Also verify that we never try to set the size of a vreg associated
to a register class.
Report an error when we encounter that in MIR. Fix a testcase that
hit that error and had a size for no reason.
llvm-svn: 276012
We skipped over ReturnInsts which didn't return an argument which would
lead us to incorrectly conclude that an argument returned by another
ReturnInst was 'returned'.
This reverts commit r275756.
This fixes PR28610.
llvm-svn: 276008
Summary:
Currently, InstCombine is already able to fold expressions of the form `logic(cast(A), cast(B))` to the simpler form `cast(logic(A, B))`, where logic designates one of `and`/`or`/`xor`. This transformation is implemented in `foldCastedBitwiseLogic()` in InstCombineAndOrXor.cpp. However, this optimization will not be performed if both `A` and `B` are `icmp` instructions. The decision to preclude casts of `icmp` instructions originates in r48715 in combination with r261707, and can be best understood by the title of the former one:
> Transform (zext (or (icmp), (icmp))) to (or (zext (cimp), (zext icmp))) if at least one of the (zext icmp) can be transformed to eliminate an icmp.
Apparently, it introduced a transformation that is a reverse of the transformation that is done in `foldCastedBitwiseLogic()`. Its purpose is to expose pairs of `zext icmp` that would subsequently be optimized by `transformZExtICmp()` in InstCombineCasts.cpp. Therefore, in order to avoid an endless loop of switching back and forth between these two transformations, the one in `foldCastedBitwiseLogic()` has been restricted to exclude `icmp` instructions which is mirrored in the responsible check:
`if ((!isa<ICmpInst>(Cast0Src) || !isa<ICmpInst>(Cast1Src)) && ...`
This check seems to sort out more cases than necessary because:
- the reverse transformation is obviously done for `or` instructions only
- and also not every `zext icmp` pair is necessarily the result of this reverse transformation
Therefore we now remove this check and replace it by a more finegrained one in `shouldOptimizeCast()` that now rejects only those `logic(zext(icmp), zext(icmp))` that would be able to be optimized by `transformZExtICmp()`, which also avoids the mentioned endless loop. That means we are now able to also simplify expressions of the form `logic(cast(icmp), cast(icmp))` to `cast(logic(icmp, icmp))` (`cast` being an arbitrary `CastInst`).
As an example, consider the following IR snippet
```
%1 = icmp sgt i64 %a, %b
%2 = zext i1 %1 to i8
%3 = icmp slt i64 %a, %c
%4 = zext i1 %3 to i8
%5 = and i8 %2, %4
```
which would now be transformed to
```
%1 = icmp sgt i64 %a, %b
%2 = icmp slt i64 %a, %c
%3 = and i1 %1, %2
%4 = zext i1 %3 to i8
```
This issue became apparent when experimenting with the programming language Julia, which makes use of LLVM. Currently, Julia lowers its `Bool` datatype to LLVM's `i8` (also see https://github.com/JuliaLang/julia/pull/17225). In fact, the above IR example is the lowered form of the Julia snippet `(a > b) & (a < c)`. Like shown above, this may introduce `zext` operations, casting between `i1` and `i8`, which could for example hinder ScalarEvolution and Polly on certain code.
Reviewers: grosser, vtjnash, majnemer
Subscribers: majnemer, llvm-commits
Differential Revision: https://reviews.llvm.org/D22511
Contributed-by: Matthias Reisinger
llvm-svn: 275989
D20859 and D20860 attempted to replace the SSE (V)CVTTPS2DQ and VCVTTPD2DQ truncating conversions with generic IR instead.
It turns out that the behaviour of these intrinsics is different enough from generic IR that this will cause problems, INF/NAN/out of range values are guaranteed to result in a 0x80000000 value - which plays havoc with constant folding which converts them to either zero or UNDEF. This is also an issue with the scalar implementations (which were already generic IR and what I was trying to match).
This patch changes both scalar and packed versions back to using x86-specific builtins.
It also deals with the other scalar conversion cases that are runtime rounding mode dependent and can have similar issues with constant folding.
A companion clang patch is at D22105
Differential Revision: https://reviews.llvm.org/D22106
llvm-svn: 275981
Recommitting after r274347 was reverted. This patch introduces some
classes to refactor the 3 and 4 register Thumb2 multiplication
instruction descriptions, plus improved tests for some of those
instructions.
Differential Revision: https://reviews.llvm.org/D21929
llvm-svn: 275979
The standard local dynamic model for TLS on ARM systems needs two
relocations:
- R_ARM_TLS_LDM32 (module idx)
- R_ARM_TLS_LDO32 (offset of object from origin of module TLS block)
In GNU style assembler we use symbol(tlsldm) and symbol(tlsldo) to
produce these relocations.
llvm-mc for ARM supports symbol(tlsldo) but does not support symbol(tlsldm).
This patch wires up the existing symbol(tlsldm) to R_ARM_TLS_LDM32.
TLS for ARM is defined in Addenda to, and Errata in, the ABI for the
ARM Architecture
Differential Revision: https://reviews.llvm.org/D22461
llvm-svn: 275977
As discussed on PR27654, this patch fixes the triples of a lot of aarch64 tests and enables lit tests on windows
This will hopefully help stop cases where windows developers break the aarch64 target
Differential Revision: https://reviews.llvm.org/D22191
llvm-svn: 275973
Summary:
N32 and N64 follow the standard ELF conventions (.L) whereas O32 uses its own
($).
This fixes the majority of object differences between -fintegrated-as and
-fno-integrated-as.
Reviewers: sdardis
Subscribers: dsanders, sdardis, llvm-commits
Differential Revision: https://reviews.llvm.org/D22412
llvm-svn: 275967
The following condition expression ( a >> n) & 1 is converted to "bt a, n" instruction. It works on all intel targets.
But on AVX-512 it was broken because the expression is modified to (truncate (a >>n) to i1).
I added the new sequence (truncate (a >>n) to i1) to the BT pattern.
Differential Revision: https://reviews.llvm.org/D22354
llvm-svn: 275950
This patch updates MemorySSA's use-optimizing walker to be more
accurate and, in some cases, faster.
Essentially, this changed our core walking algorithm from a
cache-as-you-go DFS to an iteratively expanded DFS, with all of the
caching happening at the end. Said expansion happens when we hit a Phi,
P; we'll try to do the smallest amount of work possible to see if
optimizing above that Phi is legal in the first place. If so, we'll
expand the search to see if we can optimize to the next phi, etc.
An iteratively expanded DFS lets us potentially quit earlier (because we
don't assume that we can optimize above all phis) than our old walker.
Additionally, because we don't cache as we go, we can now optimize above
loops.
As an added bonus, this patch adds a ton of verification (if
EXPENSIVE_CHECKS are enabled), so finding bugs is easier.
Differential Revision: https://reviews.llvm.org/D21777
llvm-svn: 275940
Add a "-j" option to llvm-profdata to control the number of threads used.
Auto-detect NumThreads when it isn't specified, and avoid spawning threads when
they wouldn't be beneficial.
I tested this patch using a raw profile produced by clang (147MB). Here is the
time taken to merge 4 copies together on my laptop:
No thread pool: 112.87s user 5.92s system 97% cpu 2:01.08 total
With 2 threads: 134.99s user 26.54s system 164% cpu 1:33.31 total
Changes since the initial commit:
- When handling odd-length inputs, call ThreadPool::wait() before merging the
last profile. Should fix a race/off-by-one (see r275937).
Differential Revision: https://reviews.llvm.org/D22438
llvm-svn: 275938
For instructions in uniform set, they will not have vector versions so
add them to VecValuesToIgnore.
For induction vars, those only used in uniform instructions or consecutive
ptrs instructions have already been added to VecValuesToIgnore above. For
those induction vars which are only used in uniform instructions or
non-consecutive/non-gather scatter ptr instructions, the related phi and
update will also be added into VecValuesToIgnore set.
The change will make the vector RegUsages estimation less conservative.
Differential Revision: https://reviews.llvm.org/D20474
The recommit fixed the testcase global_alias.ll.
llvm-svn: 275936
This is to help moveSILowerControlFlow to before regalloc.
There are a couple of tradeoffs with this. The complete CFG
is visible to more passes, the loop body avoids an extra copy of m0,
vcc isn't required, and immediate offsets can be shrunk into s_movk_i32.
The disadvantage is the register allocator doesn't understand that
the single lane's vector is dead within the loop body, so an extra
register is used to outlive the loop block when expanding the
VGPR -> m0 loop. This also now results in worse waitcnt insertion
before the loop instead of after for pending operations at the point
of the indexing, but that should be fixed by future improvements to
cross block waitcnt insertion.
v_movreld_b32's operands are now modeled more correctly since vdst
is not a true output. This is kind of a hack to treat vdst as a
use operand. Extra checking is required in the verifier since
I can't seem to get tablegen to emit an implicit operand for a
virtual register.
llvm-svn: 275934
Add a "-j" option to llvm-profdata to control the number of threads
used. Auto-detect NumThreads when it isn't specified, and avoid spawning
threads when they wouldn't be beneficial.
I tested this patch using a raw profile produced by clang (147MB). Here is the
time taken to merge 4 copies together on my laptop:
No thread pool: 112.87s user 5.92s system 97% cpu 2:01.08 total
With 2 threads: 134.99s user 26.54s system 164% cpu 1:33.31 total
Differential Revision: https://reviews.llvm.org/D22438
llvm-svn: 275921
For instructions in uniform set, they will not have vector versions so
add them to VecValuesToIgnore.
For induction vars, those only used in uniform instructions or consecutive
ptrs instructions have already been added to VecValuesToIgnore above. For
those induction vars which are only used in uniform instructions or
non-consecutive/non-gather scatter ptr instructions, the related phi and
update will also be added into VecValuesToIgnore set.
The change will make the vector RegUsages estimation less conservative.
Differential Revision: https://reviews.llvm.org/D20474
llvm-svn: 275912
Taking address of a byval variable in PTX is legal, but currently runs
into miscompilation by ptxas on sm_50+ (NVIDIA issue 1789042).
Work around the issue by enforcing minimum alignment on byval arguments
of device functions.
The change is a no-op on SASS level for sm_3x where ptxas already aligns
local copy by at least 4.
Differential Revision: https://reviews.llvm.org/D22428
llvm-svn: 275893
Summary:
Usually LCSSA survives this transformation, but in some cases (see
attached test) it doesn't: values from the original loop after
separating might be used from the outer loop. Before the transformation
it was the same loop, so LCSSA phis were not required.
This fixes PR28272.
Reviewers: sanjoy, hfinkel, chandlerc
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D21665
llvm-svn: 275891
This is currently only called with GEP users. A direct
alloca would only happen with current typed pointers
for arrays which are a perverse case.
Also fix crashes on 0 x and 1 x arrays.
llvm-svn: 275869
Elsewhere (particularly computeKnownBits) we assume that a global will be
aligned to the value returned by Value::getPointerAlignment. This is used to
boost the alignment on memcpy/memset, so any target-specific request can only
increase that value.
llvm-svn: 275866
DAGTypeLegalizer::CanSkipSoftenFloatOperand should allow
SELECT op code for x86_64 fp128 type for MME targets,
so SoftenFloatOperand does not abort on SELECT op code.
Differential Revision: http://reviews.llvm.org/D21758
llvm-svn: 275818
Summary:
The direct motivation for the port is to ensure that the OptRemarkEmitter
tests work with the new PM.
This remains a function pass because we not only create multiple loops
but could also version the original loop.
In the test I need to invoke opt
with -passes='require<aa>,loop-distribute'. LoopDistribute does not
directly depend on AA however LAA does. LAA uses getCachedResult so
I *think* we need manually pull in 'aa'.
Reviewers: davidxl, silvas
Subscribers: sanjoy, llvm-commits, mzolotukhin
Differential Revision: https://reviews.llvm.org/D22437
llvm-svn: 275811
Schedule a load and its use in the same packet in MISched. Previously,
isResourceAvailable was returning false for dependences in the same
packet, which prevented MISched from packetizing a load and its use in
the same packet for v60.
Patch by Ikhlas Ajbar.
llvm-svn: 275804
This patch corresponds to review:
https://reviews.llvm.org/D21354
We use direct moves for extracting integer elements from vectors. We also use
direct moves when converting integers to FP. When these operations are chained,
we get a direct move out of a VSR followed by a direct move back into a VSR.
These are redundant - all we need to do is line up the element and convert.
llvm-svn: 275796
Add parseToken and compatriot functions to stitch error checks in
straight linear code. As part of this fix some erronous handling of
directives where the EndOfStatement token either was not checked or
Lexed on termination.
Reviewers: rnk, majnemer
Subscribers: llvm-commits
Differential Revision: http://reviews.llvm.org/D22312
llvm-svn: 275795
The machine scheduler needs to account for available resources
more accurately in order to avoid scheduling an instruction that
forces a new packet to be created.
This occurs in two ways: First, an instruction without an available
resource may have a large priority due to other metrics and be
scheduled when there are other instructions with available resources.
Second, an instruction with a non-zero latency may become available
prematurely. In both these cases, we attempt change the priority
in order to allow a better instruction to be scheduled.
Patch by Brendon Cahoon.
llvm-svn: 275793
An instruction may have multiple predecessors that are candidates
for using .cur. However, only one of them can use .cur in the
packet. When this case occurs, we need to make sure that only
one of the dependences gets a 0 latency value.
Patch by Brendon Cahoon.
llvm-svn: 275790
When SelectionDAGISel transforms a node representing an inline asm
block, memory constraint information is not preserved. This can cause
constraints to be broken when a memory offset is of the form:
offset + frame index
when the frame is resolved.
By propagating the constraints all the way to the backend, targets can
enforce memory operands of inline assembly to conform to their constraints.
For MIPSR6, some instructions had their offsets reduced to 9 bits from
16 bits such as ll/sc. This becomes problematic when using inline assembly
to perform atomic operations, as an offset can generated that is too big to
encode in the instruction.
Reviewers: dsanders, vkalintris
Differential Review: https://reviews.llvm.org/D21615
llvm-svn: 275786
Summary:
The work item intrinsics are not available for the shader
calling conventions. And even if we did hook them up most
shader stages haves some extra restrictions on the amount
of available LDS.
Reviewers: tstellarAMD, arsenm
Subscribers: nhaehnle, arsenm, llvm-commits, kzhuravl
Differential Revision: https://reviews.llvm.org/D20728
llvm-svn: 275779
The current logic for handling inline asm operands in DAGToDAGISel interprets
the operands by looking for constants, which should represent the flags
describing the kind of operand we're dealing with (immediate, memory, register
def etc). The operands representing actual data are skipped only if they are
non-const, with the exception of immediate operands which are skipped explicitly
when a flag describing an immediate is found.
The oversight is that memory operands may be const too (e.g. for device drivers
reading a fixed address), so we should explicitly skip the operand following a
flag describing a memory operand. If we don't, we risk interpreting that
constant as a flag, which is definitely not intended.
Fixes PR26038
Differential Revision: https://reviews.llvm.org/D22103
llvm-svn: 275776
At higher optimization levels, we generate the libcall for DIVREM_Ix, which is
fine: aeabi_{u|i}divmod. At -O0 we generate the one for REM_Ix, which is the
default {u}mod{q|h|s|d}i3.
This commit makes sure that we don't generate REM_Ix calls for ABIs that
don't support them (i.e. where we need to use DIVREM_Ix instead). This is
achieved by bailing out of FastISel, which can't handle non-double multi-reg
returns, and letting the legalization infrastructure expand the REM_Ix calls.
It also updates the divmod-eabi.ll test to run under -O0 as well, and adds some
Windows checks to it to make sure we don't break things for it.
Fixes PR27068
Differential Revision: https://reviews.llvm.org/D21926
llvm-svn: 275773
While debugging GVNHoist, I found it confusing that the entries in a
VNtoInsns were not always value numbers. They _usually_ were except for
StoreInst in which case they were a hash of two different value numbers.
This leads to two observations:
- It is more difficult to debug things when the semantic contents of
VNtoInsns changes over time.
- Using a single value number is not much cheaper, the value of
VNtoInsns is a SmallVector.
- It is not immediately clear what the algorithm would do if there were
hash collisions in the StoreInst case.
Using a DenseMap of std::pair sidesteps all of this.
N.B. The changes in the test were due their sensitivity to the
iteration order of VNtoInsns which has changed.
llvm-svn: 275761
Don't make the test/tools/llvm-cov/demangle.test depend on the order in
which symbols are seen, or on the exact formatting llvm-cov emits after
a symbol is printed. This is an attempt to fix a Windows bot failure:
http://lab.llvm.org:8011/builders/clang-x86-win2008-selfhost/builds/9141
I don't know what the root cause of the failure is, or why the
showTemplateInstantiations test doesn't fail in the same way on the
Windows bots. However, this measure can't hurt, and it'll at least get
me on the blamelists again.
llvm-svn: 275758
This reverts also r275029, "Update Clang tests after adding inference for the returned argument attribute"
It broke LTO build. Seems miscompilation.
llvm-svn: 275756
Summary:
To enable profile-guided indirect call promotion in ThinLTO mode, we
simply add call graph edges for each profitable target from the profile
to the summaries, then the summary-guided importing will consider the
callee for importing as usual.
Also we need to enable the indirect call promotion pass creation in the
PassManagerBuilder when PerformThinLTO=true (we are in the ThinLTO
backend), so that the newly imported functions are considered for
promotion in the backends.
The IC promotion profiles refer to callees by GUID, which required
adding GUIDs to the per-module VST in bitcode (and assigning them
valueIds similar to how they are assigned valueIds in the combined
index).
Reviewers: mehdi_amini, xur
Subscribers: mehdi_amini, davidxl, llvm-commits
Differential Revision: http://reviews.llvm.org/D21932
llvm-svn: 275707
This is a partial implementation of a general fold for associative+commutative operators:
(op (cast (op X, C2)), C1) --> (cast (op X, op (C1, C2)))
(op (cast (op X, C2)), C1) --> (op (cast X), op (C1, C2))
There are 7 associative operators and 13 cast types, so this could potentially go a lot further.
Differential Revision: https://reviews.llvm.org/D22421
llvm-svn: 275684
This reverts commit r275042; the initial commit triggered self-hosting failures
on ARM/AArch64. James Molloy identified the problematic backend code, which has
been disabled in r275677. Trying again...
Original commit message:
Let FuncAttrs infer the 'returned' argument attribute
A function can have one argument with the 'returned' attribute, indicating that
the associated argument is always the return value of the function. Add
FuncAttrs inference logic.
llvm-svn: 275678
r275042 reverted function-attribute inference for the 'returned' attribute
because the feature triggered self-hosting failures on ARM and AArch64. James
Molloy determined that the this-return argument forwarding feature, which
directly ties the returned input argument to the returned value, was the cause.
It seems likely that this forwarding code contains, or triggers, a subtle bug.
Disabling for now until we can track that down.
llvm-svn: 275677
This does not schedule any passes besides the ones necessary to
construct and print the machine function. This is useful to test .mir
file reading and printing.
Differential Revision: http://reviews.llvm.org/D22432
llvm-svn: 275664
test/CodeGen/MIR/ARM/ARMLoadStoreDBG.mir is an actual test for the ARM
load store optimization pass and not a test of the mir parser/printer.
It belongs to test/CodeGen/ARM; This also updates the test to use the
new -run-pass llc syntax.
llvm-svn: 275662
Add an option to specify a symbol demangler (as well as options to the
demangler). This can be used to make reports more human-readable.
This option is especially useful in -output-dir mode, since it isn't as
easy to manually pipe reports into a demangler in this mode.
llvm-svn: 275640
Ahmed Bougacha