Refactor towards making this recursive (necessary for PR38243 rotation splat detection).
IsSplatVector returns the original vector source of the splat and the splat index.
GetSplatValue returns the scalar splatted value as an extraction from IsSplatVector.
llvm-svn: 347168
We were using the 'normalized' shuffle mask from resolveTargetShuffleInputs, which replaces zero/undef inputs with sentinel values. For SimplifyDemandedVectorElts we need the raw mask so we can correctly demand those 'zero' inputs that got normalized away, this requires an extra bit of logic to locally normalize undef inputs.
llvm-svn: 347158
The zero extend will require two stages of unpacks to implement. So its better to shrink the multiply using pmullw and then extend that result back to v4i32 using a single unpack.
llvm-svn: 347149
This tries to force the result type to vXi32 followed by a truncate. This can help avoid scalarization that would otherwise occur.
There's some annoying examples of an avx512 truncate instruction followed by a packus where we should really be able to just use one truncate. But overall this is still a net improvement.
llvm-svn: 347105
By early promoting the multiply to use an i16 element type we can avoid op legalization emit a second multiply for the 8 upper elements of the v16i8 type we would otherwise get.
llvm-svn: 347032
We aren't going to use the upper bits of the multiply result that the extend would effect. So we don't need a specific type of extend.
This makes some reduction test cases shorter because we were previously trying to sign_extend a truncate which we can't eliminate.
llvm-svn: 347011
Removing this code doesn't affect any lit tests so it doesn't appear to be tested anymore. I assume it was when it was added, but I guess something else changed? Code coverage report also says its unused.
I mostly didn't like that it seemed to count the sign bits as if it was a sign_extend, but then set isPositive as if it was a zero_extend. It feels like we should have picked one interpretation?
Differential Revision: https://reviews.llvm.org/D54596
llvm-svn: 346995
Use unsigned to calculate the subvector index to avoid a cast.
Remove an unnecessary condition and replace it with a stronger assert.
Use the InVT variable we updated when we extracted instead of grabbing it from the In SDValue.
llvm-svn: 346983
In reduceVMULWidth, we no longer need to worry about extending the vector to 128 bits first. Regular widening of extends, muls and shuffles will take care of that for us.
In combineMulToPMADDWD, we can handle v2i32 multiplies and allow the VPMADDWD to be widened to v4i32 during type legalization by adding custom widening like we do have for AVG/ADDUS/SUBUS. I had to modify that code a little to allow different and output VTs.
Differential Revision: https://reviews.llvm.org/D54512
llvm-svn: 346980
This avoids some nasty shuffles when we have avx512. It will also prevent using zmm truncate instructions when a ymm instruction that zeroes part of an xmm register will do. Also avoid using avx512 truncate instructions when the input is 128 bits or less. These instructions are 2 uops on skx so we can probably find a better single uop shuffle like pshufb.
llvm-svn: 346936
The narrow types end up requesting widening, but generic legalization will end up scalaring and using a build_vector to do the widening.
llvm-svn: 346916
On 64-bit targets the type legalizer will use i64 to legalize these. But when i64 isn't legal, the type legalizer won't try an FP type. So do it manually instead.
There are a few regressions in here due to some v2i32 operations like mul and div now being reassembled into a full vector just to store instead of storing the pieces. But this was already occuring in 64-bit mode so its not a new issue.
llvm-svn: 346908
Narrower vectors will be widened to 128 bits without changing the element size. And generic type legalization can already handle widening mulhu/mulhs.
Differential Revision: https://reviews.llvm.org/D54513
llvm-svn: 346879
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
Simon Pilgrim