It leads to some improvements, but also a regression for the simple
case, so it's not clearly a good idea.
test/CodeGen/ARM/vcvt.ll now has test coverage to show the difference.
Ultimately, the right solution is probably to custom-lower fp-to-int
conversions, to something like ARMISD::VCVT_F32_S32 plus a bitcast.
It's hard to do the right thing when the implicit bitcast isn't visible
to DAG transforms.
llvm-svn: 314169
R12 is used for the SwiftError parameter. It is no longer a CSR as it
is used for transfer the SwiftError, and the caller must preserve it if
they need to.
llvm-svn: 314165
As far as I know SUBREG_TO_REG is stating that the upper bits are 0. But if we are just converting the GR32 with no checks, then we have no reason to say the upper bits are 0.
I don't really know how to test this today since I can't find anything that looks that closely at SUBREG_TO_REG. The test changes here seems to be some perturbance of register allocation.
Differential Revision: https://reviews.llvm.org/D38001
llvm-svn: 314152
The 1st attempt at this:
https://reviews.llvm.org/rL314117
was reverted at:
https://reviews.llvm.org/rL314118
because of bot fails for clang tests that were checking optimized IR. That should be fixed with:
https://reviews.llvm.org/rL314144
...so try again.
Original commit message:
The transform to convert an extract-of-a-select-of-vectors was added at:
https://reviews.llvm.org/rL194013
And a question about the validity of this transform was raised in the review:
https://reviews.llvm.org/D1539:
...but not answered AFAICT>
Most of the motivating cases in that patch are now handled by other combines. These are the tests that were added with
the original commit, but they are not regressing even after we remove the transform in this patch.
The diffs we see after removing this transform cause us to avoid increasing the instruction count, so we don't want to do
those transforms as canonicalizations.
The motivation for not turning a vector-select-of-vectors into a scalar operation is shown in PR33301:
https://bugs.llvm.org/show_bug.cgi?id=33301
...in those cases, we'll get vector ops with this patch rather than the vector/scalar mix that we currently see.
Differential Revision: https://reviews.llvm.org/D38006
llvm-svn: 314147
This teach simplifyDemandedBits to handle constant splat vector shifts.
This required changing some uses of getZExtValue to getLimitedValue since we can't rely on legalization using getShiftAmountTy for the shift amount.
I believe there may have been a bug in the ((X << C1) >>u ShAmt) handling where we didn't check if the inner shift was too large. I've fixed that here.
I had to add new patterns to ARM because the zext/sext the patterns were trying to look for got turned into an any_extend with this patch. Happy to split that out too, but not sure how to test without this change.
Differential Revision: https://reviews.llvm.org/D37665
llvm-svn: 314139
Since now SCEV can handle 'urem', an 'urem' is a better canonical form than an 'srem' because it has well-defined behavior
This is a follow up of D34598
Differential Revision: https://reviews.llvm.org/D38072
llvm-svn: 314125
The transform to convert an extract-of-a-select-of-vectors was added at:
rL194013
And a question about the validity of this transform was raised in the review:
https://reviews.llvm.org/D1539:
...but not answered AFAICT>
Most of the motivating cases in that patch are now handled by other combines. These are the tests that were added with
the original commit, but they are not regressing even after we remove the transform in this patch.
The diffs we see after removing this transform cause us to avoid increasing the instruction count, so we don't want to do
those transforms as canonicalizations.
The motivation for not turning a vector-select-of-vectors into a scalar operation is shown in PR33301:
https://bugs.llvm.org/show_bug.cgi?id=33301
...in those cases, we'll get vector ops with this patch rather than the vector/scalar mix that we currently see.
Differential Revision: https://reviews.llvm.org/D38006
llvm-svn: 314117
Summary:
This code iterates the 'Orders' vector in parallel with the DbgValue
list, emitting all DBG_VALUEs that occurred between the last IR order
insertion point and the next insertion point. This assumes the
SDDbgValue list is sorted in IR order, which it usually is. However, it
is not sorted when a node with a debug value is replaced with another
one. When this happens, TransferDbgValues is called, and the new value
is added to the end of the list.
The problem can be solved by stably sorting the list by IR order.
Reviewers: aprantl, Ka-Ka
Reviewed By: aprantl
Subscribers: MatzeB, hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D38197
llvm-svn: 314114
This patch expands the support of lowerInterleavedStore to 8x8i stride 4.
LLVM creates suboptimal shuffle code-gen for AVX2.
In overall, this patch is a specific fix for the pattern (Strid=4 VF=8) and we plan to include more patterns in the future.
The patch goal is to optimize the following sequence:
At the end of the computation, we have xmm2, xmm0, xmm12 and xmm3 holding
each 8 chars:
c0, c1, , c7
m0, m1, , m7
y0, y1, , y7
k0, k1, ., k7
And these need to be transposed/interleaved and stored like so:
c0 m0 y0 k0 c1 m1 y1 k1 c2 m2 y2 k2 c3 m3 y3 k3 ....
Reviewers
DavidKreitzer
Farhana
zvi
igorb
guyblank
RKSimon
Ayal
Differential Revision: https://reviews.llvm.org/D36058
Change-Id: I3cc5c2ca5d6318901c192a4428493b99ef424c32
llvm-svn: 314109
As mentioned in https://reviews.llvm.org/D33718, this simply adds another
pattern to the compare elimination sequence and is committed without a
differential review.
llvm-svn: 314106
This required changing the ISD opcode for these instructions to have the commutable operands first and the addend last. This way tablegen can autogenerate the additional patterns for us.
llvm-svn: 314083
This patch acts as a reverse to combineBitcastvxi1 - bitcasting a scalar integer to a boolean vector and extending it 'in place' to the requested legal type.
Currently this doesn't handle AVX512 at all - but the current mask register approach is lacking for some cases.
Differential Revision: https://reviews.llvm.org/D35320
llvm-svn: 314076
As mentioned in https://reviews.llvm.org/D33718, this simply adds another
pattern to the compare elimination sequence and is committed without a
differential review.
llvm-svn: 314073
We use a v16i32/v16f32 compare instead and truncate the result. We already did this for the unmasked version, but were missing the version with 'and'.
llvm-svn: 314072
The code wasn't yelling at the user when there's a reference
from a DIGlobalVariableExpression. Thanks to Adrian for the
reduced testcase. Fixes PR34672.
llvm-svn: 314069
This is a follow-up from D38181 (r314023). We have to put 64-bit
constants into a register using a separate instruction, so we
should try harder to avoid that.
From what I see, we're not likely to encounter this pattern in the
DAG because the upstream setcc combines from this don't (usually?)
produce this pattern. If we fix that, then this will become more
relevant. Since the cost of handling this case is just loosening
the predicate of the existing fold, we might as well do it now.
llvm-svn: 314064
As mentioned in https://reviews.llvm.org/D33718, this simply adds another
pattern to the compare elimination sequence and is committed without a
differential revision.
llvm-svn: 314062
As mentioned in https://reviews.llvm.org/D33718, this simply adds another
pattern to the compare elimination sequence and is committed without a
differential revision.
llvm-svn: 314060
As mentioned in https://reviews.llvm.org/D33718, this simply adds another
pattern to the compare elimination sequence and is committed without a
differential revision.
llvm-svn: 314055
Fixed suboptimal encoding of instruction memory operand when assembler is used to select 32 bit fixup rather than 8 bit immediate for encoding memory offset value.
Differential Revision: https://reviews.llvm.org/D38117
llvm-svn: 314044
The (non-)obvious win comes from saving 3 bytes by using the 0x83 'and' opcode variant instead of 0x81.
There are also better improvements based on known-bits that allow us to eliminate the mask entirely.
As noted, this could be extended. There are potentially other wins from always shifting first, but doing
that reveals a tangle of problems in other pattern matching. We do this transform generically in
instcombine, but we often have icmp IR that doesn't match that pattern, so we must account for this
in the backend.
Differential Revision: https://reviews.llvm.org/D38181
llvm-svn: 314023
Summary: Conditional returns were not taken into consideration at all. Implement them by turning them into jumps and normal returns. This means there is a slightly higher performance penalty for conditional returns, but this is the best we can do, and it still disturbs little of the rest.
Reviewers: dberris, echristo
Subscribers: sanjoy, nemanjai, hiraditya, kbarton, llvm-commits
Differential Revision: https://reviews.llvm.org/D38102
llvm-svn: 314005
Usually an intrinsic is a simple target instruction, it should have a small latency. A real function call has much larger latency. So handle the intrinsic call in function getInstructionLatency().
Differential Revision: https://reviews.llvm.org/D38104
llvm-svn: 314003
If the two instructions being compared for equivalence have corresponding operands
that are integer constants, then check their values to determine equivalence.
Patch by Suyog Sarda!
llvm-svn: 313993
