Simon Pilgrim [Wed, 1 May 2019 17:13:35 +0000 (17:13 +0000)]
[X86][SSE] Fold scalar horizontal add/sub for non-0/1 element extractions
We already perform horizontal add/sub if we extract from elements 0 and 1, this patch extends it to non-0/1 element extraction indices (as long as they are from the lowest 128-bit vector).
Daniel Sanders [Wed, 1 May 2019 16:52:29 +0000 (16:52 +0000)]
[globalisel] Update the legalizer documentation
Summary:
* The getActionDefinitionsBuilder() is now documented.
* Includes descriptions of the various actions (legal*, widenScalar*, lower*,
etc).
* Includes descriptions of the various predicates (*If, *For,
*ForCartesianProduct, etc.)
* Includes the rule-order details
* Removed the out-of-date prohibition on non-power-of-2 types.
* Removed the Vector types section since it was incorrect and vectors follow the
same ruleset as scalars. They're only special in the sense that more of the
actions and predicates are meaningful for them (e.g. moreElements).
* Clarified the position on context sensitive legality (which is not permitted)
and contrasted this with context sensitive legalization (which is permitted).
Nico Weber [Wed, 1 May 2019 16:45:15 +0000 (16:45 +0000)]
Option spell checking: Penalize delimiter flags if input has no argument
If the user passes a flag like `-version` to a program, it's more likely
they mean `--version` than `-version:`, since there's no parameter
passed. Hence, give delimited arguments a penalty of 1 if the user input
doesn't contain the delimiter or no data after it.
The motivation is that with this, lld-link can suggest "--version"
instead of "-version:" for "-version" and "-nodefaultlib" instead of
"-nodefaultlib:" for "-nodefaultlibs".
Teresa Johnson [Wed, 1 May 2019 16:26:59 +0000 (16:26 +0000)]
[ThinLTO] Fix unreachable code when parsing summary entries.
Summary:
Early returns were causing some code to be skipped. This was missed
since the summary entries are typically at the end of the llvm assembly
file.
Keno Fischer [Wed, 1 May 2019 15:58:24 +0000 (15:58 +0000)]
[SCEV] Use isKnownViaNonRecursiveReasoning for smax simplification
Summary:
Commit
rL331949: SCEV] Do not use induction in isKnownPredicate for simplification umax
changed the codepath for umax from isKnownPredicate to
isKnownViaNonRecursiveReasoning to avoid compile time blow up (and as
I found out also stack overflows). However, there is an exact copy of
the code for umax that was lacking this change. In D50167 I want to unify
these codepaths, but to avoid that being a behavior change for the smax
case, pull this independent bit out of it.
Hubert Tong [Wed, 1 May 2019 15:47:16 +0000 (15:47 +0000)]
[lit][tests][AIX] Update expected form of diagnostic messages; use `not` to normalize non-zero exit values
Summary:
Various tests in the `lit` testing suite expect specific return codes
and forms of diagnostic message from utility programs. As per
POSIX.1-2017 XCU Section 1.4, Utility Description Defaults, "[the]
format of diagnostic messages for most utilities is unspecified".
The STDERR subsections of the `cat` and `wc` utilities merely indicate
that "[the] standard error shall be used only for diagnostic messages".
The corresponding EXIT STATUS subsections merely indicate, with regard
to errors, an exit value of >0.
The affected tests are updated to accept the applicable diagnostic
message as produced by the utilities on AIX. The exit value is
normalized using `not` as necessary.
Hubert Tong [Wed, 1 May 2019 15:36:18 +0000 (15:36 +0000)]
[tests] Add host-byteorder-*-endian; update XFAILs of big-endian triples
Summary:
Triple components in `XFAIL` lines are tested against the target triple.
Various tests that are expected to fail on big-endian hosts are marked
as being `XFAIL` for big-endian targets. This patch corrects these tests
by having them test against a new `host-byteorder-big-endian` feature.
Nico Weber [Wed, 1 May 2019 14:46:17 +0000 (14:46 +0000)]
Fix OptTable::findNearest() adding delimiter for free
Prior to this, OptTable::findNearest() thought that the input `--foo`
had an editing distance of 0 from an existing flag `--foo=`, which made
it suggest flags with delimiters more often than flags without one.
After this, it correctly assigns this case an editing distance of 1.
Keno Fischer [Wed, 1 May 2019 14:39:11 +0000 (14:39 +0000)]
[LoopInfo] Faster implementation of setLoopID. NFC.
Summary:
This change was part of D46460. However, in the meantime rL341926 fixed the
correctness issue here. What remained was the performance issue in setLoopID
where it would iterate through all blocks in the loop and their successors,
rather than just the predecessor of the header (the later presumably being
much faster). We already have the `getLoopLatches` to compute precisely these
basic blocks in an efficient manner, so just use it (as the original commit
did for `getLoopID`).
Tim Northover [Wed, 1 May 2019 12:37:30 +0000 (12:37 +0000)]
DAG: allow DAG pointer size different from memory representation.
In preparation for supporting ILP32 on AArch64, this modifies the SelectionDAG
builder code so that pointers are allowed to have a larger type when "live" in
the DAG compared to memory.
Pointers get zero-extended whenever they are loaded, and truncated prior to
stores. In addition, a few not quite so obvious locations need updating:
* A GEP that has not been marked inbounds needs to enforce the IR-documented
2s-complement wrapping at the memory pointer size. Inbounds GEPs are
undefined if they overflow the address space, so no additional operations
are needed.
* Signed comparisons would give incorrect results if performed on the
zero-extended values.
This shouldn't affect CodeGen for now, but will become active when the AArch64
ILP32 support is committed.
Fangrui Song [Wed, 1 May 2019 09:28:24 +0000 (09:28 +0000)]
[llvm-readobj] Change -t to --symbols in tests. NFC
-t is --symbols in llvm-readobj but --section-details (unimplemented) in readelf.
The confusing option should not be used since we aim for improving
compatibility.
Keep just one llvm-readobj -t use case in test/tools/llvm-readobj/symbols.test
Fangrui Song [Wed, 1 May 2019 05:27:20 +0000 (05:27 +0000)]
[llvm-readobj] Change -long-option to --long-option in tests. NFC
We use both -long-option and --long-option in tests. Switch to --long-option for consistency.
In the "llvm-readelf" mode, -long-option is discouraged as it conflicts with grouped short options and it is not accepted by GNU readelf.
While updating the tests, change llvm-readobj -s to llvm-readobj -S to reduce confusion ("s" is --section-headers in llvm-readobj but --symbols in llvm-readelf).
Lang Hames [Wed, 1 May 2019 02:43:52 +0000 (02:43 +0000)]
[JITLink] Make sure we explicitly deallocate memory on failure.
JITLinkGeneric phases 2 and 3 (focused on applying fixups and finalizing memory,
respectively) may fail for various reasons. If this happens, we need to
explicitly de-allocate the memory allocated in phase 1 (explicitly, because
deallocation may also fail and so is implemented as a method returning error).
No testcase yet: I am still trying to decide on the right way to test totally
platform agnostic code like this.
GNU objcopy uses bfd_elf_get_default_section_type to decide the candidate section type,
which roughly translates to our [a] (I assume SEC_COMMON implies SHF_ALLOC):
I suggest we just ignore this clause and follow the common case
behavior, which is done in this patch. Rationales to do so:
If --set-section-flags is a no-op (osec->flags == isec->flags)
(corresponds to the "readonly" test in set-section-flags.test), GNU
objcopy will require (Sec.Flags & ELF::SHF_ALLOC). [a] is essentially:
This special case is not really useful. Non-SHF_ALLOC SHT_NOBITS
sections do not make much sense and it doesn't matter if they are
SHT_NOBITS or SHT_PROGBITS.
For all other RUN lines in set-section-flags.test, the new behavior
matches GNU objcopy, i.e. this patch improves compatibility.
Philip Reames [Tue, 30 Apr 2019 23:09:26 +0000 (23:09 +0000)]
[InstCombine] Limit a vector demanded elts rule which was producing invalid IR.
The demanded elts rules introduced for GEPs in https://reviews.llvm.org/rL356293 replaced vector constants with undefs (by design). It turns out that the LangRef disallows such cases when indexing structs. The right fix is probably to relax the langref requirement, and update other passes to expect the result, but for the moment, limit the transform to avoid compiler crashes.
This should fix https://bugs.llvm.org/show_bug.cgi?id=41624.
Dan Gohman [Tue, 30 Apr 2019 23:04:49 +0000 (23:04 +0000)]
[WebAssembly] Test the "wasm32-wasi" triple
Add triple tests for "wasm32-wasi" and "wasm64-wasi", and also remove the
"-musl" component from the existing wasm triple tests as we're not using that
in practice (WASI libc is derived in part from musl, but it is not fully
musl-compatible).
[AliasAnalysis/NewPassManager] Invalidate AAManager less often.
Summary:
This is a redo of D60914.
The objective is to not invalidate AAManager, which is stateless, unless
there is an explicit invalidate in one of the AAResults.
To achieve this, this patch adds an API to PAC, to check precisely this:
is this analysis not invalidated explicitly == is this analysis not abandoned == is this analysis stateless, so preserved without explicitly being marked as preserved by everyone
[PassManagerBuilder] Add option for interleaved loops, for loop vectorize.
Summary:
Match NewPassManager behavior: add option for interleaved loops in the
old pass manager, and use that instead of the flag used to disable loop unroll.
No changes in the defaults.
Lang Hames [Tue, 30 Apr 2019 21:27:56 +0000 (21:27 +0000)]
[ORC][JITLink] Name in-memory compiled objects after their source modules.
In-memory compiled object buffer identifiers will now be derived from the
identifiers of their source IR modules. This makes it easier to connect
in-memory objects with their source modules in debugging output.
Dan Gohman [Tue, 30 Apr 2019 19:17:59 +0000 (19:17 +0000)]
[WebAssembly] Support f16 libcalls
Add support for f16 libcalls in WebAssembly. This entails adding signatures
for the remaining F16 libcalls, and renaming gnu_f2h_ieee/gnu_h2f_ieee to
truncsfhf2/extendhfsf2 for consistency between f32 and f64/f128 (compiler-rt
already supports this).
[X86] If PreprocessISelDAG reorders a load before a call, make sure we remove dead nodes from the graph
The reordering can leave at least a dead TokenFactor in the graph. This cause the linearize scheduler to fail with something like the assert seen in PR22614. This is only one of many ways we can break the linearize scheduler today so I can't say for sure that any of the other failures in that bug were caused by this issue.
This takes the heavy hammer approach of just running RemoveDeadNodes unconditionally at the end of the PreprocessISelDAG. If this turns out to be a compile time hit, we can try to refine it.
[X86] Initial cleanups on the FixupLEAs pass. Separate Atom LEA creation from other LEA optimizations.
This removes some of the class variables. Merge basic block processing into
runOnMachineFunction to keep the flags local.
Pass MachineBasicBlock around instead of an iterator. We can get the iterator in
the few places that need it. Allows a range-based outer for loop.
Separate the Atom optimization from the rest of the optimizations. This allows
fixupIncDec to create INC/DEC and still allow Atom to turn it back into LEA
when profitable by its heuristics.
I'd like to improve fixupIncDec to turn LEAs into ADD any time the base or index
register is equal to the destination register. This is profitable regardless of
the various slow flags. But again we would want Atom to be able to undo that.
Re-reland "[Option] Fix PR37006 prefix choice in findNearest"
This was first reviewed in https://reviews.llvm.org/D46776 and
landed in r332299, but got reverted because it broke the PS4
bots.
https://reviews.llvm.org/D50410 fixed this, and then this
change was re-reviewed at https://reviews.llvm.org/D50515 and
relanded in r341329. It got reverted due to causing MSan issues.
However, nobody wrote down the error message and the bot link
is dead, so I'm relanding this to capture the MSan error.
I'll then either fix it, or copy it somewhere and revert if
fixing looks difficult.
Russell Gallop [Tue, 30 Apr 2019 15:35:16 +0000 (15:35 +0000)]
Add llvm-profdata to LLVM_TOOLCHAIN_TOOLS
This is required for using PGO on Windows but isn't in the Windows
release packages. Windows packages are built with
LLVM_INSTALL_TOOLCHAIN_ONLY so only includes llvm "tools" listed here.
We don't have this restriction in IR, so it should not be here
either simply out of consistency. Code that wants to handle FP
exceptions is expected to use the 'strict' variants of these
nodes.
We don't get the frem case because frem by 0.0 produces NaN (invalid),
and that's the remaining check here (so the removed check for frem
was dead code AFAIK).
This is the only place in SDAG that uses "HasFPExceptions", so I
think we should remove that entirely as a follow-up patch.
Russell Gallop [Tue, 30 Apr 2019 12:37:10 +0000 (12:37 +0000)]
[cmake] Fix error message on simulated VS version check
For clang-cl self hosts in VS2015 environment this was reporting: "Host
Clang must have at least -fms-compatibility-version=19.00.24213.1, your
version is 9.0.0".
This check fires as CMake detects the simulated environment as _MSC_VER
1900, which is truncated. This makes it less than the required
19.00.24213.1.
Simon Pilgrim [Tue, 30 Apr 2019 11:03:09 +0000 (11:03 +0000)]
[SLP] Lit test that cannot get vectorized due to lack of look-ahead operand reordering heuristic.
The code in this test is not vectorized by SLP because its operand reordering cannot look beyond the immediate predecessors.
This will get fixed in a follow-up patch that introduces the look-ahead operand reordering heuristic.
Committed on behalf of @vporpo (Vasileios Porpodas)
George Rimar [Tue, 30 Apr 2019 11:02:09 +0000 (11:02 +0000)]
[llvm-objcopy] - Check dynamic relocation sections for broken references.
This is a fix for https://bugs.llvm.org/show_bug.cgi?id=41371.
Currently, it is possible to break the sh_link field of the dynamic relocation section
by removing the section it refers to. The patch fixes an issue and adds 2 test cases.
One of them shows that it does not seem possible to break the sh_info field.
I added an assert to verify this.
Simon Pilgrim [Tue, 30 Apr 2019 10:18:25 +0000 (10:18 +0000)]
Fix for bug 41512: lower INSERT_VECTOR_ELT(ZeroVec, 0, Elt) to SCALAR_TO_VECTOR(Elt) for all SSE flavors
Current LLVM uses pxor+pinsrb on SSE4+ for INSERT_VECTOR_ELT(ZeroVec, 0, Elt) insead of much simpler movd.
INSERT_VECTOR_ELT(ZeroVec, 0, Elt) is idiomatic construct which is used e.g. for _mm_cvtsi32_si128(Elt) and for lowest element initialization in _mm_set_epi32.
So such inefficient lowering leads to significant performance digradations in ceratin cases switching from SSSE3 to SSE4.
https://bugs.llvm.org/show_bug.cgi?id=41512
Here INSERT_VECTOR_ELT(ZeroVec, 0, Elt) is simply converted to SCALAR_TO_VECTOR(Elt) when applicable since latter is closer match to desired behavior and always efficiently lowered to movd and alike.
This was a local static funtion in SelectionDAG, which I've promoted to
TargetLowering so that I can reuse it to estimate the cost of a memory
operation in D59787.
[ARM GlobalISel] Be more careful about bailing out
Bail out on function arguments/returns with types aggregating an
unsupported type. This fixes cases where we would happily and
incorrectly lower functions taking e.g. [1 x i64] parameters, when we
don't even support plain i64 yet.
[TargetLowering] Change getOptimalMemOpType to take a function attribute list
The MachineFunction wasn't used in getOptimalMemOpType, but more importantly,
this allows reuse of findOptimalMemOpLowering that is calling getOptimalMemOpType.
This is the groundwork for the changes in D59766 and D59787, that allows
implementation of TTI::getMemcpyCost.
Summary:
When a variable goes into scope several times within a single function
or when two variables from different scopes share a stack slot it may
be incorrect to poison such scoped locals at the beginning of the
function.
In the former case it may lead to false negatives (see
https://github.com/google/sanitizers/issues/590), in the latter - to
incorrect reports (because only one origin remains on the stack).
If Clang emits lifetime intrinsics for such scoped variables we insert
code poisoning them after each call to llvm.lifetime.start().
If for a certain intrinsic we fail to find a corresponding alloca, we
fall back to poisoning allocas for the whole function, as it's now
impossible to tell which alloca was missed.
The new instrumentation may slow down hot loops containing local
variables with lifetime intrinsics, so we allow disabling it with
-mllvm -msan-handle-lifetime-intrinsics=false.
Markus Lavin [Tue, 30 Apr 2019 07:58:57 +0000 (07:58 +0000)]
[DebugInfo] DW_OP_deref_size in PrologEpilogInserter.
The PrologEpilogInserter need to insert a DW_OP_deref_size before
prepending a memory location expression to an already implicit
expression to avoid having the existing expression act on the memory
address instead of the value behind it.
The reason for using DW_OP_deref_size and not plain DW_OP_deref is that
big-endian targets need to read the right size as simply truncating a
larger read would yield the wrong result (LSB bytes are not at the lower
address).
This re-commit fixes issues reported in the first one. Namely deref was
inserted under wrong conditions and additionally the deref_size argument
was incorrectly encoded.
Kang Zhang [Tue, 30 Apr 2019 03:39:05 +0000 (03:39 +0000)]
[NFC][PowerPC] Use -check-prefixes to simplify the check in code-align.ll
Summary:
When checking the same output, we can use the `-check-prefixes` to simplify the check.
For example, if we want to check below output.
```
; GENERIC-LABEL: .globl foo
; BASIC-LABEL: .globl foo
; PWR-LABEL: .globl foo
; GENERIC: .p2align 2
; BASIC: .p2align 4
; PWR: .p2align 4
; GENERIC: @foo
; BASIC: @foo
; PWR: @foo
```
If we use `-check-prefixes`
```
... -check-prefixes=CHECK,GENERAL
... -check-prefixes=CHECK,BASIC
... -check-prefixes=CHECK,PWR
```
Above check can be simplify to:
```
; CHECK-LABEL: .globl foo
; GENERIC: .p2align 2
; BASIC: .p2align 4
; PWR: .p2align 4
; CHECK: @foo
```
Zi Xuan Wu [Tue, 30 Apr 2019 03:01:14 +0000 (03:01 +0000)]
[DAGCombiner] Do not generate ISD::ADDE node if adde is not legal for the target when combine ISD::TRUNC node
Do not combine (trunc adde(X, Y, Carry)) into (adde trunc(X), trunc(Y), Carry),
if adde is not legal for the target. Even it's at type-legalize phase.
Because adde is special and will not be legalized at operation-legalize phase later.
This fixes: PR40922
https://bugs.llvm.org/show_bug.cgi?id=40922
projects / llvm / log