| Commit message (Collapse) | Author | Age | Files | Lines |
... | |
|
|
|
|
|
|
|
|
|
|
|
| |
constant with sufficient sign bits to fit in vXi32
The gather/scatter instructions can implicitly sign extend the indices. If we're operating on 32-bit data, an v16i64 index can force a v16i32 gather to be split in two since the index needs 2 registers. If we can shrink the index to the i32 we can avoid the split. It should always be safe to shrink the index regardless of the number of elements. We have gather/scatter instructions that can use v2i32 index stored in a v4i32 register with v2i64 data size.
I've limited this to before legalize types to avoid creating a v2i32 after type legalization. We could check for it, but we'd also need testing. I'm also only handling build_vectors with no bitcasts to be sure the truncate will constant fold.
Differential Revision: https://reviews.llvm.org/D68247
llvm-svn: 373408
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
ops."
This seems to be causing some performance regresions that I'm
trying to investigate.
One thing that stands out is that this transform can increase
the live range of the operands of the earlier logic op. This
can be bad for register allocation. If there are two logic
op inputs we should really combine the one that is closest, but
SelectionDAG doesn't have a good way to do that. Maybe we need
to do this as a basic block transform in Machine IR.
llvm-svn: 373401
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
an immediate before calling getImm().
It might be a symbol instead. We can't fold those since we can't
negate them.
Similar for other SUB with immediates.
Fixes PR43529.
llvm-svn: 373397
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
broadcasted to a vector.
Summary:
This adds the ISD opcode and a DAG combine to create it. There are
probably some places where we can directly create it, but I'll
leave that for future work.
This updates all of the isel patterns to look for this new node.
I had to add a few additional isel patterns for aligned extloads
which we should probably fix with a DAG combine or something. This
does mean that the broadcast load folding for avx512 can no
longer match a broadcasted aligned extload.
There's still some work to do here for combining a broadcast of
a broadcast_load. We also need to improve extractelement or
demanded vector elements of a broadcast_load. I'll try to get
those done before I submit this patch.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D68198
llvm-svn: 373349
|
|
|
|
|
|
|
|
|
|
| |
the non-commutable VEX instruction. Use EVEX2VEX override to fix the scalar instructions.
Previously the match was ambiguous and VMAXPS/PD and VMAXCPS/PD
were mapped to the same VEX instruction. But we should keep
the commutableness when change the opcode.
llvm-svn: 373303
|
|
|
|
|
|
|
|
|
|
|
| |
Replace with the MachineFunction. X86 is the only user, and only uses
it for the function. This removes one obstacle from using this in
GlobalISel. The other is the more tolerable EVT argument.
The X86 use of the function seems questionable to me. It checks hasFP,
before frame lowering.
llvm-svn: 373292
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
forming a SELECT.
The i1 scalar would have been type legalized to i8, but that
doesn't guarantee anything about the upper bits. If we're going
to use it as condition we need to make sure the upper bits are 0.
I've special cased ISD::SETCC conditions since that should
guarantee zero upper bits. We could go further and use
computeKnownBits, but we have no tests that would need that.
Fixes PR43507.
llvm-svn: 373246
|
|
|
|
|
|
| |
See https://reviews.llvm.org/D68167
llvm-svn: 373245
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Existing clients are converted to use MachineModuleInfoWrapperPass. The
new interface is for defining a new pass manager API in CodeGen.
Reviewers: fedor.sergeev, philip.pfaffe, chandlerc, arsenm
Reviewed By: arsenm, fedor.sergeev
Differential Revision: https://reviews.llvm.org/D64183
llvm-svn: 373240
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
default handling.
ANY_EXTEND of v8i8 is marked Custom on AVX512 for handling extends
from v8i8. But the type legalization infrastructure will call
ReplaceNodeResults for v8i8 results. We should just defer it the
default handling instead of asserting in the default of the switch.
Fixes PR43509.
llvm-svn: 373234
|
|
|
|
|
|
|
| |
These are all also implemented in avx512_logical_lowering_types
with support for masking.
llvm-svn: 373181
|
|
|
|
|
|
| |
enable the use of vpshufb on the 256-bit halves.
llvm-svn: 373177
|
|
|
|
| |
llvm-svn: 373175
|
|
|
|
|
|
|
|
|
| |
This was added back to allow some performance regressions to be
investigated. The main perf issue was fixed shortly after adding
this back and no other major issues have been reported. So I
think its safe to remove this again.
llvm-svn: 373174
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
There's room from improvement here, but this is a decent
starting point.
There are a few minor regressions in the vector-rotate tests,
where we are now forming a vpternlog from an and before we get
a chance to form it for a bitselect that we were matching
previously. This results in an AND and an ANDN feeding the
vpternlog where previously we just had an AND after the
vpternlog. I think we can probably DAG combine the AND with
the bitselect to get back to similar codegen.
llvm-svn: 373172
|
|
|
|
|
|
| |
into a vpternlog.
llvm-svn: 373157
|
|
|
|
|
|
|
|
|
|
|
|
| |
This allows us to reduce the use count on the condition node before
the match. This enables load folding for that operand without
relying on the peephole pass. This will be improved on for
broadcast load folding in a subsequent commit.
This still requires a bunch of isel patterns for vXi16/vXi8 types
though.
llvm-svn: 373156
|
|
|
|
| |
llvm-svn: 373155
|
|
|
|
|
|
| |
This uses a similar isel pattern as we used for vpcmov with XOP.
llvm-svn: 373154
|
|
|
|
| |
llvm-svn: 373138
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
nodes like most other DAG combines.
Creating new nodes is what we usually do. Have to explicitly
check that we don't update to an existing node and having
to manually manage the worklist is unusual.
We can probably add a helper function to reduce the duplication
of having to check if we should create a gather or scatter, but
I wanted to just get the simple thing done.
llvm-svn: 373137
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
another version for X86 specific nodes.
The majority of the code doesn't run on the X86 nodes today since
its gated by isBeforeLegalizeOps and we don't formm X86 nodes
until after that. Except for a couple special case in type
legalization. But I think we would probably break those if
some of the transforms fire on them.
I want to remove the hardcoded operand numbers and the unusual
use of UpdateNodeOperands. Being able to know which ISD opcodes
are present should help with that.
llvm-svn: 373136
|
|
|
|
|
|
|
|
|
| |
element is wider than i1, not just when AVX512 is disabled.
The AVX2 intrinsics can still be used when AVX512 is enabled and
those go through this path. So we should simplify them.
llvm-svn: 373108
|
|
|
|
| |
llvm-svn: 373081
|
|
|
|
|
|
| |
patterns and add COPY_TO_REGCLASS to them.
llvm-svn: 373031
|
|
|
|
| |
llvm-svn: 373026
|
|
|
|
|
|
|
|
|
| |
We have isel patterns that can put an IMPLICIT_DEF on one of
the sources for these instructions. So we should make sure
we break any dependencies there. This should be done by
just using one of the other sources.
llvm-svn: 373025
|
|
|
|
|
|
|
|
|
|
|
|
| |
addr), fp32imm0) to use masked MOVSS from memory.
Similar for f64 and having a non-zero passthru value.
We were previously not trying to fold the load at all. Using
a CodeGenOnly instruction allows us to use FR32X/FR64X as the
register class to avoid a bunch of COPY_TO_REGCLASS.
llvm-svn: 373021
|
|
|
|
|
|
|
|
| |
SLM is 2 x slower for <2 x i64> comparison ops than other vector types, we should account for this like we do for SLM <2 x i64> add/sub/mul costs.
This should remove some of the SLM codegen diffs in D43582
llvm-svn: 372954
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
form instead.
This matches what's done for VRNDSCALE and most other instructions.
This mainly determines which instruction will be preferred by
disassembler and assembly parser. The printing and encoding
information is the same.
We prefer the _Int form since it uses the VR128 class due to
intrinsic interface. For some of EVEX features like embedded
rounding, we only select from intrinsics today. So there is
only a VR128 version. So making the VR128 version the preferred
is overally consistent.
llvm-svn: 372947
|
|
|
|
|
|
|
|
| |
folding of the other operand.
The SSE and VEX versions are already correct.
llvm-svn: 372941
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
This pass is only concerned with ZMM0-15 and YMM0-15. For YMM
we use VR256 which only contains YMM0-15, but for ZMM we were
using VR512 which contains ZMM0-31. Using VR512_0_15 is more
correct.
Given that the ABI and register allocator will use registers in
order, its unlikely that register from 16-31 would be used
without also using 0-15. So this probably doesn't functionally
matter.
llvm-svn: 372933
|
|
|
|
|
|
|
|
|
|
| |
Neither the base implementation of findCommutedOpIndices nor any in-tree target modifies the instruction passed in and there is no reason why they would in the future.
Committed on behalf of @hvdijk (Harald van Dijk)
Differential Revision: https://reviews.llvm.org/D66138
llvm-svn: 372882
|
|
|
|
| |
llvm-svn: 372770
|
|
|
|
|
|
|
|
|
|
| |
to a target hook (PR42863)
Reason: this caused severe compile time regressions in JAX.
See email thread of original revision on llvm-commits for details:
http://lists.llvm.org/pipermail/llvm-commits/Week-of-Mon-20190923/697042.html
llvm-svn: 372756
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Summary:
The functions different in two ways:
- getLLVMRegNum could return both "eh" and "other" dwarf register
numbers, while getLLVMRegNumFromEH only returned the "eh" number.
- getLLVMRegNum asserted if the register was not found, while the second
function returned -1.
The second distinction was pretty important, but it was very hard to
infer that from the function name. Aditionally, for the use case of
dumping dwarf expressions, we needed a function which can work with both
kinds of number, but does not assert.
This patch solves both of these issues by merging the two functions into
one, returning an Optional<unsigned> value. While the same thing could
be achieved by adding an "IsEH" argument to the (renamed)
getLLVMRegNumFromEH function, it seemed better to avoid the confusion of
two functions and put the choice of asserting into the hands of the
caller -- if he checks the Optional value, he can safely process
"untrusted" input, and if he blindly dereferences the Optional, he gets
the assertion.
I've updated all call sites to the new API, choosing between the two
options according to the function they were calling originally, except
that I've updated the usage in DWARFExpression.cpp to use the "safe"
method instead, and added a test case which would have previously
triggered an assertion failure when processing (incorrect?) dwarf
expressions.
Reviewers: dsanders, arsenm, JDevlieghere
Subscribers: wdng, aprantl, javed.absar, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67154
llvm-svn: 372710
|
|
|
|
|
|
|
|
|
|
| |
This removes the need for ConvertToTarget opcodes in the isel table.
It's also consistent with the recent changes to use TargetConstant
for intrinsic nodes that always take immediates.
Differential Revision: https://reviews.llvm.org/D67902
llvm-svn: 372645
|
|
|
|
|
|
| |
https://bugs.llvm.org/show_bug.cgi?id=43402
llvm-svn: 372606
|
|
|
|
| |
llvm-svn: 372544
|
|
|
|
|
|
|
|
|
|
|
|
| |
prevent an infinite loop.
The attached test case would previous infinite loop after
r365711.
I'm going to move this to X86ISelDAGToDAG.cpp to get the setcc
to match VPTEST in 32-bit mode in a follow up commit.
llvm-svn: 372543
|
|
|
|
|
|
|
|
|
|
|
| |
When AVX512 is available and the preferred vector width is 512-bits or
more, we should prefer AVX512 for memcpy().
https://bugs.llvm.org/show_bug.cgi?id=43240
https://reviews.llvm.org/D67874
llvm-svn: 372540
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
TargetConstant during lowering.
This allows us to use timm in the isel table which is more
consistent with other intrinsics that take an immediate now.
We can't declare the intrinsic as taking an ImmArg because we
need to match non-constants to the shift by MMX register
instruction which we do by mutating the intrinsic id during
lowering.
llvm-svn: 372537
|
|
|
|
|
|
|
| |
This goes back to when MMX was migrated to intrinsic only. The
hack referenced here has been gone for quite a while.
llvm-svn: 372536
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
intrinsics into the X86 backend.
This intrinsics should be shift by immediate, but gcc allows any
i32 scalar and clang needs to match that. So we try to detect the
non-constant case and move the data from an integer register to an
MMX register.
Previously this was done by creating a v2i32 build_vector and
bitcast in SelectionDAGBuilder. This had to be done early since
v2i32 isn't a legal type. The bitcast+build_vector would be DAG
combined to X86ISD::MMX_MOVW2D which isel will turn into a
GPR->MMX MOVD.
This commit just moves the whole thing to lowering and emits
the X86ISD::MMX_MOVW2D directly to avoid the illegal type. The
test changes just seem to be due to nodes being linearized in a
different order.
llvm-svn: 372535
|
|
|
|
|
|
|
|
| |
ImmArg to the llvm intrinsics.
Update the isel patterns to use timm instead of imm.
llvm-svn: 372534
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
to BZHI is profitable (PR43381)
Summary:
PR43381 notes that while we are good at matching `(X >> C1) & C2` as BEXTR/BEXTRI,
we only do that if we either have BEXTRI (TBM),
or if BEXTR is marked as being fast (`-mattr=+fast-bextr`).
In all other cases we don't match.
But that is mainly only true for AMD CPU's.
However, for all the CPU's for which we have sched models,
the BZHI is always fast (or the sched models are all bad.)
So if we decide that it's unprofitable to emit BEXTR/BEXTRI,
we should consider falling-back to BZHI if it is available,
and follow-up with the shift.
While it's really tempting to do something because it's cool
it is wise to first think whether it actually makes sense to do.
We shouldn't just use BZHI because we can, but only it it is beneficial.
In particular, it isn't really worth it if the input is a register,
mask is small, or we can fold a load.
But it is worth it if the mask does not fit into 32-bits.
(careful, i don't know much about intel cpu's, my choice of `-mcpu` may be bad here)
Thus we manage to fold a load:
https://godbolt.org/z/Er0OQz
Or if we'd end up using BZHI anyways because the mask is large:
https://godbolt.org/z/dBJ_5h
But this isn'r actually profitable in general case,
e.g. here we'd increase microop count
(the register renaming is free, mca does not model that there it seems)
https://godbolt.org/z/k6wFoz
Likewise, not worth it if we just get load folding:
https://godbolt.org/z/1M1deG
https://bugs.llvm.org/show_bug.cgi?id=43381
Reviewers: RKSimon, craig.topper, davezarzycki, spatel
Reviewed By: craig.topper, davezarzycki
Subscribers: andreadb, hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D67875
llvm-svn: 372532
|
|
|
|
|
|
|
| |
This seems to have completed omitted any check for the opcode
of the operand in the isel table.
llvm-svn: 372526
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
ConvertToTarget opcodes from the X86 isel table.
We're now using a lot more TargetConstant nodes in SelectionDAG.
But we were still telling isel to convert some of them
to TargetConstants even though they already are. This is because
isel emits a conversion anytime the output pattern has a an 'imm'.
I guess for patterns in instructions we take the 'timm' from the
'set' pattern, but for Pat patterns with explcicit output we
previously had to say 'imm' since 'timm' wasn't allowed in outputs.
llvm-svn: 372525
|
|
|
|
|
|
|
| |
We need to match TargetConstant, not Constant. This was broken
in r372338, but we lacked test coverage.
llvm-svn: 372523
|
|
|
|
|
|
| |
The AVX512 cases still need some work to correct recognise the PMOV truncation cases.
llvm-svn: 372514
|