The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/5EYc6r9nh - for intels `Block RThroughput: =6.0`; for ryzens, `Block RThroughput: <=3.0`
So pick cost of `6`.
For store we have:
https://godbolt.org/z/z61e5d6GE - for intels `Block RThroughput: =2.0`; for ryzens, `Block RThroughput: <=1.0`
So pick cost of `2`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D110536
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/q6GbK89br - for intels `Block RThroughput: =18.0`; for ryzens, `Block RThroughput: <=7.0`
So pick cost of `18`.
For store we have:
https://godbolt.org/z/Yzfoo5TnW - for intels `Block RThroughput: =8.0`; for ryzens, `Block RThroughput: <=4.0`
So pick cost of `8`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D110507
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/Y1E7qnjz8 - for intels `Block RThroughput: =9.0`; for ryzens, `Block RThroughput: <=3.5`
So pick cost of `9`.
For store we have:
https://godbolt.org/z/Y1E7qnjz8 - for intels `Block RThroughput: =4.0`; for ryzens, `Block RThroughput: <=2.0`
So pick cost of `4`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D110506
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/e5YE99a4P - for intels `Block RThroughput: =6.0`; for ryzens, `Block RThroughput: =2.0`
So pick cost of `6`.
For store we have:
https://godbolt.org/z/3vM4KsE1n - for intels `Block RThroughput: =3.0`; for ryzens, `Block RThroughput: <=2.0`
So pick cost of `3`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D110505
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/1j3nf3dro - for intels `Block RThroughput: =2.0`; for ryzens, `Block RThroughput: <=1.0`
So pick cost of `2`.
For store we have:
https://godbolt.org/z/4n1zvP37j - for intels `Block RThroughput: =1.0`; for ryzens, `Block RThroughput: <=0.5`
So pick cost of `1`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D110504
The only sched models that for cpu's that support avx2
but not avx512 are: haswell, broadwell, skylake, zen1-3
For load we have:
https://godbolt.org/z/M8vEKs5jY - for intels `Block RThroughput: =2.0`;
for ryzens, `Block RThroughput: <=1.0`
So pick cost of `2`.
For store we have:
https://godbolt.org/z/Kx1nKz7je - for intels `Block RThroughput: =1.0`;
for ryzens, `Block RThroughput: <=0.5`
So pick cost of `1`.
I'm directly using the shuffling asm the llc produced,
without any manual fixups that may be needed
to ensure sequential execution.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D103144
Update the costs to match the codegen from combineMulToPMADDWD - not only can we use PMADDWD is its zero-extended, but also if its a constant or sign-extended from a vXi16 (which can be replaced with a zero-extension).
Only the most recent cpus support really 1cy 64-bit multiplies, and the X64 cost table represents a realistic worst case. The 1cy value was also discouraging vectorization when most vXi64 PMULDQ expansions aren't actually slower than scalarization.
Noticed while investigating PR51436.
Based off the worse case numbers generated by D103695, the AVX2/512 bit reversing/counting costs were higher than necessary (based off instruction counts instead of actual throughput).
This renames the primary methods for creating a zero value to `getZero`
instead of `getNullValue` and renames predicates like `isAllOnesValue`
to simply `isAllOnes`. This achieves two things:
1) This starts standardizing predicates across the LLVM codebase,
following (in this case) ConstantInt. The word "Value" doesn't
convey anything of merit, and is missing in some of the other things.
2) Calling an integer "null" doesn't make any sense. The original sin
here is mine and I've regretted it for years. This moves us to calling
it "zero" instead, which is correct!
APInt is widely used and I don't think anyone is keen to take massive source
breakage on anything so core, at least not all in one go. As such, this
doesn't actually delete any entrypoints, it "soft deprecates" them with a
comment.
Included in this patch are changes to a bunch of the codebase, but there are
more. We should normalize SelectionDAG and other APIs as well, which would
make the API change more mechanical.
Differential Revision: https://reviews.llvm.org/D109483
Please refer to
https://lists.llvm.org/pipermail/llvm-dev/2021-September/152440.html
(and that whole thread.)
TLDR: the original patch had no prior RFC, yet it had some changes that
really need a proper RFC discussion. It won't be productive to discuss
such an RFC, once it's actually posted, while said patch is already
committed, because that introduces bias towards already-committed stuff,
and the tree is potentially in broken state meanwhile.
While the end result of discussion may lead back to the current design,
it may also not lead to the current design.
Therefore i take it upon myself
to revert the tree back to last known good state.
This reverts commit 4c4093e6e3.
This reverts commit 0a2b1ba33a.
This reverts commit d9873711cb.
This reverts commit 791006fb8c.
This reverts commit c22b64ef66.
This reverts commit 72ebcd3198.
This reverts commit 5fa6039a5f.
This reverts commit 9efda541bf.
This reverts commit 94d3ff09cf.
The GLM/SLM special cases still get tested first but after the the MUL/DIV/REM pattern detection - this will be necessary for when we make the SLM vXi32 MUL canonicalization generic to improve PMULLW/PMULHW/PMADDDW cost support etc.
This patch adds an initial ShuffleVectorInst::isInsertSubvectorMask helper to recognize 2-op shuffles where the lowest elements of one of the sources are being inserted into the "in-place" other operand, this includes "concat_vectors" patterns as can be seen in the Arm shuffle cost changes. This also helped fix a x86 issue with irregular/length-changing SK_InsertSubvector costs - I'm hoping this will help with D107188
This doesn't currently attempt to work with 1-op shuffles that could either be a "widening" shuffle or a self-insertion.
The self-insertion case is tricky, but we currently always match this with the existing SK_PermuteSingleSrc logic.
The widening case will be addressed in a follow up patch that treats the cost as 0.
Masks with a high number of undef elts will still struggle to match optimal subvector widths - its currently bounded by minimum-width possible insertion, whilst some cases would benefit from wider (pow2?) subvectors.
Differential Revision: https://reviews.llvm.org/D107228
I have added a new FastMathFlags parameter to getArithmeticReductionCost
to indicate what type of reduction we are performing:
1. Tree-wise. This is the typical fast-math reduction that involves
continually splitting a vector up into halves and adding each
half together until we get a scalar result. This is the default
behaviour for integers, whereas for floating point we only do this
if reassociation is allowed.
2. Ordered. This now allows us to estimate the cost of performing
a strict vector reduction by treating it as a series of scalar
operations in lane order. This is the case when FP reassociation
is not permitted. For scalable vectors this is more difficult
because at compile time we do not know how many lanes there are,
and so we use the worst case maximum vscale value.
I have also fixed getTypeBasedIntrinsicInstrCost to pass in the
FastMathFlags, which meant fixing up some X86 tests where we always
assumed the vector.reduce.fadd/mul intrinsics were 'fast'.
New tests have been added here:
Analysis/CostModel/AArch64/reduce-fadd.ll
Analysis/CostModel/AArch64/sve-intrinsics.ll
Transforms/LoopVectorize/AArch64/strict-fadd-cost.ll
Transforms/LoopVectorize/AArch64/sve-strict-fadd-cost.ll
Differential Revision: https://reviews.llvm.org/D105432
Update shl/lshr/ashr costs based on the worst case costs from the script in D103695 - many of the 128-bit shifts (usually where integer multiplies aren't used) have similar behaviour to AVX1 so we can merge them.
We know that "CVTTPS2SI" returns 0x80000000 for out of range inputs (and for FP_TO_UINT, negative float values are undefined). We can use this to make unsigned conversions from vXf32 to vXi32 more efficient, particularly on targets without blend using the following logic:
small := CVTTPS2SI(x);
fp_to_ui(x) := small | (CVTTPS2SI(x - 2^31) & ARITHMETIC_RIGHT_SHIFT(small, 31))
Even on targets where "PBLENDVPS"/"PBLENDVB" exists, it is often a latency 2, low throughput instruction so this logic is applied there too (in particular for AVX2 also). It furthermore gets rid of one high latency floating point comparison in the previous lowering.
@TomHender checked the correctness of this for all possible floats between -1 and 2^32 (both ends excluded).
Original Patch by @TomHender (Tom Hender)
Differential Revision: https://reviews.llvm.org/D89697
Update (mainly) vXf32/vXf64 -> vXi8/vXi16 fptosi/fptoui costs based on the worst case costs from the script in D103695.
Move to using legalized types wherever possible, which allows us to prune the cost tables.
Update truncation costs based on the worst case costs from the script in D103695.
Move to using legalized types wherever possible, which allows us to prune the cost tables.
This patch removes the IsPairwiseForm flag from the Reduction Cost TTI
hooks, along with some accompanying code for pattern matching reductions
from trees starting at extract elements. IsPairWise is now assumed to be
false, which was the predominant way that the value was used from both
the Loop and SLP vectorizers. Since the adjustments such as D93860, the
SLP vectorizer has not relied upon this distinction between paiwise and
non-pairwise reductions.
This also removes some code that was detecting reductions trees starting
from extract elements inside the costmodel. This case was
double-counting costs though, adding the individual costs on the
individual instruction _and_ the total cost of the reduction. Removing
it changes the costs in llvm/test/Analysis/CostModel/X86/reduction.ll to
not double count. The cost of reduction intrinsics is still tested
through the various tests in
llvm/test/Analysis/CostModel/X86/reduce-xyz.ll.
Differential Revision: https://reviews.llvm.org/D105484
Revived D101297 in its original form + added some changes in X86
legalization cehcking for masked gathers.
This solution is the most stable and the most correct one. We have to
check the legality before trying to build the masked gather in SLP.
Without this check we have incorrect cost (for SLP) in case if the masked gather
is not legal/slower than the gather. And we're missing some
vectorization opportunities.
This can be fixed in the cost model, but in this case we need to add
special checks for the cost of GEPs for ScatterVectorize node, add
special check for small trees, etc., i.e. there are a lot of corner
cases here and there, which insrease code base and make it harder to
maintain the code.
> Can't we rely on cost model to deal with this? This can be profitable for futher vectorization, when we can start from such gather loads as seed.
The question from D101297. Actually, no, it can't. Actually, simple
gather may give us better result, especially after we started
vectorization of insertelements. Plus, like I said before, the cost for
non-legal masked gathers leads to missed vectorization opportunities.
Differential Revision: https://reviews.llvm.org/D105042
Update costs based on the worst case costs from the script in D103695.
Move to using legalized types wherever possible, which allows us to prune the cost tables.
Update (mainly) vXi8/vXi16 -> vXf32/vXf64 sitofp/uitofp costs based on the worst case costs from the script in D103695.
Move to using legalized types wherever possible, which allows us to prune the cost tables.
Provide a generic fallback that performs the fptosi to i32 types, then truncates to sub-i32 scalars.
These numbers can be tweaked for specific sse levels, but we should get the default handling in place first.
Provide a generic fallback that extends sub-i32 scalars before using the existing sitofp instructions.
These numbers can be tweaked for specific sse levels, but we should get the default handling in place first.
We get the extension for free for non-vector loads.
Update v4i64 -> v4f32/v4f64 uitofp costs based on the worst case costs from the script in D103695.
Fixes a few regressions before we start adding AVX costs for legalized types.
Building on rG2a1ef8784ad9a, adjust the SSE cost tables to use the legalized types based on the worst case costs from the script in D103695.
To account for different numbers of src/dst legalized type registers we must scale the cost by maximum of the src/dst, not just use src
Move the (SSE-only) generic, legalized type conversion matching after the specific,custom conversion cases, allowing us to properly provide cost overrides.
The next step will be to clean up some of the weird existing costs and then to enable AVX+ legalized costs, which will let us strip out a lot of the cost tables entries.
Based off the worse case numbers generated by D103695, the AVX1/2/512 sitofp/uitofp/fptosi/fptoui costs were higher than necessary (based off instruction counts instead of actual throughput).
The SSE costs still need further fixes, but I hit an issue with the order in which SSE costs are checked - we need to check CUSTOM costs (with non-legal types) first, and then fallback to LEGALIZED types. I'm looking at this now, and this should let us start thinning out a lot of the duplicates in the costs tables.
Then we can finally start work on vXi64 / vXi16 / vXi8 / vXi1 integers, which should let us look at sub-128-bit vectorization (D103925).
Based off the worse case numbers generated by D103695, we were overestimating the cost of a number of vector truncations:
AVX2: v2i32->v2i8, v2i64->v2i16 + v4i64->v4i32
AVX1: v2i32->v2i8, v4i64->v4i16 + v16i16->v16i8
Once we have a working set of conversion costs, the intention is to cleanup the tables and use legalized types a lot more to reduce the number of entries we currently have.
Determined from llvm-mca analysis (btver2 vs bdver2 vs sandybridge), the split+extends+concat sequence on AVX1 capable targets are cheaper than the #ops that the cost was previously based on.
The SkylakeServer model (and later IceLake/TigerLake targets according to Agner) have the PMOV truncations as uops=2, rthroughput=2 instructions.
Noticed while trying to reduce the diffs between cost tables and llvm-mca analysis.
Determined from llvm-mca analysis, AVX1 capable targets have a higher throughput for VPBLENDVB and shuffle ops, making it cheaper to perform shift+shuffle/select shift patterns.
rG1ad4f887bd7692a9e63fb42586f0ece366f2fe01 incorrectly assumed that vXi64 non-uniform shifts were slow like vXi32 were - but llvm-mca (+Agner) both confirm that Haswell/Broadwell are full rate.
Determined from llvm-mca analysis, AVX2+ capable targets have a higher throughput for VPBLENDVB and VPMOVZX ops, making it cheaper to perform shift+select patterns for vXi8 shifts or extend/shift/truncate for vXi16 shifts. Similarly AVX512BW can perform vXi8 as extend/shift/truncate patterns.
This follows in steps of similar `getMemoryOpCost()` changes, D100099/D100684.
Intel SDM, `VPMASKMOV — Conditional SIMD Integer Packed Loads and Stores`:
```
Faults occur only due to mask-bit required memory accesses that caused the faults. Faults will not occur due to
referencing any memory location if the corresponding mask bit for that memory location is 0. For example, no
faults will be detected if the mask bits are all zero.
```
I.e., if mask is all-zeros, any address is fine.
Masked load/store's prime use-case is e.g. tail masking the loop remainder,
where for the last iteration, only first some few elements of a vector exist.
So much similarly, i don't see why must we scalarize non-power-of-two vectors,
iff the element type is something we can masked- store/load.
We simply need to legalize it, widen the mask, and be done with it.
And we even already count the cost of widening the mask.
Reviewed By: ABataev
Differential Revision: https://reviews.llvm.org/D102990
By llvm-mca analysis, Haswell/Broadwell has a non-uniform vector shift recip-throughput cost of the AVX2 targets at 2 for both 128 and 256-bit vectors - XOP capable targets have better 128-bit vector shifts so improve the fallback in those cases.
By llvm-mca analysis, Haswell/Broadwell has the worst v4i64 recip-throughput cost of the AVX2 targets at 6 (vs the currently used cost of 8). Similarly SkylakeServer (our only AVX512 target model) implements PMULLQ with an average cost of 1.5 (rounded up to 2.0), and the PMULUDQ-sequence (without AVX512DQ) as a cost of 6.
Based on worst case of sandybridge (vs btver2 + bdver2) llvm-mca analysis - which is a lot less than what we were predicting (I think based off total uop count).
BTVER2 has a 2 cycle throughput for v4i32 multiplies (same as SSE41 targets), which is only partially hidden by the subvector extracts/insert when splitting v8i32.
Now that getMemoryOpCost() correctly handles all the vector variants,
we should no longer hand-roll our own version of it, but use it directly.
The AVX512 variant probably needs a similar change,
but there it is less obvious.
This was initially landed in 69ed93a435,
but was reverted in 6b95fd199d
because the patch it depends on was reverted.
Instead of handling power-of-two sized vector chunks,
try handling the large vector in a stream mode,
decreasing the operational vector size
once it no longer works for the elements left to process.
Notably, this improves costs for overaligned loads - loading padding is fine.
This more directly tracks when we need to insert/extract the YMM/XMM subvector,
some costs fluctuate because of that.
This was initially landed in c02476f315,
but reverted in 5fddc3312b,
because the code made some very optimistic assumptions about invariants
that didn't hold in practice.
Reviewed By: RKSimon, ABataev
Differential Revision: https://reviews.llvm.org/D100684
BTVER2 has a weaker f64 multiplier that other AVX1-era targets, so we need to bump the worst case cost slightly - llvm-mca reports the new vectorization in simplebb is beneficial on btver2, bdver2 and sandybridge AVX1 targets
Haswell, Excavator and early Ryzen all have slower 256-bit non-uniform vector shifts (confirmed on AMDSoG/Agner/instlatx64 and llvm models) - so bump the worst case costs accordingly.
Noticed while investigating PR50364
Noticed while investigating PR50364, the truncation costs for v4i64->v4i16/v4i8 and v8i32->v8i8 were way too optimistic for a shuffle sequence that usually matches the AVX1 codegen (they matched AVX512 numbers which have actual truncation instructions!).
Now that getMemoryOpCost() correctly handles all the vector variants,
we should no longer hand-roll our own version of it, but use it directly.
The AVX512 variant probably needs a similar change,
but there it is less obvious.
Instead of handling power-of-two sized vector chunks,
try handling the large vector in a stream mode,
decreasing the operational vector size
once it no longer works for the elements left to process.
Notably, this improves costs for overaligned loads - loading padding is fine.
This more directly tracks when we need to insert/extract the YMM/XMM subvector,
some costs fluctuate because of that.
Reviewed By: RKSimon, ABataev
Differential Revision: https://reviews.llvm.org/D100684
Currently we model i16 bswap as very high cost (`10`),
which doesn't seem right, with all other being at `1`.
Regardless of `MOVBE`, i16 reg-reg bswap is lowered into
(an extending move plus) rot-by-8:
https://godbolt.org/z/8jrq7fMTj
I think it should at worst have throughput of `1`:
Since i32/i64 already have cost of `1`,
`MOVBE` doesn't improve their costs any further.
BUT, `MOVBE` must have at least a single memory operand,
with other being a register. Which means, if we have
a bswap of load, iff load has a single use,
we'll fold bswap into load.
Likewise, if we have store of a bswap, iff bswap
has a single use, we'll fold bswap into store.
So i think we should treat such a bswap as free,
unless of course we know that for the particular CPU
they are performing badly.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D101924
Added an extra analysis for better choosing of shuffle kind in
getShuffleCost functions for better cost estimation if mask was
provided.
Differential Revision: https://reviews.llvm.org/D100865
Added an extra analysis for better choosing of shuffle kind in
getShuffleCost functions for better cost estimation if mask was
provided.
Differential Revision: https://reviews.llvm.org/D100865
This is similar to the subvector extractions,
except that the 0'th subvector isn't free to insert,
because we generally don't know whether or not
the upper elements need to be preserved:
https://godbolt.org/z/rsxP5W4sW
This is needed to avoid regressions in D100684
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D100698
Sometimes LV has to produce really wide vectors,
and sometimes they end up being not powers of two.
As it can be seen from the diff, the cost computation
is currently completely non-sensical in those cases.
Instead of just scalarizing everything, split/factorize the wide vector
into a number of subvectors, each one having a power-of-two elements,
recurse to get the cost of op on this subvector. Also, check how we'd
legalize this subvector, and if the legalized type is scalar,
also account for the scalarization cost.
Note that for sub-vector loads, we might be able to do better,
when the vectors are properly aligned.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D100099
Added cost estimation for switch instruction, updated costs of branches, fixed
phi cost.
Had to increase `-amdgpu-unroll-threshold-if` default value since conditional
branch cost (size) was corrected to higher value.
Test renamed to "control-flow.ll".
Removed redundant code in `X86TTIImpl::getCFInstrCost()` and
`PPCTTIImpl::getCFInstrCost()`.
Reviewed By: rampitec
Differential Revision: https://reviews.llvm.org/D96805
After rG47321c311bdbe0145b9bf45d822185c37b19fa50 we promote vXi8 reductions to vXi16 to create a much faster PMULLW mul reduction, followed by a (free) truncation. This avoids the high cost of repeated vXi8 multiplications (which extend+multiply+truncate to/from vXi16 types....).
Fixes the missing vXi8 mul reduction vectorization in PR42674 (Comment #20) 'mul16' test case.
This patch changes the interface to take a RegisterKind, to indicate
whether the register bitwidth of a scalar register, fixed-width vector
register, or scalable vector register must be returned.
Reviewed By: paulwalker-arm
Differential Revision: https://reviews.llvm.org/D98874
This adds an Mask ArrayRef to getShuffleCost, so that if an exact mask
can be provided a more accurate cost can be provided by the backend.
For example VREV costs could be returned by the ARM backend. This should
be an NFC until then, laying the groundwork for that to be added.
Differential Revision: https://reviews.llvm.org/D98206
Use instead of the isa_and_nonnull<StoreInst> and use the StoreInst::getPointerOperand wrapper instead of a hardcoded Instruction::getOperand.
Looks cleaner and avoids a spurious clang static analyzer null dereference warning.
Noticed while looking at D92701 - we only really handle TCK_RecipThroughput gather/scatter costs - for now drop back to the default implementation for non-legal gathers/scatters.
Without FMF, we lower these intrinsics into something like this:
vmaxsd %xmm0, %xmm1, %xmm2
vcmpunordsd %xmm0, %xmm0, %xmm0
vblendvpd %xmm0, %xmm1, %xmm2, %xmm0
But if we can ignore NANs, the single min/max instruction is enough
because there is no need to fix up the x86 logic that corresponds to
X > Y ? X : Y.
We probably want to make other adjustments for FP intrinsics with FMF
to account for specialized codegen (for example, FSQRT).
Differential Revision: https://reviews.llvm.org/D92337
Update costs now that D92095 and D92102 have tweaked the SSE2 implementation
The SSE42 BLENDVPD cost can actually be used on SSE41 as we don't attempt to generate PCMPGT anymore
Add scalar i16/i32/i64 costs as we can do this cheaply with CMOV
We can use GF2P8AFFINEQB to reverse bits in a byte. Shuffles are needed to reverse the bytes in elements larger than i8. LegalizeVectorOps takes care of inserting the shuffle for the larger element size.
We already have Custom lowering for v16i8 with SSSE3, v32i8 with AVX, and v64i8 with AVX512BW.
I think we might be able to use this for scalars too by moving into a vector and back. But I'll save that for a follow up as its a little more involved.
Reviewed By: RKSimon, pengfei
Differential Revision: https://reviews.llvm.org/D91515
As noticed in D90554 ,
the AVX2 costs for 256-bit vectors did not include FMAXNUM entries,
so we fell back to AVX1 which assumes those ops will be split into
128-bit halves or something close to that.
Differential Revision: https://reviews.llvm.org/D90613
This reverts the revert commit 408c4408fa.
This version of the patch includes a fix for a crash caused by
treating ICmp/FCmp constant expressions as instructions.
Original message:
On some targets, like AArch64, vector selects can be efficiently lowered
if the vector condition is a compare with a supported predicate.
This patch adds a new argument to getCmpSelInstrCost, to indicate the
predicate of the feeding select condition. Note that it is not
sufficient to use the context instruction when querying the cost of a
vector select starting from a scalar one, because the condition of the
vector select could be composed of compares with different predicates.
This change greatly improves modeling the costs of certain
compare/select patterns on AArch64.
I am also planning on putting up patches to make use of the new argument in
SLPVectorizer & LV.
I'm assuming the standard size integer instructions for this end up as something like:
mulq %rsi
seto %al
And the 'mul' generally has reciprocal throughput of 1 on typical implementations
(higher latency, but that's not handled here).
The default costs may end up much higher than that, and that's what we see in the test diffs.
Vector types are left as a 'TODO'.
Differential Revision: https://reviews.llvm.org/D90431
On some targets, like AArch64, vector selects can be efficiently lowered
if the vector condition is a compare with a supported predicate.
This patch adds a new argument to getCmpSelInstrCost, to indicate the
predicate of the feeding select condition. Note that it is not
sufficient to use the context instruction when querying the cost of a
vector select starting from a scalar one, because the condition of the
vector select could be composed of compares with different predicates.
This change greatly improves modeling the costs of certain
compare/select patterns on AArch64.
I am also planning on putting up patches to make use of the new argument in
SLPVectorizer & LV.
Reviewed By: dmgreen, RKSimon
Differential Revision: https://reviews.llvm.org/D90070
In each 128-lane, if there is at least one index is demanded and not all
indices are demanded and this 128-lane is not the first 128-lane of the
legalized-vector, then this 128-lane needs a extracti128;
If in each 128-lane, there is at least one index is demanded, this 128-lane
needs a inserti128.
The following cases will help you build a better understanding:
Assume we insert several elements into a v8i32 vector in avx2,
Case#1: inserting into 1th index needs vpinsrd + inserti128
Case#2: inserting into 5th index needs extracti128 + vpinsrd +
inserti128
Case#3: inserting into 4,5,6,7 index needs 4*vpinsrd + inserti128.
Reviewed By: pengfei, RKSimon
Differential Revision: https://reviews.llvm.org/D89767
This is my first LLVM patch, so please tell me if there are any process issues.
The main observation for this patch is that we can lower UMIN/UMAX with v8i16 by using unsigned saturated subtractions in a clever way. Previously this operation was lowered by turning the signbit of both inputs and the output which turns the unsigned minimum/maximum into a signed one.
We could use this trick in reverse for lowering SMIN/SMAX with v16i8 instead. In terms of latency/throughput this is the needs one large move instruction. It's just that the sign bit turning has an increased chance of being optimized further. This is particularly apparent in the "reduce" test cases. However due to the slight regression in the single use case, this patch no longer proposes this.
Unfortunately this argument also applies in reverse to the new lowering of UMIN/UMAX with v8i16 which regresses the "horizontal-reduce-umax", "horizontal-reduce-umin", "vector-reduce-umin" and "vector-reduce-umax" test cases a bit with this patch. Maybe some extra casework would be possible to avoid this. However independent of that I believe that the benefits in the common case of just 1 to 3 chained min/max instructions outweighs the downsides in that specific case.
Patch By: @TomHender (Tom Hender) ActuallyaDeviloper
Differential Revision: https://reviews.llvm.org/D87236
Changes TTI function getIntImmCostInst to take an additional Instruction parameter,
which enables us to be able to check it is part of a min(max())/max(min()) pattern that will match SSAT.
We can then mark the constant used as free to prevent it being hoisted so SSAT can still be generated.
Required minor changes in some non-ARM backends to allow for the optional parameter to be included.
Differential Revision: https://reviews.llvm.org/D87457
Other types can be handled in future patches but their uniform / non-uniform costs are more similar and don't appear to cause many vectorization issues.
Currently, getCastInstrCost has limited information about the cast it's
rating, often just the opcode and types. Sometimes there is a context
instruction as well, but it isn't trustworthy: for instance, when the
vectorizer is rating a plan, it calls getCastInstrCost with the old
instructions when, in fact, it's trying to evaluate the cost of the
instruction post-vectorization. Thus, the current system can get the
cost of certain casts incorrect as the correct cost can vary greatly
based on the context in which it's used.
For example, if the vectorizer queries getCastInstrCost to evaluate the
cost of a sext(load) with tail predication enabled, getCastInstrCost
will think it's free most of the time, but it's not always free. On ARM
MVE, a VLD2 group cannot be extended like a normal VLDR can. Similar
situations can come up with how masked loads can be extended when being
split.
To fix that, this path adds a new parameter to getCastInstrCost to give
it a hint about the context of the cast. It adds a CastContextHint enum
which contains the type of the load/store being created by the
vectorizer - one for each of the types it can produce.
Original patch by Pierre van Houtryve
Differential Revision: https://reviews.llvm.org/D79162
These cost methods don't make much sense in X86Subtarget. Make
them methods in X86's TTI and move the feature checks from the
X86Subtarget constructor into these methods.
Reviewed By: RKSimon
Differential Revision: https://reviews.llvm.org/D84594
The main interface has been migrated to Align already but a few backends where broadening the type from Align to MaybeAlign.
This patch makes sure all implementations conform to the public API.
Differential Revision: https://reviews.llvm.org/D82465
Summary:
Get back `const` partially lost in one of recent changes.
Additionally specify explicit qualifiers in few places.
Reviewers: samparker
Reviewed By: samparker
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D82383
Have BasicTTI call the base implementation so that both agree on the
default behaviour, which the default being a cost of '1'. This has
required an X86 specific implementation as it seems to be very
reliant on those instructions being free. Changes are also made to
AMDGPU so that their implementations distinguish between cost kinds,
so that the unrolling isn't affected. PowerPC also has its own
implementation to prevent changes to the reg-usage vectorizer test.
The cost model test changes now reflect that ret instructions are not
generally free.
Differential Revision: https://reviews.llvm.org/D79164
Add the remaining arithmetic opcodes into the generic implementation
of getUserCost and then call this from getInstructionThroughput. Most
of the backends have been modified to return the base implementation
for cost kinds other RecipThroughput. The outlier here is AMDGPU
which already uses getArithmeticInstrCost for all the cost kinds.
This change means that most of the opcodes can be removed from that
backends implementation of getUserCost.
Differential Revision: https://reviews.llvm.org/D80992
Add cases for icmp, fcmp and select into the switch statement of the
generic getUserCost implementation with getInstructionThroughput then
calling into it. The BasicTTI and backend implementations have be set
to return a default value (1) when a cost other than throughput is
being queried.
Differential Revision: https://reviews.llvm.org/D80550
Use getMemoryOpCost from the generic implementation of getUserCost
and have getInstructionThroughput return the result of that for loads
and stores.
This also means that the X86 implementation of getUserCost can be
removed with the functionality folded into its getMemoryOpCost.
Differential Revision: https://reviews.llvm.org/D80984
Add the remaining cast instruction opcodes to the base implementation
of getUserCost and directly return the result. This allows
getInstructionThroughput to return getUserCost for the casts. This
has required changes to PPC and SystemZ because they implement
getUserCost and/or getCastInstrCost with adjustments for vector
operations. Adjusts have also been made in the remaining backends
that implement the method so that they still produce a cost of zero
or one for cost kinds other than throughput.
Differential Revision: https://reviews.llvm.org/D79848
Recommitting most of the remaining changes from
259eb619ff, but excluding the call to
getUserCost from getInstructionThroughput. Though there's still no
test changes, I doubt that this is an NFC...
With the two getIntrinsicInstrCosts folded into one, now fold in the
scalar/code-size orientated getIntrinsicCost. The remaining scalar
intrinsics were memcpy, cttz and ctlz which now have special handling
in the BasicTTI implementation.
This had required a change in the AMDGPU backend for fabs as it
should always be 'free'. I've also changed the X86 backend to return
the BaseT implementation when the CostKind isn't RecipThroughput.
Differential Revision: https://reviews.llvm.org/D80012
With the two getIntrinsicInstrCosts folded into one, now fold in the
scalar/code-size orientated getIntrinsicCost. This involved sinking
cost of the TTIImpl into the base implementation, as it performs no
target checks. The opcodes remaining were memcpy, cttz and ctlz which
now have special handling in the BasicTTI implementation.
getInstructionThroughput can now directly return the result of
getUserCost.
This had required a change in the AMDGPU backend for fabs and its
always 'free'. I've also changed the X86 backend to return '1' for
any intrinsic when the CostKind isn't RecipThroughput.
Though this intended to be a non-functional change, there are many
paths being combined here so I would be very surprised if this didn't
have an effect.
Differential Revision: https://reviews.llvm.org/D80012
Combine the two API calls into one by introducing a structure to hold
the relevant data. This has the added benefit of moving the boiler
plate code for arguments and flags, into the constructors. This is
intended to be a non-functional change, but the complicated web of
logic involved here makes it very hard to guarantee.
Differential Revision: https://reviews.llvm.org/D79941
getScalarizationOverhead is only ever called with vectors (and we already had a load of cast<VectorType> calls immediately inside the functions).
Followup to D78357
Reviewed By: @samparker
Differential Revision: https://reviews.llvm.org/D79341
Make the kind of cost explicit throughout the cost model which,
apart from making the cost clear, will allow the generic parts to
calculate better costs. It will also allow some backends to
approximate and correlate the different costs if they wish. Another
benefit is that it will also help simplify the cost model around
immediate and intrinsic costs, where we currently have multiple APIs.
RFC thread:
http://lists.llvm.org/pipermail/llvm-dev/2020-April/141263.html
Differential Revision: https://reviews.llvm.org/D79002
Also fix some cost tables for vXi1 types to match the costs entries for the types they will be promoted to.
Differential Revision: https://reviews.llvm.org/D79045
vpermw is 2 uops. vpermt2b/vpermt2w are two shuffle uops and a port 015 uop. Weirdly vpermb is a single uop.
This patch bumps the cost to 2 for these operations. Maybe should go to 3 for the vpermt2*, but I've started conservative.
I've also removed a few entries that were now the same as earlier subtargets or that I didn't think we really did. Like I don't think we extend v32i8 to v32i16, shuffle, and then truncate.
Differential Revision: https://reviews.llvm.org/D79148
We generate much better code these days than we used to. And we use the same sequence for AVX1 and AVX2 for these
For v4i64->v4i32 we generate:
vextractf128 xmm1, ymm0, 1
vshufps xmm0, xmm0, xmm1, 136 # xmm0 = xmm0[0,2],xmm1[0,2]
And for v8i64->v8i32 we generate:
vperm2f128 ymm2, ymm0, ymm1, 49 # ymm2 = ymm0[2,3],ymm1[2,3]
vinsertf128 ymm0, ymm0, xmm1, 1
vshufps ymm0, ymm0, ymm2, 136 # ymm0 = ymm0[0,2],ymm2[0,2],ymm0[4,6],ymm2[4,6]
Differential Revision: https://reviews.llvm.org/D79109
The improvements to the x86 vector insert/extract element costs in D74976 resulted in the estimated costs for vector initialization and scalarization increasing higher than should be expected. This is particularly noticeable on pre-SSE4 targets where the available of legal INSERT_VECTOR_ELT ops is more limited.
This patch does 2 things:
1 - it implements X86TTIImpl::getScalarizationOverhead to more accurately represent the typical costs of a ISD::BUILD_VECTOR pattern.
2 - it adds a DemandedElts mask to getScalarizationOverhead to permit the SLP's BoUpSLP::getGatherCost to be rewritten to use it directly instead of accumulating raw vector insertion costs.
This fixes PR45418 where a v4i8 (zext'd to v4i32) was no longer vectorizing.
A future patch should extend X86TTIImpl::getScalarizationOverhead to tweak the EXTRACT_VECTOR_ELT scalarization costs as well.
Reviewed By: @craig.topper
Differential Revision: https://reviews.llvm.org/D78216
I've modified isTruncateFree to get an accurate cost for types that need to be split. I'm planning to look into fixing it for all vectors, but need more cost cleanups first.
Differential Revision: https://reviews.llvm.org/D78973
There are several different types of cost that TTI tries to provide
explicit information for: throughput, latency, code size along with
a vague 'intersection of code-size cost and execution cost'.
The vectorizer is a keen user of RecipThroughput and there's at least
'getInstructionThroughput' and 'getArithmeticInstrCost' designed to
help with this cost. The latency cost has a single use and a single
implementation. The intersection cost appears to cover most of the
rest of the API.
getUserCost is explicitly called from within TTI when the user has
been explicit in wanting the code size (also only one use) as well
as a few passes which are concerned with a mixture of size and/or
a relative cost. In many cases these costs are closely related, such
as when multiple instructions are required, but one evident diverging
cost in this function is for div/rem.
This patch adds an argument so that the cost required is explicit,
so that we can make the important distinction when necessary.
Differential Revision: https://reviews.llvm.org/D78635
All avx512 truncate instructions except vXi64->vXi32 are 2 uops
on port 5. So raise their costs to 2. Except when we have an
earlier faster sequence like pshufb for 128 bit input vectors.
Add a lower cost of 3 v16i16->v16i8 with avx512f where we can
extend to v16i32 then truncate. And a cost of 2 for avx512bw with
and without avx512vl. There we can use vpmovwb with either a ymm
or zmm input. Both of these beat masking, splitting, and using
packuswb which is our avx/avx2 codegen.
We're currently getting this from the default implementation. But
I don't like how the cost model came to this answer and I might
be making some changes there.
If one of caller/callee has disabled ZMM registers due to
prefer-vector-width=256, we were previously
disabling argument promotion as the ABI might be incompatible since
one side will split 512-bit vectors in this case.
But if we can see that the types are all scalar this shouldn't be
a problem.
This patch assumes that pointer element type reflects the type that
the argument will be promoted to.
Differential Revision: https://reviews.llvm.org/D78770
The API for shuffles and reductions uses generic Type parameters,
instead of VectorType, and so assertions and casts are used a lot.
This patch makes those types explicit, which means that the clients
can't be lazy, but results in less ambiguity, and that can only be a
good thing.
Bugzilla: https://bugs.llvm.org/show_bug.cgi?id=45562
Differential Revision: https://reviews.llvm.org/D78357
Summary:
Remove usages of asserting vector getters in Type in preparation for the
VectorType refactor. The existence of these functions complicates the
refactor while adding little value.
Reviewers: craig.topper, sdesmalen, efriedma, RKSimon
Reviewed By: efriedma
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D77264
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
If we're inserting into v2i8/v4i8/v8i8/v2i16/v4i16 style sub-128bit vectors ensure we don't use the SK_PermuteTwoSrc cost of the legalized value type - this is a followup to rG12c629ec6c59 which added equivalent sub-128bit shuffle costs
This is similar to what I recently did for getArithmeticReductionCost.
I'm trying to account for the narrowing from 512->256->128 as we go.
I've also added a new helper method getMinMaxCost that tries to
handle the cases where we have native min/max instructions and
fall back to cmp+select when we don't.
Differential Revision: https://reviews.llvm.org/D76634
v2i8/v4i8/v8i8 + v2i16/v4i16 all show up in vectorizer code and by just using the legalized types (v16i8/v8i16) we're highly exaggerating the actual cost of the shuffle.
Summary:
This allows doing `memcmp(p, q, 7)` with 2 loads instead of a call to
memcmp.
This fixes part of PR45147.
Reviewers: spatel
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D76133
This patch attempts to more accurately model the reduction of
power of 2 vectors of types we natively support. This takes into
account the narrowing of vectors that occur as we go from 512
bits to 256 bits, to 128 bits. It also takes into account the use
of wider elements in the shuffles for the first 2 steps of a
reduction from 128 bits. And uses a v8i16 shift for the final step
of vXi8 reduction.
The default implementation uses the legalized type for the arithmetic
for all levels. And uses the single source permute cost of the
legalized type for all levels. This penalizes things like
lack of v16i8 pshufb on pre-sse3 targets and the splitting and
joining that needs to be done for integer types on AVX1. We never
need v16i8 shuffle for a reduction and we only need split AVX1 ops
when type the type wide and needs to be split. I think we're still
over costing splits and joins for AVX1, but we're closer now.
I've also removed all pairwise special casing because I don't
think we ever want to generate that on X86. I've also adjusted
the add handling to more accurately account for any type splitting
that occurs before we reach a legal type.
Differential Revision: https://reviews.llvm.org/D76478
Previously we multiplied the cost for the table entries by the number of splits needed. But that implies that each split goes through a reduction to scalar independently. I think what really happens is that the we AND/OR the split pieces until we're down to a single value with a legal type and then do special reduction sequence on that.
So to model that this patch takes the number of splits minus one multiplied by the cost of a AND/OR at the legal element count and adds that on top of the table lookup.
Differential Revision: https://reviews.llvm.org/D76400
Refines the gather/scatter cost model, but also changes the TTI
function getIntrinsicInstrCost to accept an additional parameter
which is needed for the gather/scatter cost evaluation.
This did require trivial changes in some non-ARM backends to
adopt the new parameter.
Extending gathers and truncating scatters are now priced cheaper.
Differential Revision: https://reviews.llvm.org/D75525
This tries to improve the accuracy of extract/insert element costs by accounting for subvector extraction/insertion for >128-bit vectors and the shuffling of elements to/from the 0'th index.
It also adds INSERTPS for f32 types and PINSR/PEXTR costs for integer types (at the moment we assume the same cost as MOVD/MOVQ - which isn't always true).
Differential Revision: https://reviews.llvm.org/D74976
D74976 will handle larger vector types, but since SLM doesn't support AVX+ then we will always be extracting from 128-bit vectors so don't need to scale the cost.
We seem to be inheriting the cost from sse4.1. But if we have 256-bit registers we should be able to do this with just one extract to split the 16i16 and two v8i16->v8i32 operations so our cost should be 3 not 4.
Differential Revision: https://reviews.llvm.org/D73646
Summary:
Previously we did this with isel patterns that used garbage in
the widened part of the source. But that's not valid for strictfp.
So now we custom widen and use zeroes for the widened elemens for
strictfp.
This replaces D71864.
Reviewers: RKSimon, spatel, andrew.w.kaylor, pengfei, LiuChen3
Reviewed By: pengfei
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D71879
Add an extra parameter so alignment can be taken under
consideration in gather/scatter legalization.
Differential Revision: https://reviews.llvm.org/D71610
Soon Intrinsic::ID will be a plain integer, so this overload will not be
possible.
Rename both overloads to ensure that downstream targets observe this as
a build failure instead of a runtime failure.
Split off from D71320
Reviewers: efriedma
Differential Revision: https://reviews.llvm.org/D71381
This attempts to teach the cost model in Arm that code such as:
%s = shl i32 %a, 3
%a = and i32 %s, %b
Can under Arm or Thumb2 become:
and r0, r1, r2, lsl #3
So the cost of the shift can essentially be free. To do this without
trying to artificially adjust the cost of the "and" instruction, it
needs to get the users of the shl and check if they are a type of
instruction that the shift can be folded into. And so it needs to have
access to the actual instruction in getArithmeticInstrCost, which if
available is added as an extra parameter much like getCastInstrCost.
We otherwise limit it to shifts with a single user, which should
hopefully handle most of the cases. The list of instruction that the
shift can be folded into include ADC, ADD, AND, BIC, CMP, EOR, MVN, ORR,
ORN, RSB, SBC and SUB. This translates to Add, Sub, And, Or, Xor and
ICmp.
Differential Revision: https://reviews.llvm.org/D70966
This is a follow-up to D70607 where we made any
extract element on SLM more costly than default. But that is
pessimistic for extract from element 0 because that corresponds
to x86 movd/movq instructions. These generally have >1 cycle
latency, but they are probably implemented as single uop
instructions.
Note that no vectorization tests are affected by this change.
Also, no targets besides SLM are affected because those are
falling through to the default cost of 1 anyway. But this will
become visible/important if we add more specializations via cost
tables.
Differential Revision: https://reviews.llvm.org/D71023
I'm not sure what the effect of this change will be on all of the affected
tests or a larger benchmark, but it fixes the horizontal add/sub problems
noted here:
https://reviews.llvm.org/D59710?vs=227972&id=228095&whitespace=ignore-most#toc
The costs are based on reciprocal throughput numbers in Agner's tables for
PEXTR*; these appear to be very slow ops on Silvermont.
This is a small step towards the larger motivation discussed in PR43605:
https://bugs.llvm.org/show_bug.cgi?id=43605
Also, it seems likely that insert/extract is the source of perf regressions on
other CPUs (up to 30%) that were cited as part of the reason to revert D59710,
so maybe we'll extend the table-based approach to other subtargets.
Differential Revision: https://reviews.llvm.org/D70607
This is no longer needed after widening legalization as we
custom legalize v8i8 ourselves.
Added entries to the cost model, but bumped the cost slightly
to account for the truncate shuffle that wasn't costed before.
This better represents the kshift+binop we'd get for each stage
before the final extract. Its likely we'll do even better by
doing a kmov and a cmp with a GPR, but this is a good start.
The default handling was costing a worst case single source
permute shuffle of the vector before the binop. This worst
case assumes the shuffle might have to be emulated with
extracts and inserts. But since we know we're doing a reduction
we can assume we'll get kshift lowering.
There's still some room for improvement here, but this is
much better than it was.
Update TargetTransformInfo to allow AVX1 to use YMM registers for memcmp.
This is a follow up to D68632 which enabled XOR compares which made this possible.
This also updates the memcmp-optsize.ll test unlike the first patch.
https://reviews.llvm.org/D69658
Update TargetTransformInfo to allow AVX1 to use YMM registers for memcmp.
This is a follow up to D68632 which enabled XOR compares which made this possible.
https://reviews.llvm.org/D69658
Summary:
We don't pattern match pairwise shuffles in SelectionDAG. So we
should only return the optimized costs if its not a pairwise
shuffle.
I think SLP vectorizer gives priority to non pairwise shuffle if
the cost is the same. And the look up for reduction intrinsics
passes false for the pairwise flag. So this probably has no real
effect today.
Reviewers: RKSimon
Reviewed By: RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D69083
Add specific scalar costs for CTLZ instructions, we can't discriminate between CTLZ and CTLZ_ZERO_UNDEF so we have to assume the worst. Given how BSR is often a microcoded nightmare on some older targets we might still be underestimating it.
For targets supporting LZCNT (Intel Haswell+ or AMD Fam10+), we provide overrides that assume 1cy costs.
llvm-svn: 374786
Add specific scalar costs for ctpop instructions, these are based on the llvm-mca's SLM throughput numbers (the oldest model we have).
For targets supporting POPCNT, we provide overrides that assume 1cy costs.
llvm-svn: 374775
Roman Lebedev