Daniel Sanders [Tue, 31 Oct 2017 21:54:52 +0000 (21:54 +0000)]
Revert r317040: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC
The same bots fail but I believe I know what the issue is now. These bots are
missing the const_iterator versions of insert/emplace/etc. that were introduced
in C++11.
Reid Kleckner [Tue, 31 Oct 2017 21:52:15 +0000 (21:52 +0000)]
[codeview] Merge file checksum entries for DIFiles with the same absolute path
Change the map key from DIFile* to the absolute path string. Computing
the absolute path isn't expensive because we already have a map that
caches the full path keyed on DIFile*.
Daniel Sanders [Tue, 31 Oct 2017 21:34:53 +0000 (21:34 +0000)]
Re-commit: [globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC
Multi-instruction emission needs to ensure the the instructions are generated
a depth-first fashion. For example:
(ADDWrr (SUBWrr a, b), c)
needs to emit the SUBWrr before the ADDWrr. However, our walk over
TreePatternNode's is highly context sensitive which makes it difficult to append
BuildMIActions in the order we want. To fix this, we now keep track of the
insertion point as we add actions. This will allow multi-insn emission to insert
BuildMI's in the correct place.
The previous commit failed on the Ubuntu bots using GCC 4.8. These bots didn't
like a call to emplace(). I've replaced it with insert() to see if it's a quirk
of the C++11 support.
Daniel Sanders [Tue, 31 Oct 2017 19:54:05 +0000 (19:54 +0000)]
[globalisel][tablegen] Keep track of the insertion point while adding BuildMIAction's. NFC
Multi-instruction emission needs to ensure the the instructions are generated
a depth-first fashion. For example:
(ADDWrr (SUBWrr a, b), c)
needs to emit the SUBWrr before the ADDWrr. However, our walk over
TreePatternNode's is highly context sensitive which makes it difficult to append
BuildMIActions in the order we want. To fix this, we now keep track of the
insertion point as we add actions. This will allow multi-insn emission to insert
BuildMI's in the correct place.
Daniel Sanders [Tue, 31 Oct 2017 19:09:29 +0000 (19:09 +0000)]
[globalisel][tablegen] Factor out implicit def/use renderers from createAndImportInstructionRenderer(). NFC
Multi-instruction emission will require that we have separate handling for
the defs between the implicitly created temporaries and the rule outputs.
The former require new temporary vregs while the latter should copy existing
operands. Factor out the implicit def/use renderers to minimize the code
duplication when we implement that.
Craig Topper [Tue, 31 Oct 2017 19:03:51 +0000 (19:03 +0000)]
[SimplifyCFG] Use a more generic name for the selects created by SpeculativelyExecuteBB to prevent long names from being created
Currently the selects are created with the names of their inputs concatenated together. It's possible to get cases that chain these selects together resulting in long names due to multiple levels of concatenation. Our internal branch of llvm managed to generate names over 100000 characters in length on a particular test due to an extreme compounding of the names.
This patch changes the name to a generic name that is not dependent on its inputs.
Craig Topper [Tue, 31 Oct 2017 19:03:49 +0000 (19:03 +0000)]
[SimplifyCFG] Regenerate some test cases using update_test_checks.py to prepare for an upcoming commit. NFC
A future commit will change how some of the value names in the IR are generated which causes these tests to break in their current form. The script generates checks with regular expressions so it should be immune.
Daniel Sanders [Tue, 31 Oct 2017 18:50:24 +0000 (18:50 +0000)]
[globalisel][tablegen] Add infrastructure to potentially allow BuildMIAction to choose a mutatable instruction. NFC
Prepare for multiple instruction emission by allowing BuildMIAction to
search for a suitable matcher that will support mutation.
This patch deliberately neglects to add matchers aside from the root to
preserve NFC. That said, it should be noted that until we support mutations
other than just the opcode the chances of finding a non-root instruction
for which canMutate() is true, is essentially zero. Furthermore in the
presence of multi-instruction emission the chances of finding any
instruction for which canMutate() is true is also zero. Nevertheless, we
can't continue to require that all BuildMIAction's consider the root of the match
to be recyclable due to the risk of recycling it twice in the same rule.
Rui Ueyama [Tue, 31 Oct 2017 17:37:20 +0000 (17:37 +0000)]
[Support] Make the default chunk size of raw_fd_ostream to 1 GiB.
Previously, we call write(2) for each 32767 byte chunk. That is not
efficient because Linux can handle much larger write requests.
This patch changes the chunk size on Linux to 1 GiB.
This patch also changes the default chunks size to SSIZE_MAX. I think
that doesn't in practice change this function's behavior on any operating
system because SSIZE_MAX on 64-bit machine is unrealistically large,
and writing 2 GiB (SSIZE_MAX on 32-bit) on a 32-bit machine by a single
call of write(2) is also unrealistic, as the userspace is usually
limited to 2 GiB. That said, it is in general a good thing to do because
a write larger than SSIZE_MAX is implementation-defined in POSIX.
Nico Weber [Tue, 31 Oct 2017 16:39:47 +0000 (16:39 +0000)]
LTOModule::isBitcodeFile() shouldn't assert when returning false.
Fixes a bunch of assert-on-invalid-bitcode regressions after 315483.
Expected<> calls assertIsChecked() in its dtor, and operator bool() only calls
setChecked() if there's no error. So for functions that don't return an error
itself, the Expected<> version needs explicit code to disarm the error that the
ErrorOr<> code didn't need.
Florian Hahn [Tue, 31 Oct 2017 14:06:31 +0000 (14:06 +0000)]
[Reassociate] Remove FIXME from looptest.ll (NFC)
Summary: The loop invariant add (i+j) is reassoicated, I think the FIXME can be removed, because this is what the test case tries to check (AFAIK). I also changed the test to use FileCheck.
This patch adds the --threads option to dsymutil to process
architectures in parallel. The feature is already present in the version
distributed with Xcode, but was not yet upstreamed.
This is NFC as far as the linking behavior is concerned. As threads are
used automatically, the current tests cover the change in
implementation.
Teresa Johnson [Tue, 31 Oct 2017 12:56:09 +0000 (12:56 +0000)]
[ThinLTO] Double bits of module hash used for renaming
Summary:
Use 64 instead of 32 bits of the module hash as the suffix when renaming
after promotion to reduce the likelihood of a collision (which we
observed in a binary when using 32 bits).
Matthew Simpson [Tue, 31 Oct 2017 12:34:02 +0000 (12:34 +0000)]
[InstCombine] Simplify selects that test cmpxchg instructions
If a select instruction tests the returned flag of a cmpxchg instruction and
selects between the returned value of the cmpxchg instruction and its compare
operand, the result of the select will always be equal to its false value.
Ayman Musa [Tue, 31 Oct 2017 11:39:31 +0000 (11:39 +0000)]
Adding a shufflevector and select LLVM IR instructions fuzz tool
Based on similar python tool - utils/shuffle-fuzz.py - this tool extends the ability of it's previous by optionally attaching select instruction to the generated shufflevector instructions.
This was mainly developed to perform exhaustive testing of the X86 AVX512 masked shuffle instructions. But yet it can be used for various other targets.
The general design of the implementation is much modular than the original shuffle_fuzz.py tool, which makes it easier for anyone to extend it further.
David Green [Tue, 31 Oct 2017 10:47:46 +0000 (10:47 +0000)]
[LoopUnroll] Clean up remarks for unroll remainder
The optimisation remarks for loop unrolling with an unrolled remainder looks something like:
test.c:7:18: remark: completely unrolled loop with 3 iterations [-Rpass=loop-unroll]
C[i] += A[i*N+j];
^
test.c:6:9: remark: unrolled loop by a factor of 4 with run-time trip count [-Rpass=loop-unroll]
for(int j = 0; j < N; j++)
^
This removes the first of the two messages.
[AVX512] Adding new patterns for extract_subvector of vXi1
extract subvector of vXi1 from vYi1 is poorly supported by LLVM and most of the time end with an assertion.
This patch fixes this issue by adding new patterns to the TD file.
Max Kazantsev [Tue, 31 Oct 2017 06:19:05 +0000 (06:19 +0000)]
[IRCE][NFC] Rename fields of InductiveRangeCheck
Rename `Offset`, `Scale`, `Length` into `Begin`, `Step`, `End` respectively
to make naming of similar entities for Ranges and Range Checks more
consistent.
Craig Topper [Tue, 31 Oct 2017 06:01:04 +0000 (06:01 +0000)]
[X86] Make AVX512_512_SET0 XMM16-31 lower to 128-bit XOR when AVX512VL is enabled. Use 128-bit VLX instruction when VLX is enabled.
Unfortunately, this weakens our ability to do domain fixing when AVX512DQ is not enabled, but it is consistent with our 256-bit behavior.
Maybe we should add custom handling to domain fixing to allow EVEX integer XOR/AND/OR/ANDN to switch to VEX encoded fp instructions if the high registers aren't being used?
Philip Reames [Tue, 31 Oct 2017 05:16:46 +0000 (05:16 +0000)]
[IndVarSimplify] Simplify code using preheader assumption
As noted in the nice block comment, the previous code didn't actually handle multi-entry loops correctly, it just assumed SCEV didn't analyze such loops. Given SCEV has comments to the contrary, that seems a bit suspect. More importantly, the pass actually requires loopsimplify form which ensures a loop-preheader is available. Remove the excessive generaility and shorten the code greatly.
Note that we do successfully analyze many multi-entry loops, but we do so by converting them to single entry loops. See the added test case.
Max Kazantsev [Tue, 31 Oct 2017 05:07:56 +0000 (05:07 +0000)]
Reapply "[GVN] Prevent LoadPRE from hoisting across instructions that don't pass control flow to successors"
This patch fixes the miscompile that happens when PRE hoists loads across guards and
other instructions that don't always pass control flow to their successors. PRE is now prohibited
to hoist across such instructions because there is no guarantee that the load standing after such
instruction is still valid before such instruction. For example, a load from under a guard may be
invalid before the guard in the following case:
int array[LEN];
...
guard(0 <= index && index < LEN);
use(array[index]);
Philip Reames [Tue, 31 Oct 2017 04:19:06 +0000 (04:19 +0000)]
[SimplifyIndVar] Extract out invariant expression handling
Previously, the code returned early from the *function* when it couldn't find a free expansion, it should be returning from the *transform*. I don't have a test case, noticed this via inspection.
As a follow up, I'm going to revisit the logic in the extract function. I think that essentially the whole helper routine can be replaced with SCEVExpander, but I wanted to do that in a series of separate commits.
Shoaib Meenai [Tue, 31 Oct 2017 01:30:46 +0000 (01:30 +0000)]
[cmake] Make check_linker_flags operate via linker flags
`check_linker_flags` currently sets the *compiler* flags (via
`CMAKE_REQUIRED_FLAGS`), and thus implicitly relies on cmake's default
behavior of passing the compiler flags to the linker. This breaks when
cmake's build rules have been altered to not pollute the link line with
compiler flags (which can be desirable for build cleanliness). Instead,
set `CMAKE_EXE_LINKER_FLAGS` explicitly and use `CMP0056` to ensure the
linker flags are passed along. Additionally, since we're inside a
function, we can just alter the variable directly (as the alteration
will be limited to the scope of the function) rather than saving and
restoring the old value.
Philip Reames [Mon, 30 Oct 2017 23:59:51 +0000 (23:59 +0000)]
[CGP] Fix crash on i96 bit multiply
Issue found by llvm-isel-fuzzer on OSS fuzz, https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=3725
If anyone actually cares about > 64 bit arithmetic, there's a lot more to do in this area. There's a bunch of obviously wrong code in the same function. I don't have the time to fix all of them and am just using this to understand what the workflow for fixing fuzzer cases might look like.
Rui Ueyama [Mon, 30 Oct 2017 21:19:54 +0000 (21:19 +0000)]
Fix -fuse-ld feature detection error.
check_cxx_compiler_flag doesn't seem to try to link a program, so
the existing code doesn't correctly detect the availability of a given
linker. This patch uses check_cxx_source_compiles instead.
due to the construction not accounting for such variable. It led to the general
confusion and prevented setting linker-specific flags inside functions defined
in AddLLVM.cmake.
Davide Italiano [Mon, 30 Oct 2017 20:20:16 +0000 (20:20 +0000)]
[NewGVN] Stop assuming PHI args ordering when looking at phi-of-ops.
It's not guaranteed. There's a bug open to sort them in predecessor
order, but it won't happen anytime soon. In the meanwhile, passes
will have to do an O(#preds) scan. Such is life.
Daniel Neilson [Mon, 30 Oct 2017 19:51:48 +0000 (19:51 +0000)]
Create instruction classes for identifying any atomicity of memory intrinsic. (NFC)
Summary:
For reference, see: http://lists.llvm.org/pipermail/llvm-dev/2017-August/116589.html
This patch fleshes out the instruction class hierarchy with respect to atomic and
non-atomic memory intrinsics. With this change, the relevant part of the class
hierarchy becomes:
This involves some class renaming:
ElementUnorderedAtomicMemCpyInst -> AtomicMemCpyInst
ElementUnorderedAtomicMemMoveInst -> AtomicMemMoveInst
ElementUnorderedAtomicMemSetInst -> AtomicMemSetInst
A script for doing this renaming in downstream trees is included below.
An example of where the Any* classes should be used in LLVM is when reasoning
about the effects of an instruction (ex: aliasing).
---
Script for renaming AtomicMem* classes:
PREFIXES="[<,([:space:]]"
CLASSES="MemIntrinsic|MemTransferInst|MemSetInst|MemMoveInst|MemCpyInst"
SUFFIXES="[;)>,[:space:]]"
Craig Topper [Mon, 30 Oct 2017 14:51:37 +0000 (14:51 +0000)]
[X86] Make sure we don't create locked inc/dec instructions when the carry flag is being used.
Summary:
INC/DEC don't update the carry flag so we need to make sure we don't try to use it.
This patch introduces new X86ISD opcodes for locked INC/DEC. Teaches lowerAtomicArithWithLOCK to emit these nodes if INC/DEC is not slow or the function is being optimized for size. An additional flag is added that allows the INC/DEC to be disabled if the caller determines that the carry flag is being requested.
The test_sub_1_cmp_1_setcc_ugt test is currently showing this bug. The other test case changes are recovering cases that were regressed in r316860.
This should fully fix PR35068 finishing the fix started in r316860.
Yaxun Liu [Mon, 30 Oct 2017 14:30:28 +0000 (14:30 +0000)]
[AMDGPU] Emit metadata for hidden arguments for kernel enqueue
Identifies kernels which performs device side kernel enqueues and emit
metadata for the associated hidden kernel arguments. Such kernels are
marked with calls-enqueue-kernel function attribute by
AMDGPUOpenCLEnqueueKernelLowering pass and later on
hidden kernel arguments metadata HiddenDefaultQueue and
HiddenCompletionAction are emitted for them.
Clement Courbet [Mon, 30 Oct 2017 14:19:33 +0000 (14:19 +0000)]
[CodeGen][ExpandMemcmp] Allow memcmp to expand to vector loads (2).
- Targets that want to support memcmp expansions now return the list of
supported load sizes.
- Expansion codegen does not assume that all power-of-two load sizes
smaller than the max load size are valid. For examples, this is not the
case for x86(32bit)+sse2.
Florian Hahn [Mon, 30 Oct 2017 10:07:42 +0000 (10:07 +0000)]
Recommit r315288: [SCCP] Propagate integer range info for parameters in IPSCCP.
This version of the patch includes a fix addressing a stage2 LTO buildbot
failure and addressed some additional nits.
Original commit message:
This updates the SCCP solver to use of the ValueElement lattice for
parameters, which provides integer range information. The range
information is used to remove unneeded icmp instructions.
For the following function, f() can be optimized to ret i32 2 with
this change
Florian Hahn [Mon, 30 Oct 2017 09:04:18 +0000 (09:04 +0000)]
Recommit r315288: [SCCP] Propagate integer range info for parameters in IPSCCP.
This version of the patch includes a fix addressing a stage2 LTO buildbot
failure and addressed some additional nits.
Original commit message:
This updates the SCCP solver to use of the ValueElement lattice for
parameters, which provides integer range information. The range
information is used to remove unneeded icmp instructions.
For the following function, f() can be optimized to ret i32 2 with
this change
Craig Topper [Mon, 30 Oct 2017 04:39:18 +0000 (04:39 +0000)]
[X86] Rearrange code in X86InstrInfo.cpp to put all the foldMemoryOperandImpl methods together without partial/undef register handling in the middle. NFC
I have a future patch that wants to make use of the one of the partial functions in one of the earlier memory folding methods and the current ordering prevents that.
Sanjay Patel [Sun, 29 Oct 2017 20:49:31 +0000 (20:49 +0000)]
[(new) Pass Manager] instantiate SimplifyCFG with the same options as the old PM
The old PM sets the options of what used to be known as "latesimplifycfg" on the
instantiation after the vectorizers have run, so that's what we'redoing here.
FWIW, there's a later SimplifyCFGPass instantiation in both PMs where we do not
set the "late" options. I'm not sure if that's intentional or not.