restore LR when we use tBfar.
This change does two things. One, it ensures compilation will abort
instead of miscompiling if ARMFrameLowering::determineCalleeSaves
chooses not to save LR in a case where it's necessary. Two, it changes
the way we estimate the size of a function to be more conservative in
the presence of constant pool entries and jump tables.
EstimateFunctionSizeInBytes probably still isn't really conservative
enough, but I'm not sure how we can come up with a reliable estimate
before constant islands runs.
Reid Kleckner [Tue, 19 Mar 2019 21:40:59 +0000 (21:40 +0000)]
Remove MSVC compat hack since the inline keyword was added in 2015
Our minimum MSVC toolchain requirement is greater than 2015, so we don't
need this conditional macro anymore. New versions of MSVC apparently
have a header, xkeycheck.h, to check that keywords haven't been
redefined.
Reland "[Remarks] Add a new Remark / RemarkParser abstraction"
This adds a Remark class that allows us to share code when working with
remarks.
The C API has been updated to reflect this. Instead of the parser
generating C structs, it's now using a C++ object that is used through
opaque pointers in C. This gives us much more flexibility on what
changes we can make to the internal state of the object and interacts
much better with scenarios where the library is used through dlopen.
* C API updates:
* move from C structs to opaque pointers and functions
* the remark type is now an enum instead of a string
* unit tests updates:
* use mostly the C++ API
* keep one test for the C API
* rename to YAMLRemarksParsingTest
* a typo was fixed: AnalysisFPCompute -> AnalysisFPCommute.
* a new error message was added: "expected a remark tag."
* llvm-opt-report has been updated to use the C++ parser instead of the
C API
Florian Hahn [Tue, 19 Mar 2019 20:37:06 +0000 (20:37 +0000)]
[DwarfDebug] Skip entries to big for 16 bit size field in Dwarf < 5.
Nothing prevents entries from being bigger than the 16 bit size field in
Dwarf < 5. For entries that are too big, just emit an empty entry
instead of crashing.
Robert Lougher [Tue, 19 Mar 2019 20:24:28 +0000 (20:24 +0000)]
[TailCallElim] Add tailcall elimination pass to LTO pipelines
LTO provides additional opportunities for tailcall elimination due to
link-time inlining and visibility of nocapture attribute. Testing showed
negligible impact on compilation times.
Philip Reames [Tue, 19 Mar 2019 20:10:00 +0000 (20:10 +0000)]
Demanded elements support for masked.load and masked.gather
Teach instcombine to propagate demanded elements through a masked load or masked gather instruction. This is in the broader context of improving vector pointer instcombine under https://reviews.llvm.org/D57140.
Matt Arsenault [Tue, 19 Mar 2019 19:33:12 +0000 (19:33 +0000)]
CodeGen: Refactor regallocator command line and target selection
This will allow targets more flexibility to replace the
register allocator core passes. In a future commit,
AMDGPU will run the core register assignment passes
twice, and will also want to disallow using the
standard -regalloc option.
Matt Arsenault [Tue, 19 Mar 2019 19:16:04 +0000 (19:16 +0000)]
RegAllocFast: Do not allocate registers for undef uses
Do not actually allocate a register for an undef use. Previously we we
would create unnecessary reload instruction for undef uses where the
register wasn't live.
Matt Arsenault [Tue, 19 Mar 2019 19:01:34 +0000 (19:01 +0000)]
RegAllocFast: Remove early selection loop, the spill calculation will report cost 0 anyway for free regs
The 2nd loop calculates spill costs but reports free registers as cost
0 anyway, so there is little benefit from having a separate early
loop.
Surprisingly this is not NFC, as many register are marked regDisabled
so the first loop often picks up later registers unnecessarily instead
of the first one available in the allocation order...
Philip Reames [Tue, 19 Mar 2019 18:27:18 +0000 (18:27 +0000)]
Allow unordered loads to be considered invariant in CodeGen
The actual code change is fairly straight forward, but exercising it isn't. First, it turned out we weren't adding the appropriate flags in SelectionDAG. Second, it turned out that we've got some optimization gaps, so obvious test cases don't work.
My first attempt (in atomic-unordered.ll) points out a deficiency in our peephole-opt folding logic which I plan to fix separately. Instead, I'm exercising this through MachineLICM.
[Remarks] Add a new Remark / RemarkParser abstraction
This adds a Remark class that allows us to share code when working with
remarks.
The C API has been updated to reflect this. Instead of the parser
generating C structs, it's now using a C++ object that is used through
opaque pointers in C. This gives us much more flexibility on what
changes we can make to the internal state of the object and interacts
much better with scenarios where the library is used through dlopen.
* C API updates:
* move from C structs to opaque pointers and functions
* the remark type is now an enum instead of a string
* unit tests updates:
* use mostly the C++ API
* keep one test for the C API
* rename to YAMLRemarksParsingTest
* a typo was fixed: AnalysisFPCompute -> AnalysisFPCommute.
* a new error message was added: "expected a remark tag."
* llvm-opt-report has been updated to use the C++ parser instead of the
C API
Nikita Popov [Tue, 19 Mar 2019 17:53:56 +0000 (17:53 +0000)]
[ValueTracking] Use computeConstantRange() for unsigned add/sub overflow
Improve computeOverflowForUnsignedAdd/Sub in ValueTracking by
intersecting the computeConstantRange() result into the ConstantRange
created from computeKnownBits(). This allows us to detect some
additional never/always overflows conditions that can't be determined
from known bits.
This revision also adds basic handling for constants to
computeConstantRange(). Non-splat vectors will be handled in a followup.
The signed case will also be handled in a followup, as it needs some
more groundwork.
Philip Reames [Tue, 19 Mar 2019 17:20:49 +0000 (17:20 +0000)]
[AtomicExpand] Fix a crash bug when lowering unordered loads to cmpxchg
Add tests for wider atomic loads and stores. In the process, fix a crasher where we appearently handled unorder stores, but not loads, when lowering to cmpxchg idioms.
Justin Bogner [Tue, 19 Mar 2019 16:52:00 +0000 (16:52 +0000)]
[DAGCombine] Fix a miscompile when reducing BUILD_VECTORs to a shuffle
In r311255 we added a case where we split vectors whose elements are
all derived from the same input vector so that we could shuffle it
more efficiently. In doing so, createBuildVecShuffle was taught to
adjust for the fact that all indices would be based off of the first
vector when this happens, but it's possible for the code that checked
that to fire incorrectly if we happen to have a BUILD_VECTOR of
extracts from subvectors and don't hit this new optimization.
Instead of trying to detect if we've split the vector by checking if
we have extracts from the same base vector, we can just pass that
information into createBuildVecShuffle, avoiding the miscompile.
Philip Reames [Tue, 19 Mar 2019 16:46:56 +0000 (16:46 +0000)]
[Tests] Update to newer ISA
There are some issues w/missed opts on older platforms, but that's not the purpose of this test. Using a newer API points out that some TODOs are already handled, and allows addition of tests to exercise other issues (future patch.)
Sanjay Patel [Tue, 19 Mar 2019 16:39:17 +0000 (16:39 +0000)]
[InstCombine] fold logic-of-nan-fcmps (PR41069)
Combine 2 fcmps that are checking for nan-ness:
and (fcmp ord X, 0), (and (fcmp ord Y, 0), Z) --> and (fcmp ord X, Y), Z
or (fcmp uno X, 0), (or (fcmp uno Y, 0), Z) --> or (fcmp uno X, Y), Z
This is an exact match for a minimal reassociation pattern.
If we want to handle this more generally that should go in
the reassociate pass and allow removing this code.
This should fix:
https://bugs.llvm.org/show_bug.cgi?id=41069
Simon Pilgrim [Tue, 19 Mar 2019 16:24:55 +0000 (16:24 +0000)]
[SelectionDAG] Handle unary SelectPatternFlavor for ABS case in SelectionDAGBuilder::visitSelect
These changes are related to PR37743 and include:
SelectionDAGBuilder::visitSelect handles the unary SelectPatternFlavor::SPF_ABS case to build ABS node.
Delete the redundant recognizer of the integer ABS pattern from the DAGCombiner.
Add promoting the integer ABS node in the LegalizeIntegerType.
Expand-based legalization of integer result for the ABS nodes.
Expand-based legalization of ABS vector operations.
Add some integer abs testcases for different typesizes for Thumb arch
Add the custom ABS expanding and change the SAD pattern recognizer for X86 arch: The i64 result of the ABS is expanded to:
tmp = (SRA, Hi, 31)
Lo = (UADDO tmp, Lo)
Hi = (XOR tmp, (ADDCARRY tmp, hi, Lo:1))
Lo = (XOR tmp, Lo)
The "detectZextAbsDiff" function is changed for the recognition of pattern with the ABS node. Given a ABS node, detect the following pattern:
(ABS (SUB (ZERO_EXTEND a), (ZERO_EXTEND b))).
Change integer abs testcases for codegen with the ABS node support for AArch64.
Indicate that the ABS is legal for the i64 type when the NEON is supported.
Change the integer abs testcases to show changing of codegen.
Add combine and legalization of ABS nodes for Thumb arch.
Extend 'matchSelectPattern' to recognize the ABS patterns with ICMP_SGE condition.
For discussion, see https://bugs.llvm.org/show_bug.cgi?id=37743
Jordan Rupprecht [Tue, 19 Mar 2019 16:09:54 +0000 (16:09 +0000)]
[llvm-ar] Support N [count] modifier
Summary:
GNU ar supports the 'N' count modifier for the extract (x) and delete (d) operations. When an archive contains multiple members with the same name, this can be used to extract (or delete) them individually. For example:
```
$ llvm-ar t archive.a
foo
foo
$ llvm-ar x archive.a
-> Writes foo twice, overwriting it the second time :( :(
$ llvm-ar xN 1 archive.a foo && mv foo foo.1
$ llvm-ar xN 2 archive.a foo && mv foo foo.2
-> Write foo twice, renaming it in between invocations to preserve all versions
```
Neil Henning [Tue, 19 Mar 2019 15:50:24 +0000 (15:50 +0000)]
[AMDGPU] Ban i8 min3 promotion.
I found this really weird WWM-related case whereby through the WWM
transformations our isel lowering was trying to promote 2 min's into a
min3 for the i8 type, which our hardware doesn't support.
The new min3_i8.ll test case would previously spew the error:
projects / llvm / log