This patch removes the last use of the constant pool shuffle decode helper and consistently uses the 'getTargetShuffleMaskIndices' versions instead. The constant pool versions are now purely used for assembly comments.
The avx512vbmi intrinsic upgrades had to be altered as they were being decoded as broadcasts, similar to what I fixed in rL346032. I don't think the change is critical - although its annoying that we lose the {k}{z} instruction test coverage as they are tricky to generate....
Differential Revision: https://reviews.llvm.org/D54083
llvm-svn: 346850
Previously, the extend_vector_inreg opcode required their input register to be the same total width as their output. But this doesn't match up with how the X86 instructions are defined. For X86 the input just needs to be a legal type with at least enough elements to cover the output.
This patch weakens the check on these nodes and allows them to be used as long as they have more input elements than output elements. I haven't changed type legalization behavior so it will still create them with matching input and output sizes.
X86 will custom legalize these nodes by shrinking the input to be a 128 bit vector and once we've done that we treat them as legal operations. We still have one case during type legalization where we must custom handle v64i8 on avx512f targets without avx512bw where v64i8 isn't a legal type. In this case we will custom type legalize to a *extend_vector_inreg with a v16i8 input. After that the input is a legal type so type legalization should ignore the node and doesn't need to know about the relaxed restriction. We are no longer allowed to use the default expansion for these nodes during vector op legalization since the default expansion uses a shuffle which required the widths to match. Custom legalization for all types will prevent us from reaching the default expansion code.
I believe DAG combine works correctly with the released restriction because it doesn't check the number of input elements.
The rest of the patch is changing X86 to use either the vector_inreg nodes or the regular zero_extend/sign_extend nodes. I had to add additional isel patterns to handle any_extend during isel since simplifydemandedbits can create them at any time so we can't legalize to zero_extend before isel. We don't yet create any_extend_vector_inreg in simplifydemandedbits.
Differential Revision: https://reviews.llvm.org/D54346
llvm-svn: 346784
This patch adds the ability to use a PALIGNR to rotate a pair of inputs to select a range containing all the referenced elements, followed by a single input permute to put them in the right location.
Differential Revision: https://reviews.llvm.org/D54267
llvm-svn: 346706
Truncate and shuffle lowering are already capable of matching to PACKUS using known bits analysis.
This features one test change where we now prefer to extend v16i16->v16i32 then trunc v16i32->v16i8 over extract_subvector+packus when avx512f is available, but avx512bw is not.
llvm-svn: 346697
getConstant will create a BUILD_VECTOR for us and use a legal type if necessary. So just create the simple node and let BUILD_VECTOR legalization do the canonicalization.
llvm-svn: 346603
This gives shuffle lowering the freedom to use zero_extend_vector_inreg for the unpckl shuffle. Shuffle combining usually makes this swap later, but not when AVX512 is enabled it seems.
While there also use DAG.getConstant to create a 0 vector instead of using the helper the forces a specific BUILD_VECTOR. I don't think that helper is usually needed. We're basically free to create a constant build_vector anytime and it will be legalized on its own.
llvm-svn: 346574
With avx512f but not avx512bw we need to extend to v16i32 then truncate that to to v16i8. Previously we emitted both nodes during lowering, but I'm trying to switch to using target independent nodes and with that switched the extend+truncate wou
This patch changes the implementation to what will be necessary with that patch which helps minimize test diffs.
llvm-svn: 346552
This makes X86ISD::VSEXT more similar to ISD::SIGN_EXTEND and ISD::ZERO_EXTEND.
I'm hoping to replace X86ISD::VSEXT/VZEXT with target independent nodes. Making the target specific nodes similar to the target independent nodes helps minimize test diffs in that patch.
llvm-svn: 346539
I noticed that we weren't generating broadcasts as much I thought we would with
D54271, and this is part of the problem.
Widening the shuffle elements means adding bitcasts and hiding the relationship
between a splatted scalar and the vector. If we can form a broadcast, do that
before going through the rest of the shuffle lowering because broadcasts should
be cheap and can often be load-folded.
Differential Revision: https://reviews.llvm.org/D54280
llvm-svn: 346498
As discussed in D54073, we have a potential regression from more aggressive vector narrowing here, so let's try to avoid that by changing build-vector lowering slightly.
Insert-vector-element lowering always does this since there's no "pinsr" for ymm/zmm:
// If the vector is wider than 128 bits, extract the 128-bit subvector, insert
// into that, and then insert the subvector back into the result.
...but we can sometimes do better for insert-into-constant-vector by using shuffle lowering.
Differential Revision: https://reviews.llvm.org/D54271
llvm-svn: 346433
The main caller of this already has an MVT and several targets called getSimpleVT inside without checking isSimple. This makes the simpleness explicit.
llvm-svn: 346180
SimplifyDemandedBits can turn a sign_extend back into an any_extend and trigger an infinite loop. So instead legalize it the same way as a sign_extend, but preserve the opcode. Then just pattern match it the same as sign_extend during isel.
I don't have a reduced test case for such an infinite loop yet.
llvm-svn: 346170
v2i8/v2i16/v2i32 are promoted to v2i64. pmuludq takes a v2i64 input and produces a v2i64 output. Since we don't about the upper bits of the type legalized multiply we can use the pmuludq to produce the multiply result for the bits we do care about.
llvm-svn: 346115
Summary: This also enables some constant folding from KnownBits propagation. This helps on some cases vXi64 case in 32-bit mode where constant vectors appear as vXi32 and a bitcast. This can prevent getNode from constant folding sra/shl/srl.
Reviewers: RKSimon, spatel
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D54069
llvm-svn: 346102
These methods were just wrappers around getNode with additional asserts (identical and repeated 3 times). But getNode already has a switch that can be used to hold these asserts that allows them to be shared for all 3 opcodes. This also enables checking on the places that create these nodes without using the wrappers.
The rest of the patch is just changing all callers to use getNode directly.
llvm-svn: 346087
The majority of the changes are because the rest of shuffle lowering/combining prefers to replace the undef input with the other operand. Using UNPCKL directly seemed to avoid this and just grabbed a randomish register for the undef which can create false dependencies.
llvm-svn: 346050
We already have custom lowering for the AVX case in LegalizeVectorOps. So its better to keep the regular extend op around as long as possible.
I had to qualify one place in DAG combine that created illegal vector extending load operations. This change by itself had no effect on any tests which is why its included here.
I've made a few cleanups to the custom lowering. The sign extend code no longer creates an identity shuffle with undef elements. The zero extend code now emits a zero_extend_vector_inreg instead of an unpckl with a zero vector.
For the high half of the custom lowering of zero_extend/any_extend, we're now using an unpckh with a zero vector or undef. Previously we used used a pshufd to move the upper 64-bits to the lower 64-bits and then used a zero_extend_vector_inreg. I think the zero vector should require less execution resources and be smaller code size.
Differential Revision: https://reviews.llvm.org/D54024
llvm-svn: 346043
This patch adds support for expanding vector CTPOP instructions and removes the x86 'bitmath' lowering which replicates the same expansion.
Differential Revision: https://reviews.llvm.org/D53258
llvm-svn: 345869
Reapplying an updated version of rL345395 (reverted in rL345451), now the issues noticed in PR39483 have been fixed.
This patch allows resolveTargetShuffleInputs to remove UNDEF inputs from cases where we have more than 2 inputs.
llvm-svn: 345824
The CONCAT_VECTORS case was using the original mask element count to determine how to adjust the broadcast index. But if we looked through a bitcast the original mask size doesn't tell us anything about the concat_vectors.
This patch switchs to using the concat_vectors input element count directly instead.
Differential Revision: https://reviews.llvm.org/D53823
llvm-svn: 345626
Summary: Previously if we had a bitcast vector output type that needs promotion and a vector input type that needs widening we would just do a stack store and load to handle the conversion. We can do a little better if we can widen the bitcast to a legal vector type the same size as the widened input type. Then we can do the bitcast between this widened type and the widened input type. Afterwards we can extract_subvector back to the original output and any_extend that. Type legalization will then circle back and handle promotion of the extract_subvector and the any_extend will just be removed. This will avoid going through the stack and allows us to remove a custom version of this legalization from X86.
Reviewers: efriedma, RKSimon
Reviewed By: efriedma
Subscribers: javed.absar, llvm-commits
Differential Revision: https://reviews.llvm.org/D53229
llvm-svn: 345567
Use SelectionDAG::EVTToAPFloatSemantics. Make the LogicVT calculation in LowerFABSorFNEG similar to LowerFCOPYSIGN. Use APInt::getSignedMaxValue instead of ~APInt::getSignMask.
llvm-svn: 345565
Add vector support to TargetLowering::expandFP_TO_UINT.
This exposes an issue in X86TargetLowering::LowerVSELECT which was assuming that the select mask was the same width as the LHS/RHS ops - as long as the result is a sign splat we can easily sext/trunk this.
llvm-svn: 345473
Makes no difference to actual shuffle decoding yet, but merges all the existing limits in one place for when proper support is fixed.
........
Its been reported that this is causing out of trunk failures.
llvm-svn: 345451
These promotions add additional bitcasts to the SelectionDAG that can pessimize computeKnownBits/computeNumSignBits. It also seems to interfere with broadcast formation.
This patch removes the promotion and adds isel patterns instead.
The increased table size is more than I would like, but hopefully we can find some canonicalizations or other tricks to start pruning out patterns going forward.
Differential Revision: https://reviews.llvm.org/D53268
llvm-svn: 345408
This is a narrow fix for 1 of the problems mentioned in PR27780:
https://bugs.llvm.org/show_bug.cgi?id=27780
I looked at more general solutions, but it's a mess. We canonicalize shuffle masks
based on the number of elements accessed from each operand, and that's not optional.
If you remove that, we'll crash because we fail to match isel patterns. So I'm
waiting until we're sure that we have blendvb with constant condition and then
commuting based on the load potential. Other cases like blend-with-immediate are
already handled elsewhere, so this is probably not a common problem anyway.
I didn't use "MayFoldLoad" because that checks for one-use and in these cases, we've
screwed that up by creating a temporary PSHUFB using these operands that we're counting
on to be killed later. Undoing that didn't look like a simple task because it's
intertwined with determining if we actually use both operands of the shuffle or not.a
Differential Revision: https://reviews.llvm.org/D53737
llvm-svn: 345390
Multiply a is complex operation so just because some bit of the output isn't used doesn't mean that bit of the input isn't used.
We might able to bound it, but it will require some more thought.
llvm-svn: 345241
Add X86 SimplifyDemandedBitsForTargetNode and use it to simplify PMULDQ/PMULUDQ target nodes.
This enables us to repeatedly simplify the node's arguments after the previous approach had to be reverted due to PR39398.
Differential Revision: https://reviews.llvm.org/D53643
llvm-svn: 345182
When implementing memset's today we often see this pattern:
$x0 = MOV 0xXYXYXYXYXYXYXYXY
store $x0, ...
$w1 = MOV 0xXYXYXYXY
store $w1, ...
We first create a 64bit constant in a 64bit register with all bytes the
same and then create a 32bit constant with all bytes the same in a 32bit
register. In many targets we could just access the lower byte of the
64bit register instead.
- Ideally this would be handled by the ConstantHoist pass but it runs
too early when memset isn't expanded yet.
- The memset expansion code already had this optimization implemented,
however SelectionDAG constantfolding would constantfold the
"trunc(bigconstnat)" pattern to "smallconstant".
- This patch makes the memset expansion mark the constant as Opaque and
stop DAGCombiner from constant folding in this situation. (Similar to
how ConstantHoisting marks things as Opaque to avoid folding
ADD/SUB/etc.)
Differential Revision: https://reviews.llvm.org/D53181
llvm-svn: 345102
We can't add the MULDQ node back to the worklist after the demanded bits change has been committed in case the node has been removed entirely. This will have to wait until we have SimplifyDemandedBitsForTargetNode.
llvm-svn: 345070
Matches the approach taken in the constant pool shuffle decoders, and uses an UndefElts mask instead of uint64_t(-1) raw mask values, which doesn't work safely for i32/i64 shuffle mask sizes (as the -1 value is legal).
This allows us to remove the constant pool shuffle decoders from most of the getTargetShuffleMask variable shuffle cases (X86ISD::VPERMV3 will be handled in a future commit).
llvm-svn: 345018
I've included a fix to DAGCombiner::ForwardStoreValueToDirectLoad that I believe will prevent the previous miscompile.
Original commit message:
Theoretically this was done to simplify the amount of isel patterns that were needed. But it also meant a substantial number of our isel patterns have to match an explicit bitcast. By making the vXi32/vXi16/vXi8 types legal for loads, DAG combiner should be able to change the load type to rem
I had to add some additional plain load instruction patterns and a few other special cases, but overall the isel table has reduced in size by ~12000 bytes. So it looks like this promotion was hurting us more than helping.
I still have one crash in vector-trunc.ll that I'm hoping @RKSimon can help with. It seems to relate to using getTargetConstantFromNode on a load that was shrunk due to an extract_subvector combine after the constant pool entry was created. So we end up decoding more mask elements than the lo
I'm hoping this patch will simplify the number of patterns needed to remove the and/or/xor promotion.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits, RKSimon
Differential Revision: https://reviews.llvm.org/D53306
llvm-svn: 344965
We can't safely assume that certain RawMask entries are UNDEF as most variable shuffles ignore non-index bits - PSHUFB only works on i8 elts so it'd be safe to use but I'm intending to come up with an alternative approach that works for all.
........
Enable this for PSHUFB constant mask decoding and remove the ConstantPool DecodePSHUFBMask
llvm-svn: 344937
We can't safely assume that certain RawMask entries are UNDEF as most variable shuffles ignore non-index bits.
........
Add support for UNDEF raw mask elements and remove the ConstantPool DecodeVPERMILPMask usage in X86ISelLowering.cpp
llvm-svn: 344936
Summary:
As discussed in D52304 / IRC, we now have pattern matching for
'bit extract' in two places - tablegen and `X86DAGToDAGISel`.
There are 4 patterns.
And we will have a problem with `x & (-1 >> (32 - y))` pattern.
* If the mask is one-use, then it is always unfolded into `x << (32 - y) >> (32 - y)` first.
Thus, the existing test coverage is already broken.
* If it is not one-use, then it is not unfolded, and is matched as BZHI.
* If it is not one-use, we will not match it as BEXTR. And if it is one-use, it will have been unfolded already.
So we will either not handle that pattern for BEXTR, or not have test coverage for it.
This is bad.
As discussed with @craig.topper, let's unify this matching, and do everything in `X86DAGToDAGISel`.
Then we will not have code duplication, and will have proper test coverage.
This indeed does not affect any tests, and this is great.
It means that for these two patterns, the `X86DAGToDAGISel` is identical to the tablegen version.
Please review carefully, i'm not fully sure about that intrinsic change, and introduction of the new `X86ISD` opcode.
Reviewers: craig.topper, RKSimon, spatel
Reviewed By: craig.topper
Subscribers: llvm-commits, craig.topper
Differential Revision: https://reviews.llvm.org/D53164
llvm-svn: 344904
Summary:
Theoretically this was done to simplify the amount of isel patterns that were needed. But it also meant a substantial number of our isel patterns have to match an explicit bitcast. By making the vXi32/vXi16/vXi8 types legal for loads, DAG combiner should be able to change the load type to remove the bitcast.
I had to add some additional plain load instruction patterns and a few other special cases, but overall the isel table has reduced in size by ~12000 bytes. So it looks like this promotion was hurting us more than helping.
I still have one crash in vector-trunc.ll that I'm hoping @RKSimon can help with. It seems to relate to using getTargetConstantFromNode on a load that was shrunk due to an extract_subvector combine after the constant pool entry was created. So we end up decoding more mask elements than the load size.
I'm hoping this patch will simplify the number of patterns needed to remove the and/or/xor promotion.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits, RKSimon
Differential Revision: https://reviews.llvm.org/D53306
llvm-svn: 344877
Summary:
These nodes exist to overcome an isel problem where we can generate a zero extend of an AH register followed by an extract subreg, and another zero extend. The first zero extend exists to avoid a partial register update copying the AH register into the low 8-bits. The second zero extend exists if the user wanted the remainder zero extended.
To make this work we had a DAG combine to morph the DIVREM opcode to a special opcode that included the extend. But then we had to add the new node to computeKnownBits and computeNumSignBits to process the extension portion.
This patch instead removes all of that and adds a late peephole to detect the two extends.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D53449
llvm-svn: 344874
D53306 exposes an issue where we sometimes use constant pool data from bigger vectors than the target shuffle mask. This should be safe to do, but we have to be certain that we're using the bottom most part of the vector as the shuffle mask decoders have no way to peek into subvectors with non-zero offsets.
llvm-svn: 344867
Summary:
I've noticed that the bitcasts we introduce for these make computeKnownBits and computeNumSignBits not work well in LegalizeVectorOps. LegalizeVectorOps legalizes bottom up while LegalizeDAG legalizes top down. The bottom up strategy for LegalizeVectorOps means operands are legalized before their uses. So we promote and/or/xor before we legalize the operands that use them making computeKnownBits/computeNumSignBits in places like LowerTruncate suboptimal. I looked at changing LegalizeVectorOps to be top down as well, but that was more disruptive and caused some regressions. I also looked at just moving promotion of binops to LegalizeDAG, but that had a few issues one around matching AND,ANDN,OR into VSELECT because I had to create ANDN as vXi64, but the other nodes hadn't legalized yet, I didn't look too hard at fixing that.
This patch seems to produce better results overall than my other attempts. We now form broadcasts of constants better in some cases. For at least some of them the AND was being introduced in LegalizeDAG, promoted to vXi64, and the BUILD_VECTOR was also legalized there. I think we got bad ordering of that. Now the promotion is out of the legalizer so we handle this better.
In the longer term I think we really should evaluate whether we should be doing this promotion at all. It's really there to reduce isel pattern count, but I'm wondering if we'd be better served just eating the pattern cost or doing C++ based isel for vector and/or/xor in X86ISelDAGToDAG. The masked and/or/xor will definitely be difficult in patterns if a bitcast gets between the vselect and the and/or/xor node. That becomes a lot of permutations to cover.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D53107
llvm-svn: 344487
Summary: This is similar to what D52528 did for loads. It should match what generic type legalization does in 64-bit mode where it uses a v2i64 cast and an i64 store.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D53173
llvm-svn: 344470
There is one remnant - AVX1 custom splitting of 256-bit vectors - which is due to a regression where the X86ISD::ANDNP is still performed as a YMM.
I've also tightened the CTLZ or CTPOP lowering in SelectionDAGLegalize::ExpandBitCount to require a legal CTLZ - it doesn't affect existing users and fixes an issue with AVX512 codegen.
llvm-svn: 344457
Use isConstantSplat instead of ISD::isConstantSplatVector to let us us peek through to illegal types (in this case for i686 targets to recognise i64 constants)
llvm-svn: 344452
If we have better CTLZ support than CTPOP, then use cttz(x) = width - ctlz(~x & (x - 1)) - and remove the CTTZ_ZERO_UNDEF handling as it no longer gives better codegen.
Similar to rL344447, this is also closer to LegalizeDAG's approach
llvm-svn: 344448
This patch changes the vector CTTZ lowering from:
cttz(x) = ctpop((x & -x) - 1)
to:
cttz(x) = ctpop(~x & (x - 1))
Not only does this make better use of the PANDN instruction, but it also matches the LegalizeDAG method which should allow us to remove the x86 specific code at some point in the future (we need to fix some issues with the bitcasted logic ops and CTPOP lowering first).
Differential Revision: https://reviews.llvm.org/D53214
llvm-svn: 344447
Add shuffle lowering for the case where we can shuffle the lanes into place followed by an in-lane permute.
This is mainly for cases where we can have non-repeating permutes in each lane, but for now I've just enabled it for v4f64 unary shuffles to fix PR39161 - there is no test coverage for other shuffles that might benefit yet.
We now have several cross-lane shuffle lowering methods that all do something similar - I've looked at merging some of these (notably by making the repeated mask mechanism in lowerVectorShuffleByMerging128BitLanes optional), but there is a lot of assertions/assumptions in the way that makes this tricky - I ended up going for adding yet another relatively simple method instead.
Differential Revision: https://reviews.llvm.org/D53148
llvm-svn: 344446
Generic legalization should be able to finish legalizing the EXTRACT_SUBVECTOR probably by turning it into a BUILD_VECTOR. But we should emit the simplest sequence.
llvm-svn: 344424
The algorithm we would do previously was identical to generic legalization. If we ever switch to legalizing integer vectors via widening we'll be able to kill off the code since it now only runs for promotion.
llvm-svn: 344423
This is the planned follow-up to D52997. Here we are reducing horizontal vector math codegen
by default. AMD Jaguar (btver2) should have no difference with this patch because it has
fast-hops. (If we want to set that bit for other CPUs, let me know.)
The code changes are small, but there are many test diffs. For files that are specifically
testing for hops, I added RUNs to distinguish fast/slow, so we can see the consequences
side-by-side. For files that are primarily concerned with codegen other than hops, I just
updated the CHECK lines to reflect the new default codegen.
To recap the recent horizontal op story:
1. Before rL343727, we were producing hops for all subtargets for a variety of patterns.
Hops were likely not optimal for all targets though.
2. The IR improvement in r343727 exposed a hole in the backend hop pattern matching, so
we reduced hop codegen for all subtargets. That was bad for Jaguar (PR39195).
3. We restored the hop codegen for all targets with rL344141. Good for Jaguar, but
probably bad for other CPUs.
4. This patch allows us to distinguish when we want to produce hops, so everyone can be
happy. I'm not sure if we have the best predicate here, but the intent is to undo the
extra hop-iness that was enabled by r344141.
Differential Revision: https://reviews.llvm.org/D53095
llvm-svn: 344361
Pull out repeated byte sum stage for popcount of vector elements > 8bits.
This allows us to simplify the LUT/BITMATH popcnt code to always assume vXi8 vectors, and also improves avx512bitalg codegen which only has access to vpopcntb/vpopcntw.
llvm-svn: 344348
Fixes PR32160 by reducing the size of PSHUFB if we only use one of the lanes.
This approach can probably be generalized to handle any target shuffle (and any subvector index) but we have no test coverage at the moment.
llvm-svn: 344336
On 64-bit targets the generic legalize will use an i64 load and a scalar_to_vector for us. But on 32-bit targets i64 isn't legal and the generic legalizer will end up emitting two 32-bit loads. We have DAG combines that try to put those two loads back together with pretty good success.
This patch instead uses f64 to avoid the splitting entirely. I've made it do the same for 64-bit mode for consistency and to keep the load in the fp domain.
There are a few things in here that look like regressions in 32-bit mode, but I believe they bring us closer to the 64-bit mode codegen. And that the 64-bit mode code could be better. I think those issues should be looked at separately.
Differential Revision: https://reviews.llvm.org/D52528
llvm-svn: 344291
This is an alternative to D53080 since I think using a BEXTR for a shifted mask is definitely an improvement when the shl can be absorbed into addressing mode. The other cases I'm less sure about.
We already have several tricks for handling an and of a shift in address matching. This adds a new case for BEXTR.
I've moved the BEXTR matching code back to X86ISelDAGToDAG to allow it to match. I suppose alternatively we could directly emit a X86ISD::BEXTR node that isel could pattern match. But I'm trying to view BEXTR matching as an isel concern so DAG combine can see 'and' and 'shift' operations that are well understood. We did lose a couple cases from tbm_patterns.ll, but I think there are ways to recover that.
I've also put back the manual load folding code in matchBEXTRFromAnd that I removed a few months ago in r324939. This gives us some more freedom to make decisions based on the ability to fold a load. I haven't done anything with that yet.
Differential Revision: https://reviews.llvm.org/D53126
llvm-svn: 344270
Summary:
As discussed in [[ https://bugs.llvm.org/show_bug.cgi?id=38938 | PR38938 ]],
we fail to emit `BEXTR` if the mask is shifted.
We can't deal with that in `X86DAGToDAGISel` `before the address mode for the inc is selected`,
and we can't really do it in the normal DAGCombine, because we don't have generic `ISD::BitFieldExtract` node,
and if we simply turn the shifted mask into a normal mask + shift-left, it will be folded back.
So it would seem X86ISelLowering is the place to handle this.
This patch only moves the matchBEXTRFromAnd()
from X86DAGToDAGISel to X86ISelLowering.
It does not add support for the 'shifted mask' pattern.
Reviewers: RKSimon, craig.topper, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D52426
llvm-svn: 344179
This is intended to restore horizontal codegen to what it looked like before IR demanded elements improved in:
rL343727
As noted in PR39195:
https://bugs.llvm.org/show_bug.cgi?id=39195
...horizontal ops can be worse for performance than a shuffle+regular binop, so I've added a TODO. Ideally, we'd
solve that in a machine instruction pass, but a quicker solution will be adding a 'HasFastHorizontalOp' feature
bit to deal with it here in the DAG.
Differential Revision: https://reviews.llvm.org/D52997
llvm-svn: 344141
Similar to what already happens in the DAGCombiner wrappers, this patch adds the root nodes back onto the worklist if the DCI wrappers' SimplifyDemandedBits/SimplifyDemandedVectorElts were successful.
Differential Revision: https://reviews.llvm.org/D53026
llvm-svn: 344132
This may give slightly better opportunities for DAG combine to simplify with the operations before the setcc. It also matches the type the xors will eventually be promoted to anyway so it saves a legalization step.
Almost all of the test changes are because our constant pool entry is now v2i64 instead of v4i32 on 64-bit targets. On 32-bit targets getConstant should be emitting a v4i32 build_vector and a v4i32->v2i64 bitcast.
There are a couple test cases where it appears we now combine a bitwise not with one of these xors which caused a new constant vector to be generated. This prevented a constant pool entry from being shared. But if that's an issue we're concerned about, it seems we need to address it another way that just relying a bitcast to hide it.
This came about from experiments I've been trying with pushing the promotion of and/or/xor to vXi64 later than LegalizeVectorOps where it is today. We run LegalizeVectorOps in a bottom up order. So the and/or/xor are promoted before their users are legalized. The bitcasts added for the promotion act as a barrier to computeKnownBits if we try to use it during vector legalization of a later operation. So by moving the promotion out we can hopefully get better results from computeKnownBits/computeNumSignBits like in LowerTruncate on AVX512. I've also looked at running LegalizeVectorOps in a top down order like LegalizeDAG, but thats showing some other issues.
llvm-svn: 344071
As noted in D52747, if we prefer IR to use trunc for bool vectors rather
than and+icmp, we can expose codegen shortcomings as seen here with masked store.
Replace a hard-coded PCMPGT simplification with the more general demanded bits call
to improve things.
Differential Revision: https://reviews.llvm.org/D52964
llvm-svn: 344048
As discussed on D52964, this adds 256-bit *_EXTEND_VECTOR_INREG lowering support for AVX1 targets to help improve SimplifyDemandedBits handling.
Differential Revision: https://reviews.llvm.org/D52980
llvm-svn: 344019
Simple types are a superset of what all in tree targets in LLVM could possibly have a legal type. This means the behavior of using isSimple to check for a supported type for X86 could change over time. For example, this could would change if a v256i1 type was added to MVT in the future.
llvm-svn: 343995
Some necessary yak shaving before lowering *_EXTEND_VECTOR_INREG 256-bit vectors on AVX1 targets as suggested by D52964.
Differential Revision: https://reviews.llvm.org/D52970
llvm-svn: 343991
The instructions are complicated, so this code will
probably never be very obvious, but hopefully this
makes it better.
As shown in PR39195:
https://bugs.llvm.org/show_bug.cgi?id=39195
...we need to improve the matching to not miss cases
where we're h-opping on 1 source vector, and that
should be a small patch after this rearranging.
llvm-svn: 343989
Prevents missing other simplifications that may occur deep in the operand chain where CommitTargetLoweringOpt won't add the PMULDQ back to the worklist itself
llvm-svn: 343922
Attempt to simplify PSHUFB masks (even non-constant ones) - we should probably be able to simplify other variable shuffles as well as the need arises.
llvm-svn: 343919
rL343853 didn't limit the number of subinputs, but we don't currently support faux shuffles with more than 2 total inputs, so put a limiter in place until this is fixed.
Found by Artem Dergachev.
llvm-svn: 343891
The comments in this code say we were trying to avoid 16-bit immediates, but if the immediate fits in 8-bits this isn't an issue. This avoids creating a zero extend that probably won't go away.
The movmskb related changes are interesting. The movmskb instruction writes a 32-bit result, but fills the upper bits with 0. So the zero_extend we were previously emitting was free, but we turned a -1 immediate that would fit in 8-bits into a 32-bit immediate so it was still bad.
llvm-svn: 343871
Decode subvector shuffles from INSERT_SUBVECTOR(SRC0, SHUFFLE(EXTRACT_SUBVECTOR(SRC1))
This was found necessary while investigating PR39161
llvm-svn: 343853
Previously we replaced the chain use ourself and return the data result. LegalizeVectorOps then detected that we'd done this and assumed the chain had already been handled.
This commit instead returns a MERGE_VALUES node with two results joined from nodes. This allows LegalizeVectorOps to do all the replacements for us without any special casing. The MERGE_VALUES will be removed by DAG combine.
llvm-svn: 343817
The additional patterns needed for this aren't overwhelming and introducing extra bitcasts during lowering limits our ability to do computeNumSignBits. Not that I have a good example of that for select. I'm just becoming increasingly grumpy about promotion of AND/OR/XOR. SELECT was just a lot easier to fix.
llvm-svn: 343723
Fix use of SSE1 registers for f32 ops in no-x87 mode.
Notably, allow use of SSE instructions for f32 operations in 64-bit
mode (but not 32-bit which is disallowed by callign convention).
Also avoid translating memset/memcopy/memmove into SSE registers
without X87 for 32-bit mode.
This fixes PR38738.
Reviewers: nickdesaulniers, craig.topper
Subscribers: hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D52555
llvm-svn: 343689
By removing demanded target shuffles that simplify to zero/undef/identity before simplifying its inputs we improve chances of further simplification, as only the immediate parent user of the combined is added back to the work list - this still doesn't help us if its passed through other ops though (bitcasts....).
llvm-svn: 343390
The shift amount might have peeked through a extract_subvector, altering the number of vector elements in the 'Amt' variable - so we were incorrectly calculating the ratio when peeking through bitcasts, resulting in incorrectly detecting splats.
llvm-svn: 343373
Similar to the existing ISD::SRL constant vector shifts from D49562, this patch adds ISD::SRA support with ISD::MULHS.
As we're dealing with signed values, we have to handle shift by zero and shift by one special cases, so XOP+AVX2/AVX512 splitting/extension is still a better solution - really we should still use ISD::MULHS if one of the special cases are used but for now I've just left a TODO and filtered by isKnownNeverZero.
Differential Revision: https://reviews.llvm.org/D52171
llvm-svn: 343093
This removes an int->fp bitcast between the surrounding code and the movmsk. I had already added a hack to combineMOVMSK to try to look through this bitcast to improve the SimplifyDemandedBits there.
But I found an additional issue where the bitcast was preventing combineMOVMSK from being called again after earlier nodes in the DAG are optimized. The bitcast gets revisted, but not the user of the bitcast. By using integer types throughout, the bitcast doesn't get in the way.
llvm-svn: 343046
This is the final (I hope!) problem pattern mentioned in PR37749:
https://bugs.llvm.org/show_bug.cgi?id=37749
We are trying to avoid an AVX1 sinkhole caused by having 256-bit bitwise logic ops but no other 256-bit integer ops.
We've already solved the simple logic ops, but 'andn' is an x86 special. I looked at alternative solutions like
extending the generic DAG combine or trying to wait until the ANDNP node is created, but those are bigger patches
that can over-reach. Ie, splitting to 128-bit does not look like a win in most cases with >1 256-bit op.
The pattern matching is cluttered with bitcasts because of our i64 element canonicalization. For the affected test,
we have this vector-type-legalized sequence:
t29: v8i32 = concat_vectors t27, t28
t30: v4i64 = bitcast t29
t18: v8i32 = BUILD_VECTOR Constant:i32<-1>, Constant:i32<-1>, ...
t31: v4i64 = bitcast t18
t32: v4i64 = xor t30, t31
t9: v8i32 = BUILD_VECTOR Constant:i32<255>, Constant:i32<255>, ...
t34: v4i64 = bitcast t9
t35: v4i64 = and t32, t34
t36: v8i32 = bitcast t35
t37: v4i32 = extract_subvector t36, Constant:i64<0>
t38: v4i32 = extract_subvector t36, Constant:i64<4>
Differential Revision: https://reviews.llvm.org/D52318
llvm-svn: 343008
This is an alternative to https://reviews.llvm.org/D37896. We can't decompose
multiplies generically without a target hook to tell us when it's profitable.
ARM and AArch64 may be able to remove some existing code that overlaps with
this transform.
This extends D52195 and may resolve PR34474:
https://bugs.llvm.org/show_bug.cgi?id=34474
(still an open question about transforming legal vector multiplies, but we
could open another bug report for those)
llvm-svn: 342844
We don't have a vXi8 shift left so we need to bitcast to a vXi16 vector to perform the shift. If we let lowering legalize the vXi8 shift we get an extra and that we don't need and fail to remove.
llvm-svn: 342795
x86 had 2 versions of peekThroughBitcast. DAGCombiner had 1. Plus, it had a 1-off implementation for the one-use variant.
Move the x86 versions of the code to SelectionDAG, so we don't have different copies of the code.
No functional change intended.
I'm putting this next to isBitwiseNot() because I am planning to use it in there. Another option is next to the
helpers in the ISD namespace (eg, ISD::isConstantSplatVector()). But if there's no good reason for those to be
there, I'd prefer to pull other helpers over to SelectionDAG in follow-up steps.
Differential Revision: https://reviews.llvm.org/D52285
llvm-svn: 342669
As the code comments suggest, these are about splitting, and they
are not necessarily limited to lowering, so that misled me.
There's nothing that's actually x86-specific in these either, so
they might be better placed in a common header so any target can
use them.
llvm-svn: 342575
This patch adds an initial x86 SimplifyDemandedVectorEltsForTargetNode implementation to handle target shuffles.
Currently the patch only decodes a target shuffle, calls SimplifyDemandedVectorElts on its input operands and removes any shuffle that reduces to undef/zero/identity.
Future work will need to integrate this with combineX86ShufflesRecursively, add support for other x86 ops, etc.
NOTE: There is a minor regression that appears to be affecting further (extractelement?) combines which I haven't been able to solve yet - possibly something to do with how nodes are added to the worklist after simplification.
Differential Revision: https://reviews.llvm.org/D52140
llvm-svn: 342564
This is an alternative to D37896. I don't see a way to decompose multiplies
generically without a target hook to tell us when it's profitable.
ARM and AArch64 may be able to remove some duplicate code that overlaps with
this transform.
As a first step, we're only getting the most clear wins on the vector examples
requested in PR34474:
https://bugs.llvm.org/show_bug.cgi?id=34474
As noted in the code comment, it's likely that the x86 constraints are tighter
than necessary, but it may not always be a win to replace a pmullw/pmulld.
Differential Revision: https://reviews.llvm.org/D52195
llvm-svn: 342554
Summary:
The IR reference for the `byval` attribute states:
```
This indicates that the pointer parameter should really be passed by value
to the function. The attribute implies that a hidden copy of the pointee is
made between the caller and the callee, so the callee is unable to modify
the value in the caller. This attribute is only valid on LLVM pointer arguments.
```
However, on Win64, this attribute is unimplemented and the raw pointer is
passed to the callee instead. This is problematic, because frontend authors
relying on the implicit hidden copy (as happens for every other calling
convention) will see the passed value silently (if mutable memory) or
loudly (by means of a crash) modified because the callee treats the
location as scratch memory space it is allowed to mutate.
At this point, it's worth taking a step back to understand the context.
In most calling conventions, aggregates that are too large to be passed
in registers, instead get *copied* to the stack at a fixed (computable
from the signature) offset of the stack pointer. At the LLVM, we hide
this hidden copy behind the byval attribute. The caller passes a pointer
to the desired data and the callee receives a pointer, but these pointers
are not the same. In particular, the pointer that the callee receives
points to temporary stack memory allocated as part of the call lowering.
In most calling conventions, this pointer is never realized in registers
or memory. The temporary memory is simply defined by an implicit
offset from the stack pointer at function entry.
Win64, uniquely, works differently. The structure is still passed in
memory, but instead of being stored at an implicit memory offset, the
caller computes a pointer to the temporary memory and passes it to
the callee as a regular pointer (taking up a register, or if all
registers are taken up, an additional stack slot). Presumably, this
was done to allow eliding the copy when passing aggregates through
several functions on the stack.
This explains why ignoring the `byval` attribute mostly works on Win64.
The argument simply gets passed as a pointer and as long as we're ok
with the callee trampling all over that memory, there are no ill effects.
However, it does contradict the documentation of the `byval` attribute
which specifies that there is to be an implicit copy.
Frontends can of course work around this by never emitting the `byval`
attribute for Win64 and creating `alloca`s for the requisite temporary
stack slots (and that does appear to be what frontends are doing).
However, the presence of the `byval` attribute is not a trap for
frontend authors, since it seems to work, but silently modifies the
passed memory contrary to documentation.
I see two solutions:
- Disallow the `byval` attribute in the verifier if using the Win64
calling convention.
- Make it work by simply emitting a temporary stack copy as we would
with any other calling convention (frontends can of course always
not use the attribute if they want to elide the copy).
This patch implements the second option (make it work), though I would
be fine with the first also.
Ref: https://github.com/JuliaLang/julia/issues/28338
Reviewers: rnk
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51842
llvm-svn: 342402
For constant non-uniform cases we'll never introduce more and/andn/or selects than already occur in generic pre-SSE41 ISD::SRL lowering.
llvm-svn: 342352
https://bugs.llvm.org/show_bug.cgi?id=38949
It's not clear to me that we even need a one-use check in this fold.
Ie, 2 independent loads might be better than a load+dependent shuffle.
Note that the existing re-use tests are not affected. We actually do form a
broadcast node in those tests now because there's no extra use of the
insert_subvector node in those cases. But something later in isel pattern
matching decides that it is not worth using a broadcast for the full load in
those tests:
Legalized selection DAG: %bb.0 'test_broadcast_2f64_4f64_reuse:'
t7: v2f64,ch = load<(load 16 from %ir.p0)> t0, t2, undef:i64
t4: i64,ch = CopyFromReg t0, Register:i64 %1
t10: ch = store<(store 16 into %ir.p1)> t7:1, t7, t4, undef:i64
t18: v4f64 = insert_subvector undef:v4f64, t7, Constant:i64<0>
t20: v4f64 = insert_subvector t18, t7, Constant:i64<2>
Becomes:
t7: v2f64,ch = load<(load 16 from %ir.p0)> t0, t2, undef:i64
t4: i64,ch = CopyFromReg t0, Register:i64 %1
t10: ch = store<(store 16 into %ir.p1)> t7:1, t7, t4, undef:i64
t21: v4f64 = X86ISD::SUBV_BROADCAST t7
ISEL: Starting selection on root node: t21: v4f64 = X86ISD::SUBV_BROADCAST t7
...
Created node: t27: v4f64 = INSERT_SUBREG IMPLICIT_DEF:v4f64, t7, TargetConstant:i32<7>
Morphed node: t21: v4f64 = VINSERTF128rr t27, t7, TargetConstant:i8<1>
llvm-svn: 342347
Summary: This unfortunately adds a move, but isn't that better than going to the int domain and back?
Reviewers: RKSimon
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D52134
llvm-svn: 342327
Summary:
MOVMSK only care about the sign bit so we don't need the setcc to fill the whole element with 0s/1s. We can just shift the bit we're looking for into the sign bit. This saves a constant pool load.
Inspired by PR38840.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: lebedev.ri, llvm-commits
Differential Revision: https://reviews.llvm.org/D52121
llvm-svn: 342326
Attempt to lower a shuffle as an unpack of elements from two inputs followed by a single-input (wider) permutation.
As long as the permutation is wider this is a win - there may be some circumstances where same size permutations would also be useful but I've left that for future work.
Differential Revision: https://reviews.llvm.org/D52043
llvm-svn: 342257
When replacing a named register input to the appropriately sized
sub/super-register. In the case of a 64-bit value being assigned to a
register in 32-bit mode, match GCC's assignment.
Reviewers: eli.friedman, craig.topper
Subscribers: nickdesaulniers, llvm-commits, hiraditya
Differential Revision: https://reviews.llvm.org/D51502
llvm-svn: 342175
Summary:
Previously we type legalized v2i32 div/rem by promoting to v2i64. But we don't support div/rem of vectors so op legalization would then scalarize it using i64 scalar ops since it doesn't know about the original promotion. 64-bit scalar divides on Intel hardware are known to be slow and in 32-bit mode they require a libcall.
This patch switches type legalization to do the scalarizing itself using i32.
It looks like the division by power of 2 optimization is still kicking in and leaving the code as a vector. The division by other constant optimization doesn't kick in pre type legalization since it ignores illegal types. And previously, after type legalization we scalarized the v2i64 since we don't have v2i64 MULHS/MULHU support.
Another option might be to widen v2i32 to v4i32 so we could do division by constant optimizations, but we'd have to be careful to only do that for constant divisors or we risk scalaring to 4 scalar divides.
Reviewers: RKSimon, spatel
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51325
llvm-svn: 342114
MOVMSKPS and MOVMSKPD both take FP types, but likely the operations before it are on integer types with just a int->fp bitcast between them. If the bitcast isn't used by anything else and doesn't change the element width we can look through it to simplify the integer ops.
llvm-svn: 341915
I'm having a hard time finding a test case for this, but we should be consistent here. The fact that we canonicalize all zeros and all ones constants to vXi32 and all other constants to loads makes this hard to hit the easy DAG combine infinite loop we get for some of the other types.
llvm-svn: 341859
We have isel patterns for v4i32/v4f64 that artificially widen to v8i32/v8f64 so just use that.
If x86-experimental-vector-widening-legalization is enabled, we don't need any custom legalization and can just return. I've modified the test RUN lines to cover this case.
llvm-svn: 341765
Summary:
This patch allows vectors with a power of 2 number of elements and i8/i16 element type to select paddus/psubus instructions. ReplaceNodeResults has been updated to custom widen these operations up to 128 bits like we already do for PAVG.
Another step towards fixing PR38691
Reviewers: RKSimon, spatel
Reviewed By: RKSimon, spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51818
llvm-svn: 341753
The generic type legalizer will scalarize vXi1 instructions getting rid of the vector entirely. Creating wider vector instructions is just going to prevent that.
llvm-svn: 341705
The type legalizer will try to scalarize this and fail.
It looks like there's some other v1iX oddities out there too since we still generated some vector instructions.
llvm-svn: 341704
Similar to what was recently done for addcarry/subborrow and has been done for rdrand/rdseed for a while. It's better to use two results and an explicit store in IR when the store isn't part of the semantics of the instruction. This allows store->load forwarding to happen in the middle end. Or the store to be removed if its never loaded.
Differential Revision: https://reviews.llvm.org/D51803
llvm-svn: 341698
We should represent the store directly in IR instead. This gives the middle end a chance to remove it if it can see a load from the same address.
Differential Revision: https://reviews.llvm.org/D51769
llvm-svn: 341677
These intrinsics use the same implementation as PTEST intrinsics, but use vXi1 vectors.
New clang builtins will be accompanying them shortly.
llvm-svn: 341259
This patch recognizes shuffles that shift elements and fill with zeros. I've copied and modified the shift matching code we use for normal vector registers to do this. I'm not sure if there's a good way to share more of this code without making the existing function more complex than it already is.
This will be used to enable kshift intrinsics in clang.
Differential Revision: https://reviews.llvm.org/D51401
llvm-svn: 341227
This assert tried to check that AND constants are only on the RHS. But its possible for both operands to be constants if one is opaque which will prevent the AND from being constant folded.
Fixes PR38771
llvm-svn: 341102
Noticed while looking at D49562 codegen - we can avoid a large constant mask load and a slow VPBLENDVB select op by using VPBLENDW+VPBLENDD instead.
TODO: As discussed on the patch, we should investigate adding VPBLENDVB handling to target shuffle combining as well, that will allow us to extend this to VPBLENDW+VPBLENDW+VPBLENDD.
Differential Revision: https://reviews.llvm.org/D50074
llvm-svn: 340913
This patch creates the shift mask and actual shift using the vXi16 vector shift ops.
Differential Revision: https://reviews.llvm.org/D51263
llvm-svn: 340813
We're using a 256-bit PACKUS to do the truncation, but that instruction operates on 128-bit lanes. So previously we shuffled first to rearrange the lanes. But that requires 2 shuffles. Instead we can shuffle after the PACKUS using a single VPERMQ. This matches what our normal LowerTRUNCATE code does when it uses PACKUS.
Differential Revision: https://reviews.llvm.org/D51284
llvm-svn: 340757
InstCombine mucks these up a bit. So we need to do some additional pattern matching to fix it. There are a still a few special cases not handled, but this covers the general case.
Differential Revision: https://reviews.llvm.org/D50952
llvm-svn: 340756
vXi32 support was recently moved from LowerMUL_LOHI to LowerMULH.
This commit shares the getOperand calls, switches both to use common IsSigned flag, and hoists the NumElems/NumElts variable.
llvm-svn: 340720
I noticed this along with the patterns in D51125, but when the index is variable,
we don't convert insertelement into a build_vector.
For x86, that means these get expanded at legalization time into the loading/spilling
code that we see in the tests. I think it's always better to avoid going to memory on
these, and we get the optimal 'broadcast' if it's available.
I suspect other targets may want to look at enabling the hook. AArch64 and AMDGPU have
regression tests that would be affected (although I did not check what would happen in
those cases). In the most basic cases shown here, AArch64 would probably do much
better with a splat.
Differential Revision: https://reviews.llvm.org/D51186
llvm-svn: 340705
Summary:
The only time vector SMUL_LOHI/UMUL_LOHI nodes are created is during division/remainder lowering. If its created before op legalization, generic DAGCombine immediately turns that SMUL_LOHI/UMUL_LOHI into a MULHS/MULHU since only the upper half is used. That node will stick around through vector op legalization and will be turned back into UMUL_LOHI/SMUL_LOHI during op legalization. It will then be custom lowered by the X86 backend. Due to this two step lowering the vector shuffles created by the custom lowering get legalized after their inputs rather than before. This prevents the shuffles from being combined with any build_vector of constants.
This patch uses changes vXi32 to use MULHS/MULHU instead. This is what the later DAG combine did anyway. But by skipping the change back to UMUL_LOHI/SMUL_LOHI we lower it before any constant BUILD_VECTORS. This allows the vector_shuffle creation to constant fold with the build_vectors. This accounts for the test changes here.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D51254
llvm-svn: 340690
Summary:
Previously the value being stored is the last operand in SDNode. This causes the type legalizer to visit the mask operand before the value operand. The type legalizer was more complicated because of this since we want the type of the value to drive the decisions.
This patch moves the value to be the first operand so we visit it first during type legalization. It also simplifies the type legalization code accordingly.
X86 is currently the only in tree target that uses this SDNode. Not sure if there are any users out of tree.
Reviewers: RKSimon, delena, hfinkel, eli.friedman
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D50402
llvm-svn: 340689
This is a preliminary step for a preliminary step for D50992.
I noticed that x86 often misses chances to load a scalar directly
into a vector register.
So this patch is just allowing more of those cases to match a
broadcast op in lowerBuildVectorAsBroadcast(). The old code comment
said it doesn't make sense to use a broadcast when we're loading a
single element and everything else is undef, but I think that's the
best case in the improved tests in insert-loaded-scalar.ll. We avoid
scalar-to-vector-register move and/or less efficient shuffling.
Note that there are some existing types that were already producing
a broadcast, but that happens semi-accidentally. Ie, it's not
happening as part of lowerBuildVectorAsBroadcast(). The build vector
gets expanded into load + shuffle, and then shuffle lowering produces
the broadcast.
Description of the other test diffs:
1. avx-basic.ll - replacing load+shufle is a win.
2. sse3-avx-addsub-2.ll - vmovddup vs. vbroadcastss is neutral
3. sse41.ll - don't care - we convert that intrinsic to generic IR now, so this test is deprecated
4. vector-shuffle-128-v8.ll / vector-shuffle-256-v16.ll - pshufb alternatives with an extra instruction are not obviously bad
Differential Revision: https://reviews.llvm.org/D51125
llvm-svn: 340685
subtarget features for indirect calls and indirect branches.
This is in preparation for enabling *only* the call retpolines when
using speculative load hardening.
I've continued to use subtarget features for now as they continue to
seem the best fit given the lack of other retpoline like constructs so
far.
The LLVM side is pretty simple. I'd like to eventually get rid of the
old feature, but not sure what backwards compatibility issues that will
cause.
This does remove the "implies" from requesting an external thunk. This
always seemed somewhat questionable and is now clearly not desirable --
you specify a thunk the same way no matter which set of things are
getting retpolines.
I really want to keep this nicely isolated from end users and just an
LLVM implementation detail, so I've moved the `-mretpoline` flag in
Clang to no longer rely on a specific subtarget feature by that name and
instead to be directly handled. In some ways this is simpler, but in
order to preserve existing behavior I've had to add some fallback code
so that users who relied on merely passing -mretpoline-external-thunk
continue to get the same behavior. We should eventually remove this
I suspect (we have never tested that it works!) but I've not done that
in this patch.
Differential Revision: https://reviews.llvm.org/D51150
llvm-svn: 340515
Previously we asumed a vector reduction add is part of a loop and one of the input is a phi. But the code in SelectionDAGBuilder that sets vector reduction flag handles more cases than that. It just requires that the use chain ends in a horizontal reduction. And there are no other uses. This means it can handle unrolled reduction loops.
If the initial value of the reduction was 0, an unrolled loop would begin with a vector reduction add that has two sad inputs. Previously we would only transform one side of the add, but for this case we need to transform both sides.
I've created a lambda to reuse some of the code for both sides. And fixed the variables names to remove reference to "phi".
Differential Revision: https://reviews.llvm.org/D50817
llvm-svn: 340478
Most of these shifts are extended to vXi16 so we don't gain anything from forcing another round of generic shift lowering - we know these extended cases are legal constant splat shifts.
llvm-svn: 340307
Due to some splat handling code in getVectorShuffle, its possible for NewV1/NewV2 to have their mask modified from what is requested. This can lead to cycles being created in the DAG.
This patch examines the returned mask and makes sure its different. Long term we may need to look closer at that splat code in getVectorShuffle, or add more splat awareness to getVectorShuffle.
Fixes PR38639
Differential Revision: https://reviews.llvm.org/D50981
llvm-svn: 340214
We can safely avoid interfering with the subus combine if both inputs are freely truncatable. Either both extends, or an extend and a constant vector.
Differential Revision: https://reviews.llvm.org/D50878
llvm-svn: 340212
We were basically assuming only one operand of the compare could be an ADD node and using that to swap operands. But we can have a normal add followed by a saturing add.
This rewrites the canonicalization to just be based on the condition code.
llvm-svn: 340134
The code already support 128 and 256 and even knows to split 256 for AVX1. So we really just needed to stop looking for specific VTs and subtarget features and just look for legal VTs with i8/i16 elements.
While there, add some curly braces around outer if statement bodies that contain only another if. It makes all the closing curly braces look more regular.
llvm-svn: 340128
Extending the concept introduced in D49562, this patch lowers constant vXi8 ISD::SRL/ISD::SRA by zero/sign extending to vXi16 and using PMULLW and then truncating the high 8 bits of the result.
Differential Revision: https://reviews.llvm.org/D50781
llvm-svn: 340062
isOnlyUserOf is a little heavier because it allows the node to be used multiple times by the other node. In this case we are looking at a truncate which only has one operand so we know it can only use it once. Thus hasOneUse is better.
llvm-svn: 340059
test/CodeGen/X86/shadow-stack.ll has the following machine verifier
errors:
```
*** Bad machine code: Using a killed virtual register ***
- function: bar
- basic block: %bb.6 entry (0x7fdc81857818)
- instruction: %3:gr64 = MOV64rm killed %2:gr64, 1, $noreg, 8, $noreg
- operand 1: killed %2:gr64
*** Bad machine code: Using a killed virtual register ***
- function: bar
- basic block: %bb.6 entry (0x7fdc81857818)
- instruction: $rsp = MOV64rm killed %2:gr64, 1, $noreg, 16, $noreg
- operand 1: killed %2:gr64
*** Bad machine code: Virtual register killed in block, but needed live out. ***
- function: bar
- basic block: %bb.2 entry (0x7fdc818574f8)
Virtual register %2 is used after the block.
```
The fix here is to only copy the machine operand's register without the
kill flags for all the instructions except the very last one of the
sequence.
I had to insert dummy PHIs in the test case to force the NoPHI function
property to be set to false. More on this here: https://llvm.org/PR38439
Differential Revision: https://reviews.llvm.org/D50260
llvm-svn: 340033
a generically extensible collection of extra info attached to
a `MachineInstr`.
The primary change here is cleaning up the APIs used for setting and
manipulating the `MachineMemOperand` pointer arrays so chat we can
change how they are allocated.
Then we introduce an extra info object that using the trailing object
pattern to attach some number of MMOs but also other extra info. The
design of this is specifically so that this extra info has a fixed
necessary cost (the header tracking what extra info is included) and
everything else can be tail allocated. This pattern works especially
well with a `BumpPtrAllocator` which we use here.
I've also added the basic scaffolding for putting interesting pointers
into this, namely pre- and post-instruction symbols. These aren't used
anywhere yet, they're just there to ensure I've actually gotten the data
structure types correct. I'll flesh out support for these in
a subsequent patch (MIR dumping, parsing, the works).
Finally, I've included an optimization where we store any single pointer
inline in the `MachineInstr` to avoid the allocation overhead. This is
expected to be the overwhelmingly most common case and so should avoid
any memory usage growth due to slightly less clever / dense allocation
when dealing with >1 MMO. This did require several ergonomic
improvements to the `PointerSumType` to reasonably support the various
usage models.
This also has a side effect of freeing up 8 bits within the
`MachineInstr` which could be repurposed for something else.
The suggested direction here came largely from Hal Finkel. I hope it was
worth it. ;] It does hopefully clear a path for subsequent extensions
w/o nearly as much leg work. Lots of thanks to Reid and Justin for
careful reviews and ideas about how to do all of this.
Differential Revision: https://reviews.llvm.org/D50701
llvm-svn: 339940
To lower this we now create a new V1 containing the low half of both sources and a new V2 containing the upper half of both sources. Then we created a repeated lane shuffle of those new sources to create the final result.
This fixes PR35833
Differential Revison: https://reviews.llvm.org/D41794
llvm-svn: 339818
AVX512 added new versions of these intrinsics that take a rounding mode. If the rounding mode is 4 the new intrinsics are equivalent to the old intrinsics.
The AVX512 intrinsics were being lowered to ISD opcodes, but the legacy SSE intrinsics were left as intrinsics. This resulted in the AVX512 instructions needing separate patterns for the ISD opcodes and the legacy SSE intrinsics.
Now we convert SSE intrinsics and AVX512 intrinsics with rounding mode 4 to the same ISD opcode so we can share the isel patterns.
llvm-svn: 339749
rL339686 added the case where a faux shuffle might have repeated shuffle inputs coming from either side of the OR().
This patch improves the insertion of the inputs into the source ops lists to account for this, as well as making it trivial to add support for shuffles with more than 2 inputs in the future.
llvm-svn: 339696
Summary: This revision improves previous version (rL330322) which has been reverted due to crashes.
This is the patch that lowers x86 intrinsics to native IR
in order to enable optimizations. The patch also includes folding
of previously missing saturation patterns so that IR emits the same
machine instructions as the intrinsics.
Reviewers: craig.topper, spatel, RKSimon
Reviewed By: craig.topper
Subscribers: mike.dvoretsky, DavidKreitzer, sroland, llvm-commits
Differential Revision: https://reviews.llvm.org/D46179
llvm-svn: 339650
As discussed on D41794, we have many cases where we fail to combine shuffles as the input operands have other uses.
This patch permits these shuffles to be combined as long as they don't introduce additional variable shuffle masks, which should reduce instruction dependencies and allow the total number of shuffles to still drop without increasing the constant pool.
However, this may mean that some memory folds may no longer occur, and on pre-AVX require the occasional extra register move.
This also exposes some poor PMULDQ/PMULUDQ codegen which was doing unnecessary upper/lower calculations which will in fact fold to zero/undef - the fix will be added in a followup commit.
Differential Revision: https://reviews.llvm.org/D50328
llvm-svn: 339335
Src0 doesn't really convey any meaning to what the operand is. Passthru matches what's used in the documentation for the intrinsic this comes from.
llvm-svn: 339101
Summary:
Expand isFNEG so that we generate the appropriate F(N)M(ADD|SUB)
instructions in more cases. For example, the following sequence
a = _mm256_broadcast_ss(f)
d = _mm256_fnmadd_ps(a, b, c)
generates an fsub and fma without this patch and an fnma with this
change.
Reviewers: craig.topper
Subscribers: llvm-commits, davidxl, wmi
Differential Revision: https://reviews.llvm.org/D48467
llvm-svn: 339043
Clang uses "ctpop & 1" to implement __builtin_parity. If the popcnt instruction isn't supported this generates a large amount of code to calculate the population count. Instead we can bisect the data down to a single byte using xor and then check the parity flag.
Even when popcnt is supported, its still a good idea to split 64-bit data on 32-bit targets using an xor in front of a single popcnt. Otherwise we get two popcnts and an add before the and.
I've specifically targeted this at the sizes supported by clang builtins, but we could generalize this if we think that's useful.
Differential Revision: https://reviews.llvm.org/D50165
llvm-svn: 338907
Move all the patterns to X86InstrVecCompiler.td so we can keep SSE/AVX/AVX512 all in one place.
To save some patterns we'll use an existing DAG combine to convert f128 fand/for/fxor to integer when sse2 is enabled. This allows use to reuse all the existing patterns for v2i64.
I believe this now makes SHA instructions the only case where VEX/EVEX and legacy encoded instructions could be generated simultaneously.
llvm-svn: 338821
We now emit a move of -1 before the cmov and do the addition after the cmov just like the case with an extra addition.
This may be slightly worse for code size, but is more consistent with other compilers. And we might be able to hoist the mov -1 outside of loops.
llvm-svn: 338613
There is nothing x86-specific about this code, so it'd be nice to make this available for other targets to use in the future (and get it out of X86ISelLowering!).
Differential Revision: https://reviews.llvm.org/D50083
llvm-svn: 338586
It's not strictly required by the transform of the cmov and the add, but it makes sure we restrict it to the cases we know we want to match.
While there canonicalize the operand order of the cmov to simplify the matching and emitting code.
llvm-svn: 338492
As was done for vector rotations, we can efficiently use ISD::MULHU for vXi8/vXi16 ISD::SRL lowering.
Shift-by-zero cases are still problematic (mainly on v32i8 due to extra AND/ANDN/OR or VPBLENDVB blend masks but v8i16/v16i16 aren't great either if PBLENDW fails) so I've limited this first patch to known non-zero cases if we can't easily use PBLENDW.
Differential Revision: https://reviews.llvm.org/D49562
llvm-svn: 338407
Summary:
Similar to D49636, but for PMADDUBSW. This instruction has the additional complexity that the addition of the two products saturates to 16-bits rather than wrapping around. And one operand is treated as signed and the other as unsigned.
A C example that triggers this pattern
```
static const int N = 128;
int8_t A[2*N];
uint8_t B[2*N];
int16_t C[N];
void foo() {
for (int i = 0; i != N; ++i)
C[i] = MIN(MAX((int16_t)A[2*i]*(int16_t)B[2*i] + (int16_t)A[2*i+1]*(int16_t)B[2*i+1], -32768), 32767);
}
```
Reviewers: RKSimon, spatel, zvi
Reviewed By: RKSimon, zvi
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D49829
llvm-svn: 338402
isFNEG was duplicating much of what was done by getTargetConstantBitsFromNode in its own calls to getTargetConstantFromNode.
Noticed while reviewing D48467.
llvm-svn: 338358
Not sure why they were being explicitly excluded, but I believe all the math inside the if works. I changed the absolute value to be uint64_t instead of int64_t so INT64_MIN+1 wouldn't be signed wrap.
llvm-svn: 338101
Summary:
This is the pattern you get from the loop vectorizer for something like this
int16_t A[1024];
int16_t B[1024];
int32_t C[512];
void pmaddwd() {
for (int i = 0; i != 512; ++i)
C[i] = (A[2*i]*B[2*i]) + (A[2*i+1]*B[2*i+1]);
}
In this case we will have (add (mul (build_vector), (build_vector)), (mul (build_vector), (build_vector))). This is different than the pattern we currently match which has the build_vectors between an add and a single multiply. I'm not sure what C code would get you that pattern.
Reviewers: RKSimon, spatel, zvi
Reviewed By: zvi
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D49636
llvm-svn: 338097
If this happens the operands aren't updated and the existing node is returned. Make sure we pass this existing node up to the DAG combiner so that a proper replacement happens. Otherwise we get stuck in an infinite loop with an unoptimized node.
llvm-svn: 338090
I'm not sure if this was trying to avoid optimizing the new nodes further or what. Or maybe to prevent a cycle if something tried to reform the multiply? But I don't think its a reliable way to do that. If the user of the expanded multiply is visited by the DAGCombiner after this conversion happens, the DAGCombiner will check its operands, see that they haven't been visited by the DAGCombiner before and it will then add the first node to the worklist. This process will repeat until all the new nodes are visited.
So this seems like an unreliable prevention at best. So this patch just returns the new nodes like any other combine. If this starts causing problems we can try to add target specific nodes or something to more directly prevent optimizations.
Now that we handle the combine normally, we can combine any negates the mul expansion creates into their users since those will be visited now.
llvm-svn: 338007
These calls were making sure some newly created nodes were added to worklist, but the DAGCombiner has internal support for ensuring it has visited all nodes. Any time it visits a node it ensures the operands have been queued to be visited as well. This means if we only need to return the last new node. The DAGCombiner will take care of adding its inputs thus walking backwards through all the new nodes.
llvm-svn: 337996
- Avoid duplication of regmask size calculation.
- Simplify allocateRegisterMask() call.
- Rename allocateRegisterMask() to allocateRegMask() to be consistent
with naming in MachineOperand.
llvm-svn: 337986
We generated a subtract for the power of 2 minus one then negated the result. The negate can be optimized away by swapping the subtract operands, but DAG combine doesn't know how to do that and we don't add any of the new nodes to the worklist anyway.
This patch makes use explicitly emit the swapped subtract.
llvm-svn: 337858
Use a left shift and 2 subtracts like we do for 30. Move this out from behind the slow lea check since it doesn't even use an LEA.
Use this for multiply by 14 as well.
llvm-svn: 337856
This seems to be a net improvement. There's still an issue under avx512f where we have a 512-bit vpaddd, but not vpmaddwd so we end up doing two 256-bit vpmaddwds and inserting the results before a 512-bit vpaddd. It might be better to do two 512-bits paddds with zeros in the upper half. Same number of instructions, but breaks a dependency.
llvm-svn: 337656
Ideally our ISD node types going into the isel table would have types consistent with their instruction domain. This prevents us having to duplicate patterns with different types for the same instruction.
Unfortunately, it seems our shuffle combining is currently relying on this a little remove some bitcasts. This seems to enable some switching between shufps and shufd. Hopefully there's some way we can address this in the combining.
Differential Revision: https://reviews.llvm.org/D49280
llvm-svn: 337590
CombineTo is most useful when you need to replace multiple results, avoid the worklist management, or you need to something else after the combine, etc. Otherwise you should be able to just return the new node and let DAGCombiner go through its usual worklist code.
All of the places changed in this patch look to be standard cases where we should be able to use the more stand behavior of just returning the new node.
Differential Revision: https://reviews.llvm.org/D49569
llvm-svn: 337589
We can safely use getConstant here as we're still lowering, which allows constant folding to kick in and simplify the vector shift codegen.
Noticed while working on D49562.
llvm-svn: 337578
This is an early step towards using SimplifyDemandedVectorElts for target shuffle combining - this merely moves the existing X86ISD::VBROADCAST simplification code to use the SimplifyDemandedVectorElts mechanism.
Adds X86TargetLowering::SimplifyDemandedVectorEltsForTargetNode to handle X86ISD::VBROADCAST - in time we can support all target shuffles (and other ops) here.
llvm-svn: 337547
We have a number of cases where we fail to reduce vector op widths, performing the op in a larger vector and then extracting a subvector. This is often because by default it would create illegal types.
This peephole patch attempts to handle a few common cases detailed in PR36761, which typically involved extension+conversion to vX2f64 types.
Differential Revision: https://reviews.llvm.org/D49556
llvm-svn: 337500
Returning SDValue() means nothing was changed. Returning the result of CombineTo returns the first argument of CombineTo. This is specially detected by DAGCombiner as meaning that something changed, but worklist management was already taken care of.
I think the only real effect of this change is that we now properly update the Statistic the counts the number of combines performed. That's the only thing between the check for null and the check for N in the DAGCombiner.
llvm-svn: 337491
As discussed on PR38197, this canonicalizes MOVS*(N0, OP(N0, N1)) --> MOVS*(N0, SCALAR_TO_VECTOR(OP(N0[0], N1[0])))
This returns the scalar-fp codegen lost by rL336971.
Additionally it handles the OP(N1, N0)) case for commutable (FADD/FMUL) ops.
Differential Revision: https://reviews.llvm.org/D49474
llvm-svn: 337419
When rL336971 removed the scalar-fp isel patterns, we lost the need for this canonicalization - commutation/folding can handle everything else.
llvm-svn: 337387
I'm trying to restrict the MOVLHPS/MOVHLPS ISD nodes to SSE1 only. With SSE2 we can use unpcks. I believe this will allow some patterns to be cleaned up to require fewer bitcasts.
I've put in an odd isel hack to still select MOVHLPS instruction from the unpckh node to avoid changing tests and because movhlps is a shorter encoding. Ideally we'd do execution domain switching on this, but the operands are in the wrong order and are tied. We might be able to try a commute in the domain switching using custom code.
We already support domain switching for UNPCKLPD and MOVLHPS.
llvm-svn: 337348
This unfortunately requires a bunch of bitcasts to be added added to SUBREG_TO_REG, COPY_TO_REGCLASS, and instructions in output patterns. Otherwise tablegen seems to default to picking f128 and then we fail when something tries to get the register class for f128 which isn't always valid.
The test changes are because we were previously mixing fr128 and vr128 due to contrainRegClass finding FR128 first and passes like live range shrinking weren't handling that well.
llvm-svn: 337147
canWidenShuffleElements can do a better job if given a mask with ZeroableElements info. Apparently, ZeroableElements was being only used to identify AllZero candidates, but possibly we could plug it into more shuffle matchers.
Original Patch by Zvi Rackover @zvi
Differential Revision: https://reviews.llvm.org/D42044
llvm-svn: 336903
Noticed while updating D42044, lowerV2X128VectorShuffle can improve the shuffle mask with the zeroable data to create a target shuffle mask to recognise more 'zero upper 128' patterns.
NOTE: lowerV4X128VectorShuffle could benefit as well but the code needs refactoring first to discriminate between SM_SentinelUndef and SM_SentinelZero for negative shuffle indices.
Differential Revision: https://reviews.llvm.org/D49092
llvm-svn: 336900
We now use llvm.fma.f32/f64 or llvm.x86.fmadd.f32/f64 intrinsics that use scalar types rather than vector types. So we don't these special ISD nodes that operate on the lowest element of a vector.
llvm-svn: 336883
This converts them to what clang is now using for codegen. Unfortunately, there seem to be a few kinks to work out still. I'll try to address with follow up patches.
llvm-svn: 336871
Summary:
These changes cover the PR#31399.
Now the ffs(x) function is lowered to (x != 0) ? llvm.cttz(x) + 1 : 0
and it corresponds to the following llvm code:
%cnt = tail call i32 @llvm.cttz.i32(i32 %v, i1 true)
%tobool = icmp eq i32 %v, 0
%.op = add nuw nsw i32 %cnt, 1
%add = select i1 %tobool, i32 0, i32 %.op
and x86 asm code:
bsfl %edi, %ecx
addl $1, %ecx
testl %edi, %edi
movl $0, %eax
cmovnel %ecx, %eax
In this case the 'test' instruction can't be eliminated because
the 'add' instruction modifies the EFLAGS, namely, ZF flag
that is set by the 'bsf' instruction when 'x' is zero.
We now produce the following code:
bsfl %edi, %ecx
movl $-1, %eax
cmovnel %ecx, %eax
addl $1, %eax
Patch by Ivan Kulagin
Reviewers: davide, craig.topper, spatel, RKSimon
Reviewed By: craig.topper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D48765
llvm-svn: 336768
These ISD nodes try to select the MOVLPS and MOVLPD instructions which are special load only instructions. They load data and merge it into the lower 64-bits of an XMM register. They are logically equivalent to our MOVSD node plus a load.
There was only one place in X86ISelLowering that used MOVLPD and no places that selected MOVLPS. The one place that selected MOVLPD had to choose between it and MOVSD based on whether there was a load. But lowering is too early to tell if the load can really be folded. So in isel we have patterns that use MOVSD for MOVLPD if we can't find a load.
We also had patterns that select the MOVLPD instruction for a MOVSD if we can find a load, but didn't choose the MOVLPD ISD opcode for some reason.
So it seems better to just standardize on MOVSD ISD opcode and manage MOVSD vs MOVLPD instruction with isel patterns.
llvm-svn: 336728
Now that rL336250 has landed, we should prefer 2 immediate shifts + a shuffle blend over performing a multiply. Despite the increase in instructions, this is quicker (especially for slow v4i32 multiplies), avoid loads and constant pool usage. It does mean however that we increase register pressure. The code size will go up a little but by less than what we save on the constant pool data.
This patch also adds support for v16i16 to the BLEND(SHIFT(v,c1),SHIFT(v,c2)) combine, and also prevents blending on pre-SSE41 shifts if it would introduce extra blend masks/constant pool usage.
Differential Revision: https://reviews.llvm.org/D48936
llvm-svn: 336642
Summary:
This adds a reverse transform for the instcombine canonicalizations
that were added in D47980, D47981.
As discussed later, that was worse at least for the code size,
and potentially for the performance, too.
https://rise4fun.com/Alive/Zmpl
Reviewers: craig.topper, RKSimon, spatel
Reviewed By: spatel
Subscribers: reames, llvm-commits
Differential Revision: https://reviews.llvm.org/D48768
llvm-svn: 336585
This replaces some asserts in lowerV2F64VectorShuffle with the similar asserts from lowerVIF64VectorShuffle which are more readable. The original asserts mentioned a blend, but there's no guarantee that it is a blend.
Also remove an if that the asserts prove is always true. Mask[0] is always less than 2 and Mask[1] is always at least 2. Therefore (Mask[0] >= 2) + (Mask[1] >= 2) == 1 must wlays be true.
llvm-svn: 336517
Pre-AVX512 (which can perform a quick extend/shift/truncate), extending to 2 v8i16 for the PMULLW and then truncating is more performant than relying on the generic PBLENDVB vXi8 shift path and uses a similar amount of mask constant pool data.
Differential Revision: https://reviews.llvm.org/D48963
llvm-svn: 336513
Splits off isKnownNeverZeroFloat to handle +/- 0 float cases.
This will make it easier to be more aggressive with the integer isKnownNeverZero tests (similar to ValueTracking), use computeKnownBits etc.
Differential Revision: https://reviews.llvm.org/D48969
llvm-svn: 336492
It's a bit neater to write T.isIntOrPtrTy() over `T.isIntegerTy() ||
T.isPointerTy()`.
I used Python's re.sub with this regex to update users:
r'([\w.\->()]+)isIntegerTy\(\)\s*\|\|\s*\1isPointerTy\(\)'
llvm-svn: 336462
The intrinsics can be implemented with a f32/f64 llvm.fma intrinsic and an insert into a zero vector.
There are a couple regressions here due to SelectionDAG not being able to pull an fneg through an extract_vector_elt. I'm not super worried about this though as InstCombine should be able to do it before we get to SelectionDAG.
llvm-svn: 336416
This upgrades all of the intrinsics to use fneg instructions to convert fma into fmsub/fnmsub/fnmadd/fmsubadd. And uses a select instruction for masking.
This matches how clang uses the intrinsics these days.
llvm-svn: 336409
Previously we could only negate the FMADD opcodes. This used to be mostly ok when we lowered FMA intrinsics during lowering. But with the move to llvm.fma from target specific intrinsics, we can combine (fneg (fma)) to (fmsub) earlier. So if we start with (fneg (fma (fneg))) we would get stuck at (fmsub (fneg)).
This patch fixes that so we can also combine things like (fmsub (fneg)).
llvm-svn: 336304
We were only doing this for basic blends, despite shuffle lowering now being good enough to handle more complex blends. This means that the two v8i16 splat shifts are performed in parallel instead of serially as the general shift case.
Reapplied with a fixed (extra null tests) version of rL336113 after reversion in rL336189 - extra test case added at rL336247.
llvm-svn: 336250
We were only doing this for basic blends, despite shuffle lowering now being good enough to handle more complex blends. This means that the two v8i16 splat shifts are performed in parallel instead of serially as the general shift case.
llvm-svn: 336113
I believe all of these are constants so legalizing them should be pretty trivial, but this saves a step.
In one case it looks like we may have been creating a shift amount larger than the shift input itself.
llvm-svn: 336052
This combine runs pretty late and causes us to introduce a shift after the op legalization phase has run. We need to be sure we create the shift with the proper type for the shift amount. If we don't do this, we will still re-legalize the operation properly, but we won't get a chance to fully optimize the truncate that gets inserted.
So this patch adds the necessary truncate when the shift is created. I've also narrowed the subtract that gets created to always be an i32 type. The truncate would have trigered SimplifyDemandedBits to optimize it anyway. But using a more appropriate VT here is free and saves an optimization step.
llvm-svn: 336051
The important part is the creation of the SHLD/SHRD nodes. The compare and the conditional move can use target independent nodes that can be legalized on their own. This gives some opportunities to trigger the optimizations present in the lowering for those things. And its just better to limit the number of places we emit target specific nodes.
The changed test cases still aren't optimal.
Differential Revision: https://reviews.llvm.org/D48619
llvm-svn: 335998
This uses the same technique as for shifts - split the rotation into 4/2/1-bit partial rotations and select those partials based on the amount bit, making use of PBLENDVB if available. This halves the use of PBLENDVB compared to expanding to shifts, which can be a slow op.
Unfortunately I haven't found a decent way to share much of this code with the shift equivalent.
Differential Revision: https://reviews.llvm.org/D48655
llvm-svn: 335957
These opcodes have a fixed type of i8 for their immediate and shouldn't have anything to do with the scalar shift amount used by target independent shift nodes.
llvm-svn: 335578
This recommits r335562 and 335563 as a single commit.
The frontend will surround the intrinsic with the appropriate marshalling to/from a scalar type to match the sigature of the builtin that software expects.
By exposing the vXi1 type directly in the llvm intrinsic we make it available to optimizers much earlier. This can enable the scalar marshalling code to be optimized away.
llvm-svn: 335568
They appear to be untested other than the test case for p37879.ll and I believe we should be using SimplifyDemandedElts here to handle these cases.
llvm-svn: 335436
Changing the logic of scalar mask folding to check for valid input types rather
than against invalid ones, making it more robust and fixing PR37879.
Differential Revision: https://reviews.llvm.org/D48366
llvm-svn: 335323
This should help in lowering the following four intrinsics:
_mm256_cvtepi32_epi8
_mm256_cvtepi64_epi16
_mm256_cvtepi64_epi8
_mm512_cvtepi64_epi8
Differential Revision: https://reviews.llvm.org/D46957
llvm-svn: 335238
I don't believe there is any real reason to have separate X86 specific opcodes for vector compares. Setcc has the same behavior just uses a different encoding for the condition code.
I had to change the CondCodeAction for SETLT and SETLE to prevent some transforms from changing SETGT lowering.
Differential Revision: https://reviews.llvm.org/D43608
llvm-svn: 335173
This patch handles back-end folding of generic patterns created by lowering the
X86 rounding intrinsics to native IR in cases where the instruction isn't a
straightforward packed values rounding operation, but a masked operation or a
scalar operation.
Differential Revision: https://reviews.llvm.org/D45203
llvm-svn: 335037
isVectorClearMaskLegal() is the TLI hook used by the generic
DAGCombiner::XformToShuffleWithZero().
We've grown to accomodate/expect this transform to shuffle
(disabling it more generally results in many regressions).
So I'm narrowly excluding the 256-bit types that clearly
are not worthwhile for AVX1.
I think in most cases we are able to recover by converting
the shuffle back into 'and' ops, but the cases in:
https://bugs.llvm.org/show_bug.cgi?id=37749
...show that there are cracks.
llvm-svn: 334759
This shortcoming was noted in D47330, and the test diffs show we already
had other examples where we failed to fold to a SHRUNKBLEND:
/// Dynamic (non-constant condition) vector blend where only the sign bits
/// of the condition elements are used. This is used to enforce that the
/// condition mask is not valid for generic VSELECT optimizations.
This patch implements an idea from D48043 and would obsolete that patch
because it catches more cases (notable the AVX1 case that was missed there).
All we're doing is allowing the existing transform to fire more often by
removing the post-legalize constraint. All of the relevant feature checks
and other predicates are left as-is.
Differential Revision: https://reviews.llvm.org/D48078
llvm-svn: 334592
Extension to D46954 (PR37426), this patch adds support for v8i16/v16i16 rotations in a similar manner - the conversion of the shift/rotate amount to a multiplication factor and the use of PMULLW to shift left and PMULHUW (ISD::MULHU) to shift the wrapped bits back around to be ORd together.
Differential Revision: https://reviews.llvm.org/D47822
llvm-svn: 334309
Simplify combineVectorTruncationWithPACKUS to mask the upper bits followed by calling truncateVectorWithPACK instead of duplicating with similar code.
This results in the codegen using (V)PACKUSDW on SSE41+ targets for vXi64/vXi32 inputs where before it always used PACKUSWB (along with a lot more bitcasting).
I've raised PR37749 as until we avoid unnecessary concats back to 256-bit for bitwise ops, we can't avoid splitting the input value into 128-bit subvectors for masking.
llvm-svn: 334289
We have some combines/lowerings that attempt to use PACKSS-then-PACKUS and others that use PACKUS-then-PACKSS.
PACKUS is much easier to combine with if we know the upper bits are zero as ComputeKnownBits can easily see through BITCASTs etc. especially now that rL333995 and rL334007 have landed. It also effectively works at byte level which further simplifies shuffle combines.
The only (minor) annoyances are that ComputeKnownBits can sometimes take longer as it doesn't fail as quickly as ComputeNumSignBits (but I'm not seeing any actual regressions in tests) and PACKUSDW only became available after SSE41 so we have more codegen diffs between targets.
llvm-svn: 334276
Simplify combineVectorTruncationWithPACKSS to just a SIGN_EXTEND_INREG followed by using the existing truncateVectorWithPACK instead of duplicating code.
llvm-svn: 334193
Only the bottom 16-bits of BEXTR's control op are required (0:8 INDEX, 15:8 LENGTH).
Differential Revision: https://reviews.llvm.org/D47690
llvm-svn: 334083
Similar to v4i32 SHL, convert v8i16 shift amounts to scale factors instead to improve performance and reduce instruction count. We were already doing this for constant shifts, this adds variable shift support.
Reduces the serial nature of the codegen, which relies on chains of plendvb/pand+pandn+por shifts.
This is a step towards adding support for vXi16 vector rotates.
Differential Revision: https://reviews.llvm.org/D47546
llvm-svn: 334023
Ideally we'd use resolveTargetShuffleInputs to handle faux shuffles as well but:
(a) that code path doesn't handle general/pre-legalized ops/types very well.
(b) I'm concerned about the compute time as they recurse to calls to computeKnownBits/ComputeNumSignBits which would need depth limiting somehow.
llvm-svn: 334007
This is the new version of D46181, allowing setjmp/longjmp
to work correctly with the Intel CET shadow stack by storing
SSP on setjmp and fixing it on longjmp. The patch has been
updated to use the cf-protection-return module flag instead
of HasSHSTK, and the bug that caused D46181 to be reverted
has been fixed with the test expanded to track that fix.
patch by mike.dvoretsky
Differential Revision: https://reviews.llvm.org/D47311
llvm-svn: 333990
Previously we just returned undef, but really we should be returning the pass thru input. We also need to make sure we preserve the chain output that the original intrinsic node had to maintain connectivity in the DAG. So we should just return the incoming chain as the output chain.
llvm-svn: 333804
Noticed while fixing PR37426, for splat rotations (rotation by an uniform value) its better to just expand back to shift ops than performing as a general non-uniform rotation.
llvm-svn: 333661
This improves splat rotations (rotation by an uniform value), to avoid having to use the generic non-uniform shift code (extension to PR37426).
llvm-svn: 333641
Created the IsSplatValue helper from the splat detection code in LowerScalarVariableShift as a first NFC step towards improving support for splat rotations, which is an extension of PR37426.
llvm-svn: 333580
Support for Clang lowering of fused intrinsics. This patch:
1. Removes bindings to clang fma intrinsics.
2. Introduces new LLVM unmasked intrinsics with rounding mode:
int_x86_avx512_vfmadd_pd_512
int_x86_avx512_vfmadd_ps_512
int_x86_avx512_vfmaddsub_pd_512
int_x86_avx512_vfmaddsub_ps_512
supported with a new intrinsic type (INTR_TYPE_3OP_RM).
3. Introduces new x86 fmaddsub/fmsubadd folding.
4. Introduces new tests for code emitted by sequentions introduced in Clang part.
Patch by tkrupa
Reviewers: craig.topper, sroland, spatel, RKSimon
Reviewed By: craig.topper, RKSimon
Differential Revision: https://reviews.llvm.org/D47443
llvm-svn: 333554
There seems to be no real reason to have these separate copies.
The existing implementations just copy each other for x86.
For Mips there is a subtle difference, which is just a bug
since it changes based on the context where which one was called.
Dropping this version, all tests pass. If I try to merge them
to match the removed version, a test fails.
llvm-svn: 333440
1. Introduction of mask scalar TableGen patterns.
2. Introduction of new scalar move TableGen patterns
and refactoring of existing ones.
3. Folding of pattern created by introducing scalar
masking in Clang header files.
Patch by tkrupa
Differential Revision: https://reviews.llvm.org/D47012
llvm-svn: 333419
Summary: We already get this right if the i64 didn't come from a load.
Reviewers: RKSimon
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D47439
llvm-svn: 333393
We have unmasked intrinsics now and wrap them with a select. This is a net reduction of 36 intrinsics from before the unmasked intrinsics were added.
llvm-svn: 333388
These do the same thing with the first and second sources swapped. They previously came from separate intrinsics that specified different masking behavior. But we can cover that with isel patterns and a single node.
This is a step towards reducing the number of intrinsics needed.
A bunch of tests change because we are now biased to choosing VPERMT over VPERMI when there is nothing to signal that commuting is beneficial.
llvm-svn: 333383
This basically reverts r280696 in favor of using extra patterns as mentioned as an alternative in that commit message. For now I've only added the cases we have test cases for, but it should be easy to add more in the future.
This will help to convert VPERMI2PS/VPERMT2PS intrinsics to use a single ISD node opcode. And hopefully allow some intrinsics to be removed.
llvm-svn: 333365
Summary:
This **appears** to be the last missing piece for the masked merge pattern handling in the backend.
This is [[ https://bugs.llvm.org/show_bug.cgi?id=37104 | PR37104 ]].
[[ https://bugs.llvm.org/show_bug.cgi?id=6773 | PR6773 ]] will introduce an IR canonicalization that is likely bad for the end assembly.
Previously, `andps`+`andnps` / `bsl` would be generated. (see `@out`)
Now, they would no longer be generated (see `@in`), and we need to make sure that they are generated.
Differential Revision: https://reviews.llvm.org/D46528
llvm-svn: 332904
This removes 6 intrinsics since we no longer need separate mask and maskz intrinsics.
Differential Revision: https://reviews.llvm.org/D47124
llvm-svn: 332890
As suggested by Fabian on PR37426, we can use PMULUDQ to perform v4i32 vector rotations as the upper 32bits of the multiply will contain the 'wrapped' bits of the rotation.
v8i16/v16i8 rotations would be straightforward to add to lowerRotate in the future - ideally we'd mostly share code with the vector shifts lowering.
Differential Revision: https://reviews.llvm.org/D46954
llvm-svn: 332832
The intrinsic legalization for masked truncate uses ISD::TRUNCATE which can be constant folded by getNode. This prevents getVectorMaskingNode from seeing the ISD::TRUNCATE special case where it should emit X86ISD::SELECT instead of ISD::VSELECT. This causes a vselect with a v16i1 or v8i1 condition to be emitted during vector legalization. but vector legalization doesn't revisit nodes it creates. DAG combine will then promote this condition to match the result type. Then op legalization will try to legalize it, but the custom lowering hook returned SDValue(). But op legalization doesn't have an Expand for VSELECT because it expects vector legalization to have taken care of it. So the operation sticks around and fails in isel.
This patch adds a custom legalization hook to morph it to a vXi8 vselect instead.
This also simplifies the normal vXi16 vselect handling because vector legalization was normally expanding to AND/ANDN/OR and DAG combine was turning that into VBLENDVB. So we can skip a step by doing it directly.
Fixes PR37499
Differential Revision: https://reviews.llvm.org/D47025
llvm-svn: 332743
As suggested by Fabian on PR37441, use PSHUFLW to extend shift amount types for use with PSRAD/PSRLD to reduce register pressure.
Some of this ideally would be done by combineTargetShuffle but its tricky to do as most of the shuffles are sharing inputs.
Differential Revision: https://reviews.llvm.org/D46959
llvm-svn: 332524
As i64 types are not legal on 32-bit targets, insert these into a suitable zero vector and use the packed vXi64<->FP conversion instructions instead.
Fixes PR3163.
Differential Revision: https://reviews.llvm.org/D43441
llvm-svn: 332498
Summary:
New unsigned saturation downconvert patterns detection was implemented in
X86 Codegen:
(truncate (smin (smax (x, C1), C2)) to dest_type),
where C1 >= 0 and C2 is unsigned max of destination type.
(truncate (smax (smin (x, C2), C1)) to dest_type)
where C1 >= 0, C2 is unsigned max of destination type and C1 <= C2.
These two patterns are equivalent to:
(truncate (umin (smax(x, C1), unsigned_max_of_dest_type)) to dest_type)
Reviewers: RKSimon
Subscribers: llvm-commits, a.elovikov
Differential Revision: https://reviews.llvm.org/D45315
llvm-svn: 332336
With nnan, there's no need for the masked merge / blend
sequence (that probably costs much more than the min/max
instruction).
Somewhere between clang 5.0 and 6.0, we started producing
these intrinsics for fmax()/fmin() in C source instead of
libcalls or fcmp/select. The backend wasn't prepared for
that, so we regressed perf in those cases.
Note: it's possible that other targets have similar problems
as seen here.
Noticed while investigating PR37403 and related bugs:
https://bugs.llvm.org/show_bug.cgi?id=37403
The IR FMF propagation cases still don't work. There's
a proposal that might fix those cases in D46563.
llvm-svn: 331992
Because we create a new kind of debug instruction, DBG_LABEL, we need to
check all passes which use isDebugValue() to check MachineInstr is debug
instruction or not. When expelling debug instructions, we should expel
both DBG_VALUE and DBG_LABEL. So, I create a new function,
isDebugInstr(), in MachineInstr to check whether the MachineInstr is
debug instruction or not.
This patch has no new test case. I have run regression test and there is
no difference in regression test.
Differential Revision: https://reviews.llvm.org/D45342
Patch by Hsiangkai Wang.
llvm-svn: 331844
This is a fix for PR30290: by marking all byval stack slots as being aliased,
the instruction scheduler is more conservative about rescheduling memory
accesses to such stack slots as an LLVM Value* might alias it. This fixes
errors such as in the patched test case, where reads and writes to a data
structure are illegally mixed.
This could be fixed better in the future with better analysis for the
instruction scheduler to know what Values alias what stack slots.
Differential Revision: https://reviews.llvm.org/D45022
llvm-svn: 331749
This patch adds a shadow stack fix when compiling
setjmp/longjmp with the shadow stack enabled. This
allows setjmp/longjmp to work correctly with CET.
Patch by mike.dvoretsky
Differential Revision: https://reviews.llvm.org/D46181
llvm-svn: 331748
Summary:
Split off from D46031.
In masked merge case, this degrades IPC by decreasing instruction count.
{F6108777}
The next patch should be able to recover and improve this.
This also affects the transform @spatel have added in D27489 / rL289738,
and the test coverage for X86 was missing.
But after i have added it, and looked at the changes in MCA, i'm somewhat confused.
{F6093591} {F6093592} {F6093593}
I'd say this regression is an improvement, since `IPC` increased in that case?
Reviewers: spatel, craig.topper
Reviewed By: spatel
Subscribers: andreadb, llvm-commits, spatel
Differential Revision: https://reviews.llvm.org/D46493
llvm-svn: 331684
Summary:
The legacy VRCPPS/VRSQRTPS instructions aren't available in 512-bit versions. The new increased precision versions are. So we can use those to implement v16f32 reciprocal estimates.
For KNL CPUs we can probably use VRCP28PS/VRSQRT28PS and avoid the NR step altogether, but I leave that for a future patch.
Reviewers: spatel
Reviewed By: spatel
Subscribers: RKSimon, llvm-commits, mehdi_amini
Differential Revision: https://reviews.llvm.org/D46498
llvm-svn: 331606
We've been running doxygen with the autobrief option for a couple of
years now. This makes the \brief markers into our comments
redundant. Since they are a visual distraction and we don't want to
encourage more \brief markers in new code either, this patch removes
them all.
Patch produced by
for i in $(git grep -l '\\brief'); do perl -pi -e 's/\\brief //g' $i & done
Differential Revision: https://reviews.llvm.org/D46290
llvm-svn: 331272
Summary:
Previously the flag intrinsics always used the index instructions even if a mask instruction also exists.
To fix fix this I've created a single ISD node type that returns index, mask, and flags. The SelectionDAG CSE process will merge all flavors of intrinsics with the same inputs to a s ingle node. Then during isel we just have to look at which results are used to know what instruction to generate. If both mask and index are used we'll need to emit two instructions. But for all other cases we can emit a single instruction.
Since I had to do manual isel anyway, I've removed the pseudo instructions and custom inserter code that was working around tablegen limitations with multiple implicit defs.
I've also renamed the recently added sse42.ll test case to sttni.ll since it focuses on that subset of the sse4.2 instructions.
Reviewers: chandlerc, RKSimon, spatel
Reviewed By: chandlerc
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D46202
llvm-svn: 331091
Previously we only formed MUL_IMM when we split a constant. This blocked load folding on those cases. We should also form MUL_IMM for 3/5/9 to favor LEA over load folding.
Differential Revision: https://reviews.llvm.org/D46040
llvm-svn: 330850
Summary:
If attribute "use-soft-float"="true" is set then X86ISelLowering.cpp sets
'Promote' action for ISD::SINT_TO_FP operation on type i32.
But 'Promote' action is not proper in this case since lib function
__floatsidf is available for casting from signed int to float type.
Thus Expand action is more suitable here.
The Expand action should be set for ISD::UINT_TO_FP for soft float as well.
If function attribute "use-soft-float"="true" is set then infinite looping
can happen in DAG combining, function visitSINT_TO_FP() replaces SINT_TO_FP
node with UINT_TO_FP node and function combineUIntToFP() replace vice versa in cycle.
The fix prevents it.
Patch by vrybalov
Differential Revision: https://reviews.llvm.org/D45572
llvm-svn: 330711
Three new instructions:
umonitor - Sets up a linear address range to be
monitored by hardware and activates the monitor.
The address range should be a writeback memory
caching type.
umwait - A hint that allows the processor to
stop instruction execution and enter an
implementation-dependent optimized state
until occurrence of a class of events.
tpause - Directs the processor to enter an
implementation-dependent optimized state
until the TSC reaches the value in EDX:EAX.
Also modifying the description of the mfence
instruction, as the rep prefix (0xF3) was allowed
before, which would conflict with umonitor during
disassembly.
Before:
$ echo 0xf3,0x0f,0xae,0xf0 | llvm-mc -disassemble
.text
mfence
After:
$ echo 0xf3,0x0f,0xae,0xf0 | llvm-mc -disassemble
.text
umonitor %rax
Reviewers: craig.topper, zvi
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D45253
llvm-svn: 330462
This is the patch that lowers x86 intrinsics to native IR
in order to enable optimizations. The patch also includes folding
of previously missing saturation patterns so that IR emits the same
machine instructions as the intrinsics.
Patch by tkrupa
Differential Revision: https://reviews.llvm.org/D44785
llvm-svn: 330322
Summary:
Add an LLVM intrinsic for type discriminated event logging with XRay.
Similar to the existing intrinsic for custom events, but also accepts
a type tag argument to allow plugins to be aware of different types
and semantically interpret logged events they know about without
choking on those they don't.
Relies on a symbol defined in compiler-rt patch D43668. I may wait
to submit before I can see demo everything working together including
a still to come clang patch.
Reviewers: dberris, pelikan, eizan, rSerge, timshen
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D45633
llvm-svn: 330219
This completes the work started in r329604 and r329605 when we changed clang to no longer use the intrinsics.
We lost some InstCombine SimplifyDemandedBit optimizations through this change as we aren't able to fold 'and', bitcast, shuffle very well.
llvm-svn: 329990
This cleans up a number of operations that only claimed te use EFLAGS
due to using DF. But no instructions which we think of us setting EFLAGS
actually modify DF (other than things like popf) and so this needlessly
creates uses of EFLAGS that aren't really there.
In fact, DF is so restrictive it is pretty easy to model. Only STD, CLD,
and the whole-flags writes (WRFLAGS and POPF) need to model this.
I've also somewhat cleaned up some of the flag management instruction
definitions to be in the correct .td file.
Adding this extra register also uncovered a failure to use the correct
datatype to hold X86 registers, and I've corrected that as necessary
here.
Differential Revision: https://reviews.llvm.org/D45154
llvm-svn: 329673
The key idea is to lower COPY nodes populating EFLAGS by scanning the
uses of EFLAGS and introducing dedicated code to preserve the necessary
state in a GPR. In the vast majority of cases, these uses are cmovCC and
jCC instructions. For such cases, we can very easily save and restore
the necessary information by simply inserting a setCC into a GPR where
the original flags are live, and then testing that GPR directly to feed
the cmov or conditional branch.
However, things are a bit more tricky if arithmetic is using the flags.
This patch handles the vast majority of cases that seem to come up in
practice: adc, adcx, adox, rcl, and rcr; all without taking advantage of
partially preserved EFLAGS as LLVM doesn't currently model that at all.
There are a large number of operations that techinaclly observe EFLAGS
currently but shouldn't in this case -- they typically are using DF.
Currently, they will not be handled by this approach. However, I have
never seen this issue come up in practice. It is already pretty rare to
have these patterns come up in practical code with LLVM. I had to resort
to writing MIR tests to cover most of the logic in this pass already.
I suspect even with its current amount of coverage of arithmetic users
of EFLAGS it will be a significant improvement over the current use of
pushf/popf. It will also produce substantially faster code in most of
the common patterns.
This patch also removes all of the old lowering for EFLAGS copies, and
the hack that forced us to use a frame pointer when EFLAGS copies were
found anywhere in a function so that the dynamic stack adjustment wasn't
a problem. None of this is needed as we now lower all of these copies
directly in MI and without require stack adjustments.
Lots of thanks to Reid who came up with several aspects of this
approach, and Craig who helped me work out a couple of things tripping
me up while working on this.
Differential Revision: https://reviews.llvm.org/D45146
llvm-svn: 329657
LowerIntUnary as its name says has an assert for integer types. But for the bitcast case one side might be an FP type.
Rather than making sure the function really works for fp types and renaming it. Just do really basic splitting directly. The LowerIntUnary has the advantage that it can peek through BUILD_VECTOR because every other call is during Lowering. But these calls are during legalization and will be followed by a DAG combine round.
Revert some change to LowerVectorIntUnary that were originally made just to make these two calls work even in pure integer cases.
This was found purely by compiling the avx512f-builtins.c test from clang so I've copied over the offending function from that.
llvm-svn: 329616
Summary:
r327219 added wrappers to std::sort which randomly shuffle the container before sorting.
This will help in uncovering non-determinism caused due to undefined sorting
order of objects having the same key.
To make use of that infrastructure we need to invoke llvm::sort instead of std::sort.
Note: This patch is one of a series of patches to replace *all* std::sort to llvm::sort.
Refer the comments section in D44363 for a list of all the required patches.
Reviewers: chandlerc, craig.topper, RKSimon
Reviewed By: chandlerc, craig.topper
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D44874
llvm-svn: 329534
Previously we used a custom lowering for this because of the AVX1 splitting requirement. But we can do the split during DAG combine if we check the types and subtarget
llvm-svn: 329510
This Promote flag was alwasys set to true except in the default case. But in the default case we don't need to set PVT and can just return false.
llvm-svn: 328926
We are re-adding all the bitcasts, constant masks and target shuffles to the work list for no apparent gain.
Found while investigating adding SimplifyDemandedVectorElts to target shuffles.
Differential Revision: https://reviews.llvm.org/D44942
llvm-svn: 328771
Summary:
Re-lands r328386 and r328443, reverting r328482.
Incorporates fixes from @mstorsjo in D44876 (thanks!) so that small
parameters in i8 and i16 do not end up in the SysV register parameters
(EDI, ESI, etc).
I added tests for how we receive small parameters, since that is the
important part. It's always safe to store more bytes than will be read,
but the assumptions you make when loading them are what really matter.
I also tested this by self-hosting clang and it passed tests on win64.
Reviewers: mstorsjo, hans
Subscribers: hiraditya, mstorsjo, llvm-commits
Differential Revision: https://reviews.llvm.org/D44900
llvm-svn: 328570
This broke Chromium (see crbug.com/825748). It looks like mstorsjo's follow-up
patch at D44876 fixes this, but let's revert back to green for now until that's
ready to land.
(Also reverts r328443.)
> Both GCC and MSVC only look at the low byte of a boolean when it is
> passed.
llvm-svn: 328482
These nodes only use the lower 32 bits of their inputs so we can use SimplifyDemandedBits to simplify them.
Differential Revision: https://reviews.llvm.org/D44375
llvm-svn: 328405
Mingw uses the same stack protector functions as GCC provides
on other platforms as well.
Patch by Valentin Churavy!
Differential Revision: https://reviews.llvm.org/D27296
llvm-svn: 328039
We don't need to create an ISD::TRUNCATE node to return, we started with one and can return it. Also remove the call to getExtendInVec, the result is just going to be a getNode of that value passed in.
llvm-svn: 327914
With the SRAs removed from the SSE2 code in D44267, then there doesn't appear to be any advantage to the sse41 code. The punpcklbw instruction and pmovsx seem to have the same latency and throughput on most CPUs. And the SSE41 code requires moving the upper 64-bits into the lower 64-bit before the sign extend can be done. The unpckhbw in sse2 code can do better than that.
llvm-svn: 327869
X86 Supports Indirect Branch Tracking (IBT) as part of Control-Flow Enforcement Technology (CET).
IBT instruments ENDBR instructions used to specify valid targets of indirect call / jmp.
The `nocf_check` attribute has two roles in the context of X86 IBT technology:
1. Appertains to a function - do not add ENDBR instruction at the beginning of the function.
2. Appertains to a function pointer - do not track the target function of this pointer by adding nocf_check prefix to the indirect-call instruction.
This patch implements `nocf_check` context for Indirect Branch Tracking.
It also auto generates `nocf_check` prefixes before indirect branchs to jump tables that are guarded by range checks.
Differential Revision: https://reviews.llvm.org/D41879
llvm-svn: 327767
Previously, we called the same functions twice with a bool flag determining whether we should look for ADDSUB or SUBADD. It would be more efficient to run the code once and detect either pattern with a flag to tell which type it found.
Differential Revision: https://reviews.llvm.org/D44540
llvm-svn: 327730
Previously if getSetccResultType returned an illegal type we just fell back to using the default promoted type. This appears to have been to handle the case where for vectors getSetccResultType returns the input type, but the input type itself isn't legal and will need to be promoted. Without the legality check we would never reach a legal type.
But just picking the promoted type to be the setcc type can create strange setccs where the result type is 128 bits and the operand type is 256 bits. If for example the result type was promoted to v8i16 from v8i1, but the input type was promoted from v8i23 to v8i32. We currently handle this with custom lowering code in X86.
This legality check also caused us reject the getSetccResultType when the input type needed to be widened or split. Even though that result wouldn't have caused legalization to get stuck.
This patch tries to fix this by detecting the getSetccResultType needs to be promoted. If its input type also needs to be promoted we'll try a ask for a new setcc result type based on its eventual promoted value. Otherwise we fall back to default type to promote to.
For any other illegal values we might get back from the initial call to getSetccResultType we just keep and allow it to be re-legalized later via splitting or widening or scalarizing.
llvm-svn: 327683
The FADD part of the addsub/subadd pattern can have its operands commuted, but when checking for fsubadd we were using the fadd as reference and commuting the fsub node.
llvm-svn: 327660
Rather than enumerating all specific types, for the DAG combine we can just use TLI::isTypeLegal and an SSE3 check. For the BUILD_VECTOR version we already know the type is legal so we just need to check SSE3.
llvm-svn: 327649
I had to modify the bswap recognition to allow unshrunk masks to make this work.
Fixes PR36689.
Differential Revision: https://reviews.llvm.org/D44442
llvm-svn: 327530
We now only create recursive concats if we have more than two non-zero values. This keeps our subvector broadcast DAG combine functioning.
llvm-svn: 327457
This better able to detect undef and zeros pieces in the concat. Or cases when only one subvector is non-zero. This allows us to avoid silly things like double inserts into progressively larger undefs.
This still builds 512 bit concats of 128 bits by building up through 256 bits first. But I don't know if that's best.
We probably want to merge this with the vXi1 concat code since they are very similar.
llvm-svn: 327454
Summary: Unless you were intentionally avoiding this syntax? I saw you mentioned makeArrayRef in your commit that added SplitOpsAndApply.
Reviewers: RKSimon
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D44403
llvm-svn: 327418
MVT belongs to the CodeGen layer, but ShuffleDecode is used by the X86 InstPrinter which is part of the MC layer. This only worked because MVT is completely implemented in a header file with no other library dependencies.
Differential Revision: https://reviews.llvm.org/D44353
llvm-svn: 327292
We called MaskedValueIsZero with two different masks, but underneath that calls computeKnownBits before applying the mask. This means we compute the same known bits twice due to the two calls. Instead just call computeKnownBits directly and apply the two masks ourselves.
llvm-svn: 327251
64-bit MMX vector generation usually ends up lowering into SSE instructions before being spilled/reloaded as a MMX type.
This patch creates a MMX vector from MMX source values, taking the lowest element from each source and constructing broadcasts/build_vectors with direct calls to the MMX PUNPCKL/PSHUFW intrinsics.
We're missing a few consecutive load combines that could be handled in a future patch if that would be useful - my main interest here is just avoiding a lot of the MMX/SSE crossover.
Differential Revision: https://reviews.llvm.org/D43618
llvm-svn: 327247
Same as the VPERMILPS/VPERMILPD approach for v8f32/v4f64 cases, rely on PSHUFB using bits[3:0] for indexing - we can ignore the sign bit (zero element) as those index vector values are considered undefined. The select between the lo/hi permute results based on the index size.
llvm-svn: 327242
As VPERMILPS/VPERMILPD only selects elements based on the bits[1:0]/bit[1] then we can permute both the (repeated) lo/hi 128-bit vectors in each case and then select between these results based on whether the index was for for lo/hi.
For v4i64/v4f64 this avoids some rather nasty v4i64 multiples on the AVX2 implementation, which seems to be worse than the extra port5 pressure from the additional shuffles/blends.
llvm-svn: 327239
Helper function to insert a subvector into the bottom elements of a larger zero/undef vector with the same scalar type.
I've converted a couple of INSERT_SUBVECTOR calls to use it, there are plenty more although in some cases I was worried it might make the code more ambiguous.
llvm-svn: 327236
Previously we unpacked the even bytes of each input into the high byte of 16-bit elements then did an v8i16 arithmetic shift right by 8 bits to fill the upper bits of each word with sign bits. Then we did the v8i16 multiply and then masked to zero the upper 8-bits of each result. The similar was done for all the odd bytes. The results are then packed together with packuswb
Since we are masking each multiply result element to 8-bits, and those 8-bits are determined only by the lower 8-bits of each of the inputs, we don't need to fill the upper bits with sign bits. So we can just unpack into the low byte of each element and treat the upper bits as garbage. This is what gcc also does.
Differential Revision: https://reviews.llvm.org/D44267
llvm-svn: 327093
This instruction can be thought of as reading either the even elements of a vXi32 input or the lower half of each element of a vXi64 input. We currently use the vXi32 interpretation, but vXi64 matches better with its broadcast behavior in EVEX.
I'm looking at moving MULDQ/MULUDQ creation to a DAG combine so we can do it when AVX512DQ is enabled without having to go through Custom lowering. But in some of the test cases we failed to use a broadcast load due to the size difference. This should help with that.
I'm also wondering if we can model these instructions in native IR and remove the intrinsics and I think using a vXi64 type will work better with that.
llvm-svn: 326991
The v8i32 conversion on AVX1 targets was only working after LowerMUL splits 256-bit vectors.
While I was there I've also made it so we don't have to check for AVX2 and BWI directly and instead just ask if the type is legal.
Differential Revision: https://reviews.llvm.org/D44190
llvm-svn: 326917
The code checks Level == AfterLegalizeDAG which is the fourth and last of the possible DAG combine stages that we have.
There is a Level called AfterLegalVectorOps, but that's the third DAG combine and it doesn't always run.
A function called isAfterLegalVectorOps should imply it returns true in either of the DAG combines that runs after the legalize vector ops stage, but that's not what this function does.
llvm-svn: 326832
Almost none of these usages were FP specific. And we had no clear guideliness on when to use hasAVX vs hasFP256.
I might also remove hasInt256 too since its an alias for hasAVX2.
llvm-svn: 326682
We were previously doing this with isel patterns. Moving it to op legalization gives us chance to see the required bitcast earlier. And it lets us remove some isel patterns.
llvm-svn: 326669
64-bit MMX constant generation usually ends up lowering into SSE instructions before being spilled/reloaded as a MMX type.
This patch bitcasts the constant to a double value to allow correct loading directly to the MMX register.
I've added MMX constant asm comment support to improve testing, it's better to always print the double values as hex constants as MMX is mainly an integer unit (and even with 3DNow! its just floats).
Differential Revision: https://reviews.llvm.org/D43616
llvm-svn: 326497
Emulated TLS is enabled by llc flag -emulated-tls,
which is passed by clang driver.
When llc is called explicitly or from other drivers like LTO,
missing -emulated-tls flag would generate wrong TLS code for targets
that supports only this mode.
Now use useEmulatedTLS() instead of Options.EmulatedTLS to decide whether
emulated TLS code should be generated.
Unit tests are modified to run with and without the -emulated-tls flag.
Differential Revision: https://reviews.llvm.org/D42999
llvm-svn: 326341
An extract_element where the result type is larger than the scalar element type is semantically an any_extend of from the scalar element type to the result type. If we expect zeroes in the upper bits of the i8/i32 we need to mae sure those zeroes are explicit in the DAG.
For these cases the best way to accomplish this is use an insert_subvector to pad zeroes to the upper bits of the v1i1 first. We extend to either v16i1(for i32) or v8i1(for i8). Then bitcast that to a scalar and finish with a zero_extend up to i32 if necessary. We can't extend past v16i1 because that's the largest mask size on KNL. But isel is smarter enough to know that a zext of a bitcast from v16i1 to i16 can use a KMOVW instruction. The insert_subvectors will be dropped during isel because we can determine that the producing instruction already zeroed the upper bits of the k-register.
llvm-svn: 326308
While the description for the instruction does mention OR, its talking about how the individual classification test results are ORed together.
The incoming mask is used as a zeroing write mask. If the bit is 1 the classification is written to the output. The bit is 0 the output is 0. This equivalent to an AND.
Here is pseudocode from the intrinsics guide
FOR j := 0 to 1
i := j*64
IF k1[j]
k[j] := CheckFPClass_FP64(a[i+63:i], imm8[7:0])
ELSE
k[j] := 0
FI
ENDFOR
k[MAX:2] := 0
llvm-svn: 326306
There's still some shortcoming in our ability to combine binops of constants with different sizes separated by an extend. I'll try to look at that next.
llvm-svn: 326128
Summary:
We have an early DAG combine to turn these patterns into MOVMSK, but that combine doesn't work if the vXi1 type has more elements than the widest legal vXi8 type. Type legalization will eventually split it down to v16i1 or v32i1 and then the bitcast gets legalized to a truncstore and a scalar load. The truncstore will get lowered to a series of extracts and bit math.
This patch adds a custom legalization to use a sign extend and MOVMSK instead. This prevents the eventual scalarization.
Reviewers: spatel, RKSimon, zvi
Reviewed By: RKSimon
Subscribers: mgorny, llvm-commits
Differential Revision: https://reviews.llvm.org/D43593
llvm-svn: 326119
This code seemed to try to widen to 128, 256, or 512 bit vectors, but we only create X86ISD::AVG with a power of 2 number of elements. This means the only nodes that need to be legalized are less than 128-bits and need to be widened up to 128 bits.
llvm-svn: 326064
Which types are considered 'simple' is a function of the requirements of all targets that LLVM supports. That shouldn't directly affect what types we are able to handle. The remainder of this code checks that the number of elements is a power of 2 and takes care of splitting down to a legal size.
llvm-svn: 326063
Our UMIN/UMAX, vector truncation and shuffle combining is good enough to efficiently handle v8i64 with the number of leading zeros that are necessary for PSUBUS.
llvm-svn: 326034
Now that UMIN etc are Legal/Custom for SSE2+, we can efficiently match SUBUS v8i32 cases from SSSE3 which can perform efficient truncation with PSHUFB.
llvm-svn: 326033
These can be created by type legalization promoting the inputs to select to match scalar boolean contents.
We were trying to pattern match them away during isel, but its better to just remove them from the DAG.
I've cleaned up some patterns to not check for this 'and' anymore. But I suspect this has also opened up opportunities for pattern removal.
llvm-svn: 325949
The test changes you can see are related to the changes in ReplaceNodeResults. Though shuffle-vs-trunc-512.ll does have a test that exercises the code in LowerBITCAST. Looks like the test output didn't change because DAG combining is able to clean up the resulting type legalization. Adding the custom hook just makes type legalization work less hard.
Differential Revision: https://reviews.llvm.org/D43447
llvm-svn: 325933
We won't be able to fold the constant pool load, but its still better than materialing ones and xoring for the invert if we used PCMPEQ.
This will fix another regression from D42948.
llvm-svn: 325845
Previously this code overrode the flags and opcode used by the later code in LowerVSETCC. This makes the code difficult to read and follow.
This patch moves all the SUBUS code into its own function and makes it responsible for creating its own SDNodes on success.
Differential Revision: https://reviews.llvm.org/D43530
llvm-svn: 325827
SimplifyDemandedBits forces the demanded mask to all 1s if the node has multiple uses, unless the AssumeSingleUse flag is set.
So previously we were only really likely to simplify something if the condition had a single use. And on the off chance we did simplify with multiple uses the demanded mask being used was all ones so there was no reason to create a shrunkblend.
This patch now checks that the condition is only used by selects first, and then sets the AssumeSingleUse flag for the simplifcation. Then we convert the selects to shrunkblend, and finally replace condition.
Differential Revision: https://reviews.llvm.org/D43446
llvm-svn: 325604
This allows us to avoid an opsize prefix. And forcing some move immediates to i32 avoids a length changing prefix on those instructions.
This mostly replaces the existing combine we had for zext/sext+cmov of constants. I left in a case for sign extending a 32 bit cmov of constants to 64 bits.
Differential Revision: https://reviews.llvm.org/D43327
llvm-svn: 325601
Previously we used vptestmd, but the scheduling data for SKX says vpmovq2m/vpmovd2m is lower latency. We already used vpmovb2m/vpmovw2m for byte/word truncates. So this is more consistent anyway.
llvm-svn: 325534
We swapped the operands and used setle, but I don't see any reason to do that. I think this is a holdover from SSE where we swap and the invert to use pcmpgt. But with AVX512 we don't want an invert so we won't use pcmpgt. So there's no need to swap.
llvm-svn: 325527
Canonicalize EQ/NE PCMPM to have build vector all zeros on the RHS so we don't have to pattern match it in both locations. This significantly reduces the number of isel patterns needed since we also had to multiply it out with loads being in either operand of the 'and' input node and in the 'and' masking node.
This removes over 24000 bytes from the isel table.
llvm-svn: 325526
We're accidentally checking that the same node is a constant twice instead of checking the other node.
This isn't a functional problem since we didn't do anything below that explicitly requires constants. It just means we may have introduced a sign_extend or zero_extend that won't fold out.
llvm-svn: 325469
Summary:
Currently we convert to shuffles during lowering. This moves it to DAG combine so hopefully we can get it done before type legalization has to extend the condition.
I believe in some cases we're creating SHRUNKBLENDs that end up with constant conditions because we see the extended on the condition and think its a dynamic selelect before DAG combine gets a chance to constant fold the extend. We could add combines to turn SHRUNKBLENDs with constant condition back to vselect. But it seemed like it might be better to just send them to shuffles as early as possible so they never get a chance to become SHRUNKBLENDs. This the reason some tests went from blends controlled by a constant pool load to just move.
Some of the constant pool entries changed because the sign_extend introduced by type legalization turned undef elements in select condition into 0s. While the select->shuffle used -1 in the shuffle mask. So now the shuffle lowering can do what it wants with them.
I'll remove the lowering code as a follow up. We might be able to simplify some of the pre-checks for SHRUNKBLEND as the FIXME there says.
Reviewers: spatel, RKSimon, efriedma, zvi, andreadb
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D43367
llvm-svn: 325417
Undef in select condition means we should pick the element from one side or the other. An undef in a shuffle mask means pick any element from either source or worse.
I suspect by the time we get here most of the undefs in a constant vector have been removed by other things, but doing this for safety.
llvm-svn: 325394
This seems to interfere with a target independent brcond combine that looks for the (srl (and X, C1), C2) pattern to enable TEST instructions. Once we flip, that combine doesn't fire and we end up exposing it to the X86 specific BT combine which causes us to emit a BT instruction. BT has lower throughput than TEST.
We could try to make the brcond combine aware of the alternate pattern, but since the flip was just a code size reduction and not likely to enable other combines, it seemed easier to just delay it until after lowering.
Differential Revision: https://reviews.llvm.org/D43201
llvm-svn: 325371
We already do this for 64-bit when it won't fit into a 64-bit AND/TEST's immediate field. This adds an additional qualifier to do it for any single bit constant larger than 8-bits under optsize
Differential Revision: https://reviews.llvm.org/D43346
llvm-svn: 325290
We can use PACKSS to saturate each stage of the chain: PACKSSDW down to [-32768,32767] and then PACKSSWB to [-128,127].
PACKUS is a little trickier and will be handled in a separate patch.
llvm-svn: 325235
Try to keep PACK*SDW/PACK*SWB as wide as possible, this helps ComputeNumSignBits as it can only peek through bitcasts to wider types, pre-AVX2 codegen was already doing this as it could peek through bitcasts/subvectors more easily than AVX2 could through shuffles.
This shouldn't affect existing results as calls to truncateVectorWithPACK ensure we have enough sign bits to pack to the same value, but it should make it possible to use truncateVectorWithPACK chains to perform saturation in combineTruncateWithSat with a future patch.
llvm-svn: 325149
While the AVX512 VTRUNCS/VTRUNCUS instructions require legal types, truncateVectorWithPACK handles cases with multiples of legal types through splitting/concatenation. So we just need to ensure that the src/dst scalar types are correct and leave truncateVectorWithPACK to handle the rest of it.
llvm-svn: 325127
Summary:
Instead of solving the hard problem of how to pass the callee to the indirect
jump thunk without a register, just use a CSR. At a call boundary, there's
nothing stopping us from using a CSR to hold the callee as long as we save and
restore it in the prologue.
Also, add tests for this mregparm=3 case. I wrote execution tests for
__llvm_retpoline_push, but they never got committed as lit tests, either
because I never rewrote them or because they got lost in merge conflicts.
Reviewers: chandlerc, dwmw2
Subscribers: javed.absar, kristof.beyls, hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D43214
llvm-svn: 325049
ISD::ADD implies individual vector element addition with no carries between elements. But for a vXi1 type that would be the same as XOR. And we already turn ISD::ADD into ISD::XOR for all vXi1 types during lowering. So the ISD::ADD pattern would never be able to match anyway.
KADD is different, it adds the elements but also propagates a carry between them. This just a way of doing an add in k-register without bitcasting to the scalar domain. There's still no way to match the pattern, but at least its not obviously wrong.
llvm-svn: 324861
Summary:
Currently we only use min/max to help with ule/uge compares because it removes an invert of the result that would otherwise be needed. But we can also use it for ult/ugt compares if it will prevent the need for a sign bit flip needed to use pcmpgt at the cost of requiring an invert after the compare.
I also refactored the code so that the max/min code is self contained and does its own return instead of setting up a flag to manipulate the rest of the function's behavior.
Most of the test cases look ok with this. I did notice that we added instructions when one of the operands being sign flipped is a constant vector that we were able to constant fold the flip into.
I also noticed that sometimes the SSE min/max clobbers a register that is needed after the compare. This resulted in an extra move being inserted before the min/max to preserve the register. We could try to detect this and switch from min to max and change the compare operands to use the operand that gets reused in the compare.
Reviewers: spatel, RKSimon
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D42935
llvm-svn: 324842
This allows us to recognise more saturation patterns and also simplify some MINMAX codegen that was failing to combine CMPGE comparisons to a legal CMPGT.
Differential Revision: https://reviews.llvm.org/D43014
llvm-svn: 324837
This patch adds a new function attribute "required-vector-width" that can be set by the frontend to indicate the maximum vector width present in the original source code. The idea is that this would be set based on ABI requirements, intrinsics or explicit vector types being used, maybe simd pragmas, etc. The backend will then use this information to determine if its save to make 512-bit vectors illegal when the preference is for 256-bit vectors.
For code that has no vectors in it originally and only get vectors through the loop and slp vectorizers this allows us to generate code largely similar to our AVX2 only output while still enabling AVX512 features like mask registers and gather/scatter. The loop vectorizer doesn't always obey TTI and will create oversized vectors with the expectation the backend will legalize it. In order to avoid changing the vectorizer and potentially harm our AVX2 codegen this patch tries to make the legalizer behavior similar.
This is restricted to CPUs that support AVX512F and AVX512VL so that we have good fallback options to use 128 and 256-bit vectors and still get masking.
I've qualified every place I could find in X86ISelLowering.cpp and added tests cases for many of them with 2 different values for the attribute to see the codegen differences.
We still need to do frontend work for the attribute and teach the inliner how to merge it, etc. But this gets the codegen layer ready for it.
Differential Revision: https://reviews.llvm.org/D42724
llvm-svn: 324834
We promote these via a DAG combine now before lowering gets the chance.
Also remove the v2i1 custom handling since it will no longer be triggered.
llvm-svn: 324833
These were added as part of the refactoring for prefer vector width. At the time I thought the hasAVX512 here would be replaced with "allow 512 bit vectors" so that it would read "allow 512 bit vectors OR VLX". But now the plan is to only give the option of disabling 512 bit vectors when VLX is enabled. So we don't need this qualification at all
llvm-svn: 324831
Summary:
This patch changes the signature of the avx512 packed fp compare intrinsics to return a vXi1 vector and no longer take a mask as input. The casts to scalar type will now need to be explicit in the IR. The masking node will now be an explicit and in the IR.
This makes the intrinsic look much more similar to an fcmp instruction that we wish we could use for these but can't. We already use icmp instructions for integer compares.
Previously the lowering step of isel would turn the intrinsic into an X86 specific ISD node and a emit the masking nodes as well as some bitcasts. This means DAG combines can't see the vXi1 type until somewhat late, making it more difficult to combine out gpr<->mask transition sequences. By exposing the vXi1 type explicitly in the IR and initial SelectionDAG we give earlier DAG combines and even InstCombine the chance to see it and optimize it.
This should make any issues with gpr<->mask sequences the same between integer and fp. Meaning we only have to fix them once.
Reviewers: spatel, delena, RKSimon, zvi
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D43137
llvm-svn: 324827
Undef VLX, getSetCCResultType returns v2i1/v4i1 for v2f32/v4f32 so default type legalization will end up changing the setcc result type back to vXi1 if it had been extended. The resulting extend gets messed up further by type legalization and is difficult to recombine back to (v4i32 (setcc (v4f32))) after legalization.
I went ahead and enabled this for SSE2 and later since its always the result we want and this helps type legalization get there in less steps.
llvm-svn: 324822
This prevents extends of masks being introduced during lowering where it become difficult to combine them out.
There are a few oddities in here.
We sometimes concatenate two k-registers produced by two compares, sign_extend the combined pair, then extract two halves. This worked better previously because the sign_extend wasn't created until after the fp_to_sint was split which led to a split sign_extend being created.
We probably also need to custom type legalize (v2i32 (sext v2i1)) via widening.
llvm-svn: 324820
This avoids a constant pool load to create 1.
The int->float are showing converts to mask and back. We probably need to widen inputs to sint_to_fp/uint_to_fp before type legalization.
llvm-svn: 324805
Previously we extracted two subvectors and concatenate. But the concatenate will be lowered to two insert subvectors. Then DAG combine will merge once of the inserts and one of the extracts back into the original vector. We might as well just directly use one extract and one insert.
llvm-svn: 324710
This regresses a couple cases in the shuffle combining test. But those cases use intrinsics that InstCombine knows how to turn into a generic shuffle earlier. This should give opportunities to fold this earlier in InstCombine or DAG combine.
llvm-svn: 324709
Most vxi1 constant build vectors have to be implemented in the scalar domain anyway so we'll probably end up with a cast there later. But by then its too late to do the combine to get rid of it.
llvm-svn: 324662
The KTEST instruction sets the C flag if the result of anding both operands together is all 1s. We can use this to lower (icmp eq/ne (bitcast (vXi1 X), -1)
Differential Revision: https://reviews.llvm.org/D42772
llvm-svn: 324577
Summary:
KTEST has weird flag behavior. The Z flag is set for all bits in the AND of the k-registers being 0, and the C flag is set for all bits being 1. All other flags are cleared.
We currently emit this instruction in EmitTEST and don't check the condition code. This can lead to strange things like using the S flag after a KTEST for a signed compare.
The domain reassignment pass can also transform TEST instructions into KTEST and is not protected against the flag usage either. For now I've disabled this part of the domain reassignment pass. I tried to comment out the checks in the mir test so that we could recover them later, but I couldn't figure out how to get that to work.
This patch moves the KTEST handling into LowerSETCC and now creates a ktest+x86setcc. I've chosen this approach because I'd like to add support for the C flag for all ones in a followup patch. To do that requires that I can rewrite the condition code going in the x86setcc to be different than the original SETCC condition code.
This fixes PR36182. I'll file a PR to fix domain reassignment once this goes in. Should this be merged to 6.0?
Reviewers: spatel, guyblank, RKSimon, zvi
Reviewed By: guyblank
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D42770
llvm-svn: 324576
We were doing a lot of whitelisting of what we handle in these routines, but setOperationAction constrains what we can get here. So just add some asserts and prune the unreachable paths.
llvm-svn: 324538
that happened to end up in GCC.
This is really unfortunate, as the names don't have much rhyme or reason
to them. Originally in the discussions it seemed fine to rely on aliases
to map different names to whatever external thunk code developers wished
to use but there are practical problems with that in the kernel it turns
out. And since we're discovering this practical problems late and since
GCC has already shipped a release with one set of names, we are forced,
yet again, to blindly match what is there.
Somewhat rushing this patch out for the Linux kernel folks to test and
so we can get it patched into our releases.
Differential Revision: https://reviews.llvm.org/D42998
llvm-svn: 324449
X86 currently has a late DAG combine after cttz/ctlz are turned into BSR+BSF+CMOV to detect this and remove the CMOV. But we should be able to do this much earlier and avoid creating the cmov all together.
For the changed AMDGPU test case it appears that previously the i8 cttz was type legalized to i16 which introduced an OR with 256 in order to limit the result to 8 on the widened type. At this point the result is known to never be zero, but nothing checked that. Then operation legalization is told to promote all i16 cttz to i32. This introduces an extend and a truncate and another OR with 65536 to limit the result to 16. With the DAG combiner change we are able to prevent the creation of the second OR since the opcode will have been changed to cttz_zero_undef after the first OR. I the lack of the OR caused the instruction to change to v_ffbl_b32_sdwa
Differential Revision: https://reviews.llvm.org/D42985
llvm-svn: 324427
Followup to D42544 that matches PACKUSWB cases for non-AVX512, SSE and PACKUSDW cases will have to wait until we can add support for general SMIN/SMAX matching.
llvm-svn: 324347
Followup to D42544 that matches PACKSSWB cases for non-AVX512, SSE and PACKSSDW cases will have to wait until we can add support for general SMIN/SMAX matching.
llvm-svn: 324339
We now allow all signed comparisons and not equal. The complement that needs to be added for this is no worse than the extend. And the vector output forms of pcmpeq/pcmpgt have better latency than the k-register version on SKX.
llvm-svn: 324294
We always created X86ISD::SHUF128 with a 64-bit element type so we can use isel patterns to detect a bitconvert to 32-bit to handle masking.
The test changes are because we also match the bitconvert even if there is no masking. This leads to unnecessary isel pattern, but it requires more multiclass hackery in tablegen to get rid of it.
llvm-svn: 324205
This reduces the number of transitions between k-registers and GPRs, reducing the number of instructions.
There's still some room for improvement to remove more transitions, but this is a good start.
llvm-svn: 324184
Clang already stopped using these a couple months ago.
The test cases aren't great as there is nothing forcing the operations to stay in k-registers so some of them moved back to scalar ops due to the bitcasts being moved around.
llvm-svn: 324177
This is running pre-legalize, we should try to use target independent nodes. This will give the best opportunity for target independent optimizations.
llvm-svn: 324147
Every instruction that has the word TEST in its name seems to have been buried into EmitTest. But that code is largely concerned with trying to reuse the flags from instructions that update flags in a pretty normal way.
PTEST/TESTP/KTEST do not update flags in a normal way. They only update Z and C and the C flag update is non-standard. Rather than try to bend EmitTest's already complex logic to accomodate this, just move the call up to LowerSETCC and replicate the few pre-checks that are needed.
While there add a FIXME for using the C flag for checking for all 1s which we definitely couldn't do from EmitTEST.
llvm-svn: 324029
This allows us to use PSHUFB for v8i16/v4i32 and VPERMD/PERMPS for v4i64/v4f64 variable shuffles.
Differential Revision: https://reviews.llvm.org/D42487
llvm-svn: 323987
Summary: Now that v2i1/v4i1 are legal without VLX. And v32i1 is legalized by splitting rather than widening. And isVectorLoadExtDesirable returns false for vXi1. It appears this handling is dead because the operations simply don't exist.
Reviewers: RKSimon, zvi, guyblank, delena, spatel
Reviewed By: delena
Subscribers: llvm-commits, rengolin
Differential Revision: https://reviews.llvm.org/D42781
llvm-svn: 323983
Summary:
EmitTest sometimes creates X86ISD::AND specifically to hide the AND from DAG combine. But this prevents isel patterns that look for (cmp (and X, Y), 0) from being able to see it. So we end up with an AND and a TEST. The TEST gets removed by compare instruction optimization during the peephole pass.
This patch attempts to fix this by converting X86ISD::AND with no flag users back into ISD::AND during the DAG preprocessing just before isel.
In order to do this correctly I had to make the X86ISD::AND node created by EmitTest in this case really have a flag output. Which arguably it should have had anyway so that the number of operands would be consistent for the opcode in all cases. Then I had to modify the ReplaceAllUsesWith to understand that we might be looking at an instruction with 2 outputs. Though in this case there are no uses to replace since we just created the node, but that's what the code did before so I just made it keep working.
Reviewers: spatel, RKSimon, niravd, deadalnix
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D42764
llvm-svn: 323982
Summary:
This change expands the amount of registers stashed by the entry and
`__xray_CustomEvent` trampolines.
We've found that since the `__xray_CustomEvent` trampoline calls can show up in
situations where the scratch registers are being used, and since we don't
typically want to affect the code-gen around the disabled
`__xray_customevent(...)` intrinsic calls, that we need to save and restore the
state of even the scratch registers in the handling of these custom events.
Reviewers: pcc, pelikan, dblaikie, eizan, kpw, echristo, chandlerc
Reviewed By: echristo
Subscribers: chandlerc, echristo, hiraditya, davide, dblaikie, llvm-commits
Differential Revision: https://reviews.llvm.org/D40894
llvm-svn: 323940
This code currently uses isSimple and getSizeInBits in an attempt to prune types. But isSimple will return true for any type that any target supports natively. I don't think that's a good way to prune types. I also don't think the dest element type checks are very robust since we didn't do an isSimple check on the dest type.
This patch adds a check for the input type being legal to the one caller that didn't already check that. Then we explicitly check the element types for the destination are i8, i16, or i32
Differential Revision: https://reviews.llvm.org/D42706
llvm-svn: 323924
Similar to D42437, XOP supports variable shift for v16i8/v8i16/v4i32/v2i64 types.
Differential Revision: https://reviews.llvm.org/D42526
llvm-svn: 323797
We can use the same input for both operands to get a free compare with zero.
We already use this trick in a couple places where we explicitly create PTESTM with the same input twice. This generalizes it.
I'm hoping to remove the ISD opcodes and move this to isel patterns like we do for scalar cmp/test.
llvm-svn: 323605
Legalization is still biased to turn LT compares in to GT by swapping operands to avoid needing extra isel patterns to commute.
I'm hoping to remove TESTM/TESTNM next and this should simplify that by making EQ/NE more similar.
llvm-svn: 323604
If broadcasting from another shuffle, attempt to simplify it.
We can probably generalize this a lot more (embedding in combineX86ShufflesRecursively), but BROADCAST is one of the more troublesome as it accepts inputs of different sizes to the result.
llvm-svn: 323602
X86ISelLowering.cpp:34130:5: error: return type 'llvm::SDValue' must
match previous return type 'const llvm::SDValue' when lambda expression
has unspecified explicit return type
llvm-svn: 323557
We currently coalesce v4i32 extracts from all 4 elements to 2 v2i64 extracts + shifts/sign-extends.
This seems to have been added back in the days when we tended to spill vectors and reload scalars, or ended up with repeated shuffles moving everything down to 0'th index. I don't think either of these are likely these days as we have better EXTRACT_VECTOR_ELT and VECTOR_SHUFFLE handling, and the existing code tends to make it very difficult for various vector and load combines.
Differential Revision: https://reviews.llvm.org/D42308
llvm-svn: 323541
Type legalization would prevent any i64 operands to the build_vector from existing before we get here. The coverage bots show this code as uncovered.
llvm-svn: 323506
The original autoupgrade for kunpck intrinsics used a bitcasted scalar shift, or, and. This combine would turn this into a concat_vectors. Now the kunpck intrinsics are autoupgraded to a vector shuffle that will become a concat_vectors.
llvm-svn: 323504
This listed all legal 128-bit integer types individually, but since we already know we have a legal type and its integer, we can just check is128BitVector.
llvm-svn: 323502
As discussed in D41484, PMADDWD for 'zero extended' vXi32 is nearly always a better option than PMULLD:
On SNB it will result in code that isn't any faster, but not any slower so we may as well keep it.
On KNL it only has half the throughput, so I've disabled it on there - ideally there'd be a better way than this.
Differential Revision: https://reviews.llvm.org/D42258
llvm-svn: 323367
There are a couple tricky things with this patch.
I had to add an override of isVectorLoadExtDesirable to stop DAG combine from combining sign_extend with loads after legalization since we legalize sextload using a load+sign_extend. Overriding this hook actually prevents a lot sextloads from being created in the first place.
I also had to add isel patterns because DAG combine blindly combines sign_extend+truncate to a smaller sign_extend which defeats what legalization was trying to do.
Differential Revision: https://reviews.llvm.org/D42407
llvm-svn: 323301
Minor refactor to make it possible for LowerBUILD_VECTORAsVariablePermute to be used with a wider variety of shuffles op and types.
I'd have liked to add v4i32/v4f32 support as well but we don't see v4i32 index extractions at the moment (which is why I created D42308)
After this I intend to begin adding scaling support for PSHUFB (v8i16, v4i32, v2i64)) and VPERMPS (v4f64, v4i64).
Differential Revision: https://reviews.llvm.org/D42431
llvm-svn: 323260
The existing code was already doing something very similar to subvector insertion so this allows us to remove the nearly duplicate code.
This patch is a little larger than it should be due to differences between the DQI handling between the two today.
llvm-svn: 323212
Summary:
For the most part its better to keep v32i1 as a mask type of a narrower width than trying to promote it to a ymm register.
I had to add some overrides to the methods that get the types for the calling convention so that we still use v32i8 for argument/return purposes.
There are still some regressions in here. I definitely saw some around shuffles. I think we probably should move vXi1 shuffle from lowering to a DAG combine where I think the extend and truncate we have to emit would be better combined.
I think we also need a DAG combine to remove trunc from (extract_vector_elt (trunc))
Overall this removes something like 13000 CHECK lines from lit tests.
Reviewers: zvi, RKSimon, delena, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D42031
llvm-svn: 323201
As detailed in rL317463, PSHUFB (like most variable shuffle instructions) uses Op[0] for the source vector and Op[1] for the shuffle index vector, VPERMV works in reverse which is probably where the confusion comes from.
Differential Revision: https://reviews.llvm.org/D42380
llvm-svn: 323190
Summary:
If we can match as a zero extend there's no need to flip the order to get an encoding benefit. As movzx is 3 bytes with independent source/dest registers. The shortest 'and' we could make is also 3 bytes unless we get lucky in the register allocator and its on AL/AX/EAX which have a 2 byte encoding.
This patch was more impressive before r322957 went in. It removed some of the same Ands that got deleted by that patch.
Reviewers: spatel, RKSimon
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D42313
llvm-svn: 323175
Add missing patterns for inserting v1i1 into a zero vector. Use insert_subvector to zero upper bits before inserting an element into a vXi1 vector. Replace kshift based isel pattern with insert_subvector based pattern now that code that caused the pattern has been fixed to emit insert_subvector.
llvm-svn: 323173
Summary:
First, we need to explain the core of the vulnerability. Note that this
is a very incomplete description, please see the Project Zero blog post
for details:
https://googleprojectzero.blogspot.com/2018/01/reading-privileged-memory-with-side.html
The basis for branch target injection is to direct speculative execution
of the processor to some "gadget" of executable code by poisoning the
prediction of indirect branches with the address of that gadget. The
gadget in turn contains an operation that provides a side channel for
reading data. Most commonly, this will look like a load of secret data
followed by a branch on the loaded value and then a load of some
predictable cache line. The attacker then uses timing of the processors
cache to determine which direction the branch took *in the speculative
execution*, and in turn what one bit of the loaded value was. Due to the
nature of these timing side channels and the branch predictor on Intel
processors, this allows an attacker to leak data only accessible to
a privileged domain (like the kernel) back into an unprivileged domain.
The goal is simple: avoid generating code which contains an indirect
branch that could have its prediction poisoned by an attacker. In many
cases, the compiler can simply use directed conditional branches and
a small search tree. LLVM already has support for lowering switches in
this way and the first step of this patch is to disable jump-table
lowering of switches and introduce a pass to rewrite explicit indirectbr
sequences into a switch over integers.
However, there is no fully general alternative to indirect calls. We
introduce a new construct we call a "retpoline" to implement indirect
calls in a non-speculatable way. It can be thought of loosely as
a trampoline for indirect calls which uses the RET instruction on x86.
Further, we arrange for a specific call->ret sequence which ensures the
processor predicts the return to go to a controlled, known location. The
retpoline then "smashes" the return address pushed onto the stack by the
call with the desired target of the original indirect call. The result
is a predicted return to the next instruction after a call (which can be
used to trap speculative execution within an infinite loop) and an
actual indirect branch to an arbitrary address.
On 64-bit x86 ABIs, this is especially easily done in the compiler by
using a guaranteed scratch register to pass the target into this device.
For 32-bit ABIs there isn't a guaranteed scratch register and so several
different retpoline variants are introduced to use a scratch register if
one is available in the calling convention and to otherwise use direct
stack push/pop sequences to pass the target address.
This "retpoline" mitigation is fully described in the following blog
post: https://support.google.com/faqs/answer/7625886
We also support a target feature that disables emission of the retpoline
thunk by the compiler to allow for custom thunks if users want them.
These are particularly useful in environments like kernels that
routinely do hot-patching on boot and want to hot-patch their thunk to
different code sequences. They can write this custom thunk and use
`-mretpoline-external-thunk` *in addition* to `-mretpoline`. In this
case, on x86-64 thu thunk names must be:
```
__llvm_external_retpoline_r11
```
or on 32-bit:
```
__llvm_external_retpoline_eax
__llvm_external_retpoline_ecx
__llvm_external_retpoline_edx
__llvm_external_retpoline_push
```
And the target of the retpoline is passed in the named register, or in
the case of the `push` suffix on the top of the stack via a `pushl`
instruction.
There is one other important source of indirect branches in x86 ELF
binaries: the PLT. These patches also include support for LLD to
generate PLT entries that perform a retpoline-style indirection.
The only other indirect branches remaining that we are aware of are from
precompiled runtimes (such as crt0.o and similar). The ones we have
found are not really attackable, and so we have not focused on them
here, but eventually these runtimes should also be replicated for
retpoline-ed configurations for completeness.
For kernels or other freestanding or fully static executables, the
compiler switch `-mretpoline` is sufficient to fully mitigate this
particular attack. For dynamic executables, you must compile *all*
libraries with `-mretpoline` and additionally link the dynamic
executable and all shared libraries with LLD and pass `-z retpolineplt`
(or use similar functionality from some other linker). We strongly
recommend also using `-z now` as non-lazy binding allows the
retpoline-mitigated PLT to be substantially smaller.
When manually apply similar transformations to `-mretpoline` to the
Linux kernel we observed very small performance hits to applications
running typical workloads, and relatively minor hits (approximately 2%)
even for extremely syscall-heavy applications. This is largely due to
the small number of indirect branches that occur in performance
sensitive paths of the kernel.
When using these patches on statically linked applications, especially
C++ applications, you should expect to see a much more dramatic
performance hit. For microbenchmarks that are switch, indirect-, or
virtual-call heavy we have seen overheads ranging from 10% to 50%.
However, real-world workloads exhibit substantially lower performance
impact. Notably, techniques such as PGO and ThinLTO dramatically reduce
the impact of hot indirect calls (by speculatively promoting them to
direct calls) and allow optimized search trees to be used to lower
switches. If you need to deploy these techniques in C++ applications, we
*strongly* recommend that you ensure all hot call targets are statically
linked (avoiding PLT indirection) and use both PGO and ThinLTO. Well
tuned servers using all of these techniques saw 5% - 10% overhead from
the use of retpoline.
We will add detailed documentation covering these components in
subsequent patches, but wanted to make the core functionality available
as soon as possible. Happy for more code review, but we'd really like to
get these patches landed and backported ASAP for obvious reasons. We're
planning to backport this to both 6.0 and 5.0 release streams and get
a 5.0 release with just this cherry picked ASAP for distros and vendors.
This patch is the work of a number of people over the past month: Eric, Reid,
Rui, and myself. I'm mailing it out as a single commit due to the time
sensitive nature of landing this and the need to backport it. Huge thanks to
everyone who helped out here, and everyone at Intel who helped out in
discussions about how to craft this. Also, credit goes to Paul Turner (at
Google, but not an LLVM contributor) for much of the underlying retpoline
design.
Reviewers: echristo, rnk, ruiu, craig.topper, DavidKreitzer
Subscribers: sanjoy, emaste, mcrosier, mgorny, mehdi_amini, hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D41723
llvm-svn: 323155
Primarily, this allows us to use the aggressive extraction mechanisms in combineExtractWithShuffle earlier and make use of UNDEF elements that may be lost during lowering.
Reapplied after rL322279 was reverted at rL322335 due to PR35918, underlying issue was fixed at rL322644.
llvm-svn: 323104
1. ReachingDefsAnalysis - Allows to identify for each instruction what is the “closest” reaching def of a certain register. Used by BreakFalseDeps (for clearance calculation) and ExecutionDomainFix (for arbitrating conflicting domains).
2. ExecutionDomainFix - Changes the variant of the instructions in order to minimize domain crossings.
3. BreakFalseDeps - Breaks false dependencies.
4. LoopTraversal - Creatws a traversal order of the basic blocks that is optimal for loops (introduced in revision L293571). Both ExecutionDomainFix and ReachingDefsAnalysis use this to determine the order they will traverse the basic blocks.
This also included the following changes to ExcecutionDepsFix original logic:
1. BreakFalseDeps and ReachingDefsAnalysis logic no longer restricted by a register class.
2. ReachingDefsAnalysis tracks liveness of reg units instead of reg indices into a given reg class.
Additional changes in affected files:
1. X86 and ARM targets now inherit from ExecutionDomainFix instead of ExecutionDepsFix. BreakFalseDeps also was added to the passes they activate.
2. Comments and references to ExecutionDepsFix replaced with ExecutionDomainFix and BreakFalseDeps, as appropriate.
Additional refactoring changes will follow.
This commit is (almost) NFC.
The only functional change is that now BreakFalseDeps will break dependency for all register classes.
Since no additional instructions were added to the list of instructions that have false dependencies, there is no actual change yet.
In a future commit several instructions (and tests) will be added.
This is the first of multiple patches that fix bugzilla https://bugs.llvm.org/show_bug.cgi?id=33869
Most of the patches are intended at refactoring the existent code.
Additional relevant reviews:
https://reviews.llvm.org/D40331https://reviews.llvm.org/D40332https://reviews.llvm.org/D40333https://reviews.llvm.org/D40334
Differential Revision: https://reviews.llvm.org/D40330
Change-Id: Icaeb75e014eff96a8f721377783f9a3e6c679275
llvm-svn: 323087
Summary:
This patch adds an implementation of targetShrinkDemandedConstant that tries to keep shrinkdemandedbits from removing bits that would otherwise have been recognized as a movzx.
We still need a follow patch to stop moving ands across srl if the and could be represented as a movzx before the shift but not after. I think this should help with some of the cases that D42088 ended up removing during isel.
Reviewers: spatel, RKSimon
Reviewed By: spatel
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D42265
llvm-svn: 323048
This change applies to places where we would turn 128/256-bit code into 512-bit in order to get a wider element type through sext/zext. Any 512-bit types that already existed in the IR/DAG will be left that way.
The width preference has no effect on codegen behavior when the target does not have AVX512 enabled. So AVX/AVX2 codegen cannot be limited via this mechanism yet.
If the preference is lower than 256 we may still use a 256 bit type to do the operation. Constraining to 128 bits makes it much more difficult to support some operations. For many of these cases we need to change element width while keeping element count constant which is easiest done by switching between 256 and 128 bit.
The preference is only obeyed when AVX512 and VLX are available. This means the preference is not obeyed for KNL, but is obeyed for SKX, Cannonlake, and Icelake. For KNL, the only way to do masked operation is on 512-bit registers so we would have to completely disable masking to obey the preference. We would also lose support for gather, scatter, ctlz, vXi64 multiplies, etc. This may change in the future, but this simplifies the initial implementation.
Differential Revision: https://reviews.llvm.org/D41895
llvm-svn: 323016
If we are splatting pairs of 32-bit elements, we can use a 64-bit broadcast to get the job done.
We could probably could probably do this with other sizes too, for example four 16-bit elements. Or we could broadcast pairs of 16-bit elements using a 32-bit element broadcast. But I've left that as a future improvement.
I've also restricted this to AVX2 only because we can only broadcast loads under AVX.
Differential Revision: https://reviews.llvm.org/D42086
llvm-svn: 322730
We legalize selects of masks with scalar conditions using a bitcast to an integer type. But if we are in 32-bit mode we can't convert v64i1 to i64. So instead split the v64i1 to v32i1 and concat it back together. Each half will then be legalized by bitcasting to i32 which is fine.
The test case is a little indirect. If we have the v64i1 select in IR it will get legalized by legalize vector ops which has a run of type legalization after it. That type legalization run is able to fix this i64 bitcast. So in order to avoid that we need a build_vector of a splat which legalize vector ops will ignore. Legalize DAG will then turn that into a select via LowerBUILD_VECTORvXi1. And the select will get legalized. In this case there is no type legalizer run to cleanup the bitcast.
This fixes pr35972.
llvm-svn: 322724
The match* functions have the annoying behavior of modifying its inputs.
Save and restore the inputs, just in case the early out for AVX512 is
hit. This is still not great and its only a matter of time this kind of
bug happens again, but I couldn't come up with a better pattern without
rewriting significant chunks of this code. Fixes PR35977.
llvm-svn: 322644
As mentioned on PR35869, (and came up recently on D41517) we don't create a MMX zero register via the PXOR but instead perform a spill to stack from a XMM zero register.
This patch adds support for direct MMX zero vector creation and should make it easier to add better constant vector creation in the future as well.
Differential Revision: https://reviews.llvm.org/D41908
llvm-svn: 322525
We have to take special care to avoid the cases where the result of the truncate would be padded with zero elements.
Ideally we'd just use ISD::TRUNCATE for these cases instead.
llvm-svn: 322454
Extend vXi1 conditions of vXi8/vXi16 selects even before type legalization gets a chance to split wide vectors. Previously we would only extend 128 and 256 bit vectors. But if we start with a 512 bit vector or wider that needs to be split we wouldn't extend until after the split had taken place. By extending early we improve the results of type legalization.
Don't widen condition of 128/256 bit vXi16/vXi8 selects when we have BWI but not VLX. We can still use a mask register by widening the select to 512-bits instead. This is similar to what we do for compares already.
llvm-svn: 322450
In addition to the existing match as part of a loop-reduction, add a
straightforward pattern match for DAG-contained patterns.
Reviewers: RKSimon, craig.topper
Subscribers: llvm-commits
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D41811
llvm-svn: 322446
This avoids having the result type stick around until lowering where we have to extend the setcc and insert a truncate. If we get the types converted early we can do more to optimize it.
llvm-svn: 322432
Summary: This is a preparatory step for D41811: refactoring code for breaking vector operands of binary operation to legal-types.
Reviewers: RKSimon, craig.topper, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D41925
llvm-svn: 322296
Primarily, this allows us to use the aggressive extraction mechanisms in combineExtractWithShuffle earlier and make use of UNDEF elements that may be lost during lowering.
llvm-svn: 322279
Summary:
As RKSimon suggested in pr35820, in the case that Src is smaller in
bit-size than Indices, need to widen Src to avoid type mismatch.
Fixes pr35820
Reviewers: RKSimon, craig.topper
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D41865
llvm-svn: 322272
If the index is v2i64 we can use the scatter instruction that has v4i32/v4f32 data register, v2i64 index, and v2i1 mask. Similar was already done for gather.
Implement custom widening for v2i32 data to remove the code that reverses type legalization during lowering.
llvm-svn: 322254
Currently we infer the scale at isel time by analyzing whether the base is a constant 0 or not. If it is we assume scale is 1, else we take it from the element size of the pass thru or stored value. This seems a little weird and I think it makes more sense to make it explicit in the DAG rather than doing tricky things in the backend.
Most of this patch is just making sure we copy the scale around everywhere.
Differential Revision: https://reviews.llvm.org/D40055
llvm-svn: 322210
Normally target independent DAG combine would do this combine based on getSetCCResultType, but with VLX getSetCCResultType returns a vXi1 type preventing the DAG combining from kicking in.
But doing this combine can allow us to remove the explicit sign extend that would otherwise be emitted.
This patch adds a target specific DAG combine to combine the sext+setcc when the result type is the same size as the input to the setcc. I've restricted this to FP compares and things that can be represented with PCMPEQ and PCMPGT since we don't have full integer compare support on the older ISAs.
Differential Revision: https://reviews.llvm.org/D41850
llvm-svn: 322101
I had to drop fast-isel-abort from a test because we can't fast isel some of the mask stuff. When we used intrinsics we implicitly fell back to SelectionDAG for the intrinsic call without triggering the abort error. But with native IR that doesn't happen the same way.
llvm-svn: 322050
Summary:
There are few oddities that occur due to v1i1, v8i1, v16i1 being legal without v2i1 and v4i1 being legal when we don't have VLX. Particularly during legalization of v2i32/v4i32/v2i64/v4i64 masked gather/scatter/load/store. We end up promoting the mask argument to these during type legalization and then have to widen the promoted type to v8iX/v16iX and truncate it to get the element size back down to v8i1/v16i1 to use a 512-bit operation. Since need to fill the upper bits of the mask we have to fill with 0s at the promoted type.
It would be better if we could just have the v2i1/v4i1 types as legal so they don't undergo any promotion. Then we can just widen with 0s directly in a k register. There are no real v4i1/v2i1 instructions anyway. Everything is done on a larger register anyway.
This also fixes an issue that we couldn't implement a masked vextractf32x4 from zmm to xmm properly.
We now have to support widening more compares to 512-bit to get a mask result out so new tablegen patterns got added.
I had to hack the legalizer for widening the operand of a setcc a bit so it didn't try create a setcc returning v4i32, extract from it, then try to promote it using a sign extend to v2i1. Now we create the setcc with v4i1 if the original setcc's result type is v2i1. Then extract that and don't sign extend it at all.
There's definitely room for improvement with some follow up patches.
Reviewers: RKSimon, zvi, guyblank
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D41560
llvm-svn: 321967
This custom inserter was added in r124272 at which time it added about bunch of Defs for Win64. In r150708, those defs were removed leaving only the "return BB". So I think this means the custom inserter is a NOP these days.
This patch removes the remaining code and stops tagging the instructions for custom insertion
Differential Revision: https://reviews.llvm.org/D41671
llvm-svn: 321747
Currently we use SIGN_EXTEND in lowerMasksToReg as part of calling convention setup, but we don't require a specific value for the upper bits.
This patch changes it to ANY_EXTEND which will be lowered as SIGN_EXTEND if it ends up sticking around.
llvm-svn: 321746
This is an extension of D31156 with the goal that we'll allow memcmp() == 0 expansion
for x86 to use 2 pairs of loads per block.
The memcmp expansion pass (formerly part of CGP) will generate this kind of pattern
with oversized integer compares, so we want to transform these into x86-specific vector
nodes before legalization splits things into scalar chunks.
See PR33325 for more details:
https://bugs.llvm.org/show_bug.cgi?id=33325
Differential Revision: https://reviews.llvm.org/D41618
llvm-svn: 321656
Currently the promotion for these ignores the normal getTypeToPromoteTo and instead just tries to double the element width. This is because the default behavior of getTypeToPromote to just adds 1 to the SimpleVT, which has the affect of increasing the element count while keeping the scalar size the same.
If multiple steps are required to get to a legal operation type, int_to_fp will be promoted multiple times. And fp_to_int will keep trying wider types in a loop until it finds one that works.
getTypeToPromoteTo does have the ability to query a promotion map to get the type and not do the increasing behavior. It seems better to just let the target specify the promotion type in the map explicitly instead of letting the legalizer iterate via widening.
FWIW, it's worth I think for any other vector operations that need to be promoted, we have to specify the type explicitly because the default behavior of getTypeToPromote isn't useful for vectors. The other types of promotion already require either the element count is constant or the total vector width is constant, but neither happens by incrementing the SimpleVT enum.
Differential Revision: https://reviews.llvm.org/D40664
llvm-svn: 321629
The CONCAT_VECTORS will be lowered to INSERT_SUBVECTOR later. In the modified cases this seems to be enough to trick a later DAG combine into running in a different order than allows the ANDs to be removed.
I'll admit this is a bit of a hack that happens to work, but using CONCAT_VECTORS is more consistent with other legalization code anyway.
llvm-svn: 321611
Don't combine buildvector(binop(),binop(),binop(),binop()) -> binop(buildvector(), buildvector()) if its a splat - keep the binop scalar and just splat the result to avoid large vector constants.
llvm-svn: 321607
We end up using an i8 load via an isel pattern from v8i1 anyway. This just makes it more explicit. This seems to improve codgen in some cases and I'd like to kill off some of the load patterns.
llvm-svn: 321598
As noted in PR34686, we are relying on a PSHUFD+PSHUFLW+PSHUFHW shuffle chain for most general vXi16 unary shuffles.
This patch checks for simpler PSHUFLW+PSHUFD and PSHUFHW+PSHUFD cases beforehand, building on some existing code that just handled splat shuffles.
By doing so we also prevent premature use of PSHUFB shuffles which can be slower and require the creation/loading of constant shuffle masks.
We now have the 'fast-variable-shuffle' option for hardware that prefers combining 2 or more shuffles to VPSHUFB etc.
Differential Revision: https://reviews.llvm.org/D38318
llvm-svn: 321553
Previously we used an extend from v8i1 to v8i32/v8i64. Then extracted to the final width. But if we have VLX we should extract first. This way we don't end up with an overly large extend.
This allows us to use vcmpeq to make all ones for the sign extend when DQI isn't available. Otherwise we get a VPTERNLOG.
If we make v2i1/v4i1 legal like proposed in D41560, we could always do this and rely on the lowering of the extend to widen when necessary.
llvm-svn: 321538
-Use MinAlign instead of std::min.
-Use SelectionDAG::getMemBasePlusOffset.
-Apply offset to the pointer info for the second load/store created.
llvm-svn: 321536
If there are 17 or more leading zeros to the v4i32 elements, then we can use PMADD for the integer multiply when PMULLD is unavailable or slow.
The 17 bits need to be zero as the PMADDWD performs a v8i16 signed-mul-extend + pairwise-add - the upper 16 so we're adding a zero pair and the 17th bit so we don't incorrectly sign extend.
Differential Revision: https://reviews.llvm.org/D41484
llvm-svn: 321516
My original implementation ran as a DAG combine post type legalization, but it turns out we don't run that DAG combine step if type legalization didn't change anything. Attempts to make the combine run before type legalization as well hit other issues.
So just do it in LowerMUL where we can catch more cases.
llvm-svn: 321496
Returning SDValue() means nothing changed, SDValue(N,0) means there was a change but the worklist management was taken care of.
I don't know if this has a real effect other than making sure the combine counter in the DAG combiner gets updated, but it is the correct thing to do.
llvm-svn: 321463
Simon Pilgrim