Drop the pow2 vector limitation for AVG generation by padding the vector to the next pow2, creating the PAVG nodes and then extracting the final subvector.
Fixes some poor codegen that has been annoying me for years.....
This patch uses partial DemandedElts masks to further simplify target shuffle chains and finally starts making target shuffle combining part of SimplifyDemandedBits/SimplifyDemandedVectorElts.
We already manage this for Depth == 0 cases, where combineX86ShuffleChain would early-out if the shuffle combined to the same op, but the patch generalizes this by manipulating the depth handling of combineX86ShufflesRecursively - calling with a new Depth = 0 and reducing the maximum shuffle combine depth accordingly.
Differential Revision: https://reviews.llvm.org/D66004
If the mask input to getV4X86ShuffleImm8 only refers to a single source element (+ undefs) then canonicalize to a full broadcast.
getV4X86ShuffleImm8 defaults to inline values for undefs, which can be useful for shuffle widening/narrowing but does leave SimplifyDemanded* calls thinking the shuffle depends on unnecessary elements.
I'm still investigating what we should do more generally to avoid these undemanded elements, but broadcast cases was a simpler win.
If we lower a v2i64 shuffle to PSHUFD, we currently clamp undef elements to 0, (elements 0,1 of the v4i32) which can result in the shuffle referencing more elements of the source vector than expected, affecting later shuffle combines and KnownBits/SimplifyDemanded calls.
By ensuring we widen the undef mask element we allow getV4X86ShuffleImm8 to use inline elements as the default, which are more likely to fold.
Truncations lowered as shuffles of multiple (concatenated) vectors often leave us with lane-crossing shuffles that feed a PACKSS/PACKUS, if both shuffles are fed from the same 2 vector sources, then we can PACK the sources directly and shuffle the result instead.
This is currently limited to whole i128 lanes in a 256-bit vector, but we can extend this if the need arises (but I'm not seeing many examples in real world code).
We generate PACK instructions with an undef second source when we are truncating from a 128-bit vector to something narrower and we don't care about the upper bits of the vector register. The register allocation process will always assign untied undef uses to xmm0. This creates a false dependency on xmm0.
By adding these instructions to hasUndefRegUpdate, we can get the BreakFalseDeps pass to reassign the source to match the other input. Normally this interface is used for instructions that might need an xor inserted to break the dependency. But the pass also has a heuristic that tries to use the same register as other sources. That should always be possible for these instructions so we'll never trigger the xor dependency break.
Differential Revision: https://reviews.llvm.org/D79032
This moves v32i16/v64i8 to a model consistent with how we
treat integer types with avx1.
This does change the ABI for types vXi16/vXi8 vectors larger than
512 bits to pass in multiple zmms instead of multiple ymms. We'd
already hacked some code to make v64i8/v32i16 pass in zmm.
Cost model is still a bit of a mess. In some place I tried to
match existing behavior. But really we need to account for
splitting and concating costs. Cost model for shuffles is
especially pessimistic.
Differential Revision: https://reviews.llvm.org/D76212
We can improve computeKnownBits results by avoiding excess bitcasts.
For this pattern we were doing:
(v16i8 PACKUS(v8i16 BITCAST(v16i8 AND(V1, MASK)), v8i16 BITCAST(v16i8 AND(V2, MASK))))
By performing the MASK/AND with a v8i16 type and bitcasting V1/V2 directly we can help computeKnownBits see that the mask is clearing the upper bits and allows shuffle combining to peek through later on.
This will be necessary to extend rG9d1721ce3926 to AVX2+ targets in a future patch.
If we're extracting the 0'th index of a v16i8 vector we're better off using MOVD than PEXTRB, unless we're storing the value or we require the implicit zero extension of PEXTRB.
The biggest perf diff is on SLM targets where MOVD (uops=1, lat=3 tp=1) is notably faster than PEXTRB (uops=2, lat=5, tp=4).
This matches what we already do for PEXTRW.
Differential Revision: https://reviews.llvm.org/D76138
This patch reuses the existing MatchRotate ROTL/ROTR rotation pattern code to also recognize the more general FSHL/FSHR funnel shift patterns when we have constant shift amounts.
Differential Revision: https://reviews.llvm.org/D75114
For pre-AVX512 targets, combine binary shuffles to X86ISD::VPERM2X128 if possible. This mainly helps optimize the blend(extract_subvector(x,1),y) pattern.
At some point soon we're going to have make a decision about when to combine AVX512 shuffles more aggressively - we bail out if there is any change in element size (to protect predicate mask merging) which means we miss out on a lot of optimizations.
If we go with D75412, we no longer depend on the scalar type directly. So we don't need to avoid using i64. We already have AVX1 fallback patterns with i32 and i64 scalar types so we don't need to avoid using integer types on AVX1.
Differential Revision: https://reviews.llvm.org/D75413
The unseen logic diff occurs because MayFoldLoad() is defined like this:
static bool MayFoldLoad(SDValue Op) {
return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
}
The test diffs here all seem ok to me on screen/paper, but it's hard to know
if that will lead to universally better perf for all targets. For example,
if a target implements broadcast from mem as multiple uops, we would have to
weigh the potential reduction of instructions and register pressure vs.
possible increase in number of uops. I don't know if we can make a truly
informed decision on this at compile-time.
The motivating case that I'm looking at in PR42024:
https://bugs.llvm.org/show_bug.cgi?id=42024
...resembles the diff in extract-concat.ll, but we're not going to change the
larger example there without at least 1 other fix.
Differential Revision: https://reviews.llvm.org/D74088
If we are extracting a chunk of a vector that's a fraction of an
operand of the concatenated vector operand, we can extract directly
from one of those original operands.
This is another suggestion from PR42024:
https://bugs.llvm.org/show_bug.cgi?id=42024#c2
But I'm not sure yet if it will make any difference on those patterns.
It seems to help a few existing AVX512 tests though.
Differential Revision: https://reviews.llvm.org/D72361
gcc and icc pass these types in zmm registers in zmm registers.
This patch implements a quick hack to override the register
type before calling convention handling to one that is legal.
Longer term we might want to do something similar to 256-bit
integer registers on AVX1 where we just split all the operations.
Fixes PR42957
Differential Revision: https://reviews.llvm.org/D66708
llvm-svn: 370495
The motivating case are the changes in vector-reduce-add.ll where
we were doing extra work in the scalar domain instead of shuffling.
There may be some one use check that needs to be looked into there,
but this patch sidesteps the issue by avoiding broadcasts that
aren't really broadcasting.
Differential Revision: https://reviews.llvm.org/D66071
llvm-svn: 369287
The assert that caused this to be reverted should be fixed now.
Original commit message:
This patch changes our defualt legalization behavior for 16, 32, and
64 bit vectors with i8/i16/i32/i64 scalar types from promotion to
widening. For example, v8i8 will now be widened to v16i8 instead of
promoted to v8i16. This keeps the elements widths the same and pads
with undef elements. We believe this is a better legalization strategy.
But it carries some issues due to the fragmented vector ISA. For
example, i8 shifts and multiplies get widened and then later have
to be promoted/split into vXi16 vectors.
This has the potential to cause regressions so we wanted to get
it in early in the 10.0 cycle so we have plenty of time to
address them.
Next steps will be to merge tests that explicitly test the command
line option. And then we can remove the option and its associated
code.
llvm-svn: 368183
This patch changes our defualt legalization behavior for 16, 32, and
64 bit vectors with i8/i16/i32/i64 scalar types from promotion to
widening. For example, v8i8 will now be widened to v16i8 instead of
promoted to v8i16. This keeps the elements widths the same and pads
with undef elements. We believe this is a better legalization strategy.
But it carries some issues due to the fragmented vector ISA. For
example, i8 shifts and multiplies get widened and then later have
to be promoted/split into vXi16 vectors.
This has the potential to cause regressions so we wanted to get
it in early in the 10.0 cycle so we have plenty of time to
address them.
Next steps will be to merge tests that explicitly test the command
line option. And then we can remove the option and its associated
code.
llvm-svn: 367901
The build_vector will become a constant pool load. By using the
desired type initially, it ensures we don't generate a bitcast
of the constant pool load which will need to be folded with
the load.
While experimenting with another patch, I noticed that when the
load type and the constant pool type don't match, then
SimplifyDemandedBits can't handle it. While we should probably
fix that, this was a simple way to fix the issue I saw.
llvm-svn: 366732
Ideally this needs to be a generic combine in DAGCombiner::visitEXTRACT_SUBVECTOR but there's some nasty regressions in aarch64 due to neon shuffles not handling bitcasts at all.....
llvm-svn: 364407
This shows up as a side issue to the main problem for the AVX target example from PR37428:
https://bugs.llvm.org/show_bug.cgi?id=37428 - https://godbolt.org/z/7tpRa3
But as we can see in the pile of existing test diffs, it's actually a widespread problem
that affects any AVX or later target. Apart from a couple of oddballs, I think these are
all improvements for the reasons stated in the code comment: we do not want to enable YMM
unnecessarily (avoid vzeroupper and frequency throttling) and some cores split 256-bit
stores anyway.
We could say that MergeConsecutiveStores() is going overboard on some of these examples,
but that won't solve the problem completely. But that is a reason I'm proposing this as
a lowering rather than a combine: we will infinite loop fighting the merge code if we try
this earlier.
Differential Revision: https://reviews.llvm.org/D62498
llvm-svn: 362524
This shows up as a side issue to the main problem for the AVX target example from PR37428:
https://bugs.llvm.org/show_bug.cgi?id=37428 - https://godbolt.org/z/7tpRa3
But as we can see in the pile of existing test diffs, it's actually a widespread problem
that affects any AVX or later target. Apart from a couple of oddballs, I think these are
all improvements for the reasons stated in the code comment: we do not want to enable YMM
unnecessarily (avoid vzeroupper and frequency throttling) and some cores split 256-bit
stores anyway.
We could say that MergeConsecutiveStores() is going overboard on some of these examples,
but that won't solve the problem completely. But that is the reason I'm proposing this as
a lowering rather than a combine: we will infinite loop fighting the merge code if we try
this earlier.
Differential Revision: https://reviews.llvm.org/D62498
llvm-svn: 361822
Basic "revectorization" combine, we can probably do more opcodes here but it can be a tricky cost-benefit depending on where the subvectors came from - but this case helps shuffle combining.
llvm-svn: 360134
SUBREG_TO_REG is supposed to be used to assert that we know the upper bits are
zero. But that isn't the case here. We've done no analysis of the inputs.
llvm-svn: 357673
Fixes PR41316 where the expanded PAVG intrinsic had had one of its ADDs turned into an OR due to its operands having no conflicting bits.
llvm-svn: 357351
Just enable this for AVX for now as SSE41 introduces extra register moves for the PMOVZX(PSHUFD(V)) -> UNPCKH(V,0) pattern (but otherwise helps reduce port5 usage on Intel targets).
Only AVX support is required for PR40685 as the issue is due to 8i8->8i32 zext shuffle leftovers.
llvm-svn: 356858
This isn't the final fix for our reduction/horizontal codegen, but it takes care
of a lot of the problems. After we narrow the shuffle, existing combines for
insert/extract and binops kick in, and we end up with cheaper 128-bit ops.
The avg and mul reduction tests show an existing shuffle lowering hole for
AVX2/AVX512. I think in its most minimal form this is:
https://bugs.llvm.org/show_bug.cgi?id=40434
...but we might need multiple fixes to get it right.
Differential Revision: https://reviews.llvm.org/D57156
llvm-svn: 352209
This code takes a truncate, fp_to_int, or int_to_fp with a legal result type and an input type that needs to be split and enlarges the elements in the result type before doing the split. Then inserts a follow up truncate or fp_round after concatenating the two halves back together.
But if the input type of the original op is being split on its way to ultimately being scalarized we're just going to end up building a vector from scalars and then truncating or rounding it in the vector register. Seems kind of silly to enlarge the result element type of the operation only to end up with scalar code and then building a vector with large elements only to make the elements smaller again in the vector register. Seems better to just try to get away producing smaller result types in the scalarized code.
The X86 test case that changes is a pretty contrived test case that exists because of a bug we used to have in our AVG matching code. I think the code is better now, but its not realistic anyway.
llvm-svn: 347482
SplitVecOp_TruncateHelper tries to introduce a multilevel truncate to avoid scalarization. But if splitting the result type would still be a legal type we don't need to do that.
The comment block at the top of the function implied that this was already implemented. I looked back through the history and it doesn't look to have ever been checked.
llvm-svn: 347479
Summary: This is an improvement over the two pshufbs and punpcklqdq we'd get otherwise.
Reviewers: RKSimon, spatel
Reviewed By: RKSimon
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D54671
llvm-svn: 347171
This is a long-awaited follow-up suggested in D33578. Since then, we've picked up even more
opportunities for vector narrowing from changes like D53784, so there are a lot of test diffs.
Apart from 2-3 strange cases, these are all wins.
I've structured this to be no-functional-change-intended for any target except for x86
because I couldn't tell if AArch64, ARM, and AMDGPU would improve or not. All of those
targets have existing regression tests (4, 4, 10 files respectively) that would be
affected. Also, Hexagon overrides the shouldReduceLoadWidth() hook, but doesn't show
any regression test diffs. The trade-off is deciding if an extra vector load is better
than a single wide load + extract_subvector.
For x86, this is almost always better (on paper at least) because we often can fold
loads into subsequent ops and not increase the official instruction count. There's also
some unknown -- but potentially large -- benefit from using narrower vector ops if wide
ops are implemented with multiple uops and/or frequency throttling is avoided.
Differential Revision: https://reviews.llvm.org/D54073
llvm-svn: 346595
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 brings back the MOV64r0 pseudo instruction for zeroing a 64-bit register. This replaces the SUBREG_TO_REG MOV32r0 sequence we use today. Post register allocation we will rewrite the MOV64r0 to a 32-bit xor with an implicit def of the 64-bit register similar to what we do for the various XMM/YMM/ZMM zeroing pseudos.
My main motivation is to enable the spill optimization in foldMemoryOperandImpl. As we were seeing some code that repeatedly did "xor eax, eax; store eax;" to spill several registers with a new xor for each store. With this optimization enabled we get a store of a 0 immediate instead of an xor. Though I admit the ideal solution would be one xor where there are multiple spills. I don't believe we have a test case that shows this optimization in here. I'll see if I can try to reduce one from the code were looking at.
There's definitely some other machine CSE(and maybe other passes) behavior changes exposed by this patch. So it seems like there might be some other deficiencies in SUBREG_TO_REG handling.
Differential Revision: https://reviews.llvm.org/D52757
llvm-svn: 345165
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
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
Enable enableMultipleCopyHints() on X86.
Original Patch by @jonpa:
While enabling the mischeduler for SystemZ, it was discovered that for some reason a test needed one extra seemingly needless COPY (test/CodeGen/SystemZ/call-03.ll). The handling for that is resulted in this patch, which improves the register coalescing by providing not just one copy hint, but a sorted list of copy hints. On SystemZ, this gives ~12500 less register moves on SPEC, as well as marginally less spilling.
Instead of improving just the SystemZ backend, the improvement has been implemented in common-code (calculateSpillWeightAndHint(). This gives a lot of test failures, but since this should be a general improvement I hope that the involved targets will help and review the test updates.
Differential Revision: https://reviews.llvm.org/D38128
llvm-svn: 342578
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
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
In DAGCombiner, we try to simplify this pattern:
([s|z]ext (load ...))
Conceptually, a new extload which is created while splitting the load
should have the same debug location as the load.
Making this change affects the IROrder of the new load, causing some
test case churn.
In practice, the new location is never different from the location of
the [s|z]ext, at least not during check-llvm or a stage2 build.
Part of: llvm.org/PR37262
Differential Revision: https://reviews.llvm.org/D46156
llvm-svn: 331301
Simon Pilgrim