Cache the LoadSDNode nodes so we can easily map to/from the element index instead of packing them together - this will be useful for future patches for PR16739 etc.
llvm-svn: 365620
This patch checks to see if the vector element loads are based off a dereferenceable pointer that covers the entire vector width, in which case we don't need to have element loads at both extremes of the vector width - just the start (base pointer) of it.
Another step towards partial vector loads......
Differential Revision: https://reviews.llvm.org/D64205
llvm-svn: 365614
This should prevent doing this on pre-sse4.1 targets or for 256
bit vectors without avx2.
I don't know of a failure from this. Op legalization will probably
take care of, but seemed better to be safe.
llvm-svn: 365577
Basically the problem is that X86 doesn't set the Fast flag from
allowsMemoryAccess on certain CPUs due to slow unaligned memory
subtarget features. This prevents bitcasts from being folded into
loads and stores. But all vector loads and stores of the same width
are the same cost on X86.
This patch merges the allowsMemoryAccess call into isLoadBitCastBeneficial to allow X86 to skip it.
Differential Revision: https://reviews.llvm.org/D64295
llvm-svn: 365549
Summary:
This makes it so that IR files using triples without an environment work
out of the box, without normalizing them.
Typically, the MSVC behavior is more desirable. For example, it tends to
enable things like constant merging, use of associative comdats, etc.
Addresses PR42491
Reviewers: compnerd
Subscribers: hiraditya, dexonsmith, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D64109
llvm-svn: 365387
Bitcast v4i32 to v8f32 and back again - it might be worth adding isel patterns for X86PShufd v8i32 on AVX1 targets like we did for X86Blendi to avoid the bitcasts?
llvm-svn: 365125
If we have more then 2 shuffle ops to combine, try to use combineX86ShuffleChainWithExtract to see if some are from the same super vector.
llvm-svn: 365050
iff the number of elements doesn't change.
This gets around an issue with combineX86ShuffleChain not being able to hint which domain is preferred for shuffles that can be done with either.
Fixes regression introduced in rL365041
llvm-svn: 365044
This better accounts for the cost/benefit of removing extract_subvectors from the shuffle and will be more useful in future patches.
The vpermq predicate regression will be fixed shortly.
llvm-svn: 365041
Assert that the shift amount is in range and create vXi8 shift masks in a way that doesn't cause MSVC/cppcheck shift result is truncated then extended warnings.
llvm-svn: 365024
Don't use APInt::getZExtValue() if you can avoid it - eventually someone will call it with i128 or something that doesn't fit into 64-bits.
In this case it was completely superfluous as we'd moved the rest of the code to always use APInt.
Fixes the <1 x i128> addition bug in PR42486
llvm-svn: 364953
Pull out CombineShuffleWithExtract lambda to new combineX86ShuffleChainWithExtract wrapper and refactored it to handle more than 2 shuffle inputs - this will allow combineX86ShufflesRecursively to call this in a future patch.
llvm-svn: 364924
We were relying on combineX86ShufflesRecursively to handle this - this patch gets it done earlier which should make it easier for other code to use resolveTargetShuffleInputsAndMask.
llvm-svn: 364906
These instructions only read 64-bits of memory so we shouldn't
allow a full vector width load to be pattern matched in case it
is marked volatile.
Instead allow vzload or scalar_to_vector+load.
Also add a DAG combine to turn full vector loads into vzload when
used by one of these instructions if the load isn't volatile.
This fixes another case for PR42079
llvm-svn: 364838
We can already widenSubVector to a specific type (of the same scalar type) - this variant just specifies the target vector size.
This will be useful when CombineShuffleWithExtract relaxes the need to have the same scalar type for all shuffle operand subvector sources.
llvm-svn: 364803
But only when the load isn't volatile.
This improves load folding during isel where we only have vzload
and scalar_to_vector+load patterns. We can't have full vector load
isel patterns for the same volatile load issue.
Also add some missing masked cvtsi2fp/cvtui2fp with vzload patterns.
llvm-svn: 364728
AVX masked loads only support 0 as the value for masked off elements.
So we need an extra blend to support other values. Previously we
expanded the masked load to two instructions with isel patterns.
With this patch we now insert the vselect during lowering and it
will be separately selected as a blend.
llvm-svn: 364718
We were requiring that both shuffle operands were EXTRACT_SUBVECTORs, but we can relax this to only require one of them to be.
Also, we shouldn't bother attempting this if both operands are from the lowest subvector (or not EXTRACT_SUBVECTOR at all).
llvm-svn: 364644
We already had the infrastructure for this, but were waiting for the fix for a number of regressions which were handled by the recent shuffle(extract_subvector(),extract_subvector()) -> extract_subvector(shuffle()) shuffle combines
llvm-svn: 364569
While lowering calls, collect info about registers that forward arguments
into following function frame. We store such info into the MachineFunction
of the call. This is used very late when dumping DWARF info about
call site parameters.
([9/13] Introduce the debug entry values.)
Co-authored-by: Ananth Sowda <asowda@cisco.com>
Co-authored-by: Nikola Prica <nikola.prica@rt-rk.com>
Co-authored-by: Ivan Baev <ibaev@cisco.com>
Differential Revision: https://reviews.llvm.org/D60715
llvm-svn: 364516
Without the fix gcc 7.4.0 complains with
../lib/Target/X86/X86ISelLowering.cpp: In function 'bool getFauxShuffleMask(llvm::SDValue, llvm::SmallVectorImpl<int>&, llvm::SmallVectorImpl<llvm::SDValue>&, llvm::SelectionDAG&)':
../lib/Target/X86/X86ISelLowering.cpp:6690:36: error: enumeral and non-enumeral type in conditional expression [-Werror=extra]
int Idx = (ZeroMask[j] ? SM_SentinelZero : (i + j + Ofs));
~~~~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
cc1plus: all warnings being treated as errors
llvm-svn: 364507
This patch rewrites the loop iteration to only visit every other element starting with element 0. And we work on the "even" element and "next" element at the same time. The "First" logic has been moved to the bottom of the loop and doesn't run on every element. I believe it could create dangling nodes previously since we didn't check if we were going to use SCALAR_TO_VECTOR for the first insertion. I got rid of the "First" variable and just do a null check on V which should be equivalent. We also no longer use undef as the starting V for vectors with no zeroes to avoid false dependencies. This matches v8i16.
I've changed all the extends and OR operations to use MVT::i32 since that's what they'll be promoted to anyway. I've tried to use zero_extend only when necessary and use any_extend otherwise. This resulted in some improvements in tests where we are now able to promote aligned (i32 (extload i8)) to a 32-bit load.
Differential Revision: https://reviews.llvm.org/D63702
llvm-svn: 364469
This was trying to optimize concat_vectors with zero of setcc or
kand instructions. But I think it produced the same code we
produce for a concat_vectors with 0 even it it doesn't come from
one of those operations.
llvm-svn: 364463
Create a per-byte shuffle mask based on the computeKnownBits from each operand - if for each byte we have a known zero (or both) then it can be safely blended.
Fixes PR41545
llvm-svn: 364458
We have (almost) no target opcodes that have scalar/vector equivalents - for now assume we can't scalarize them (we can add exceptions if we need to).
llvm-svn: 364429
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
truncateVectorWithPACK is often used in conjunction with ComputeNumSignBits which struggles when peeking through bitcasts.
This fix tries to avoid bitcast(shuffle(bitcast())) patterns in the 256-bit 64-bit sublane shuffles so we can still see through at least until lowering when the shuffles will need to be bitcasted to widen the shuffle type.
llvm-svn: 364401
We currently have some isel patterns for treating vzmovl+load the same as vzload, but that shrinks the load which we shouldn't do if the load is volatile.
Rather than adding isel checks for volatile. This patch removes the patterns and teachs DAG combine to merge them into vzload when its legal to do so.
Differential Revision: https://reviews.llvm.org/D63665
llvm-svn: 364333
In LowerBuildVectorv16i8 we took care to use an any_extend if the first pair is in the lower 16-bits of the vector and no elements are 0. So bits [31:16] will be undefined. But we still emitted a vzext_movl to ensure that bits [127:32] are 0. If we don't need any zeroes we should be consistent and make all of 127:16 undefined.
In LowerBuildVectorv8i16 we can just delete the vzext_movl code because we only use the scalar_to_vector when there are no zeroes. So the vzext_movl is always unnecessary.
Found while investigating whether (vzext_movl (scalar_to_vector (loadi32)) patterns are necessary. At least one of the cases where they were necessary was where the loadi32 matched 32-bit aligned 16-bit extload. Seemed weird that we required vzext_movl for that case.
Differential Revision: https://reviews.llvm.org/D63700
llvm-svn: 364207
This patch does a few things to start cleaning up the isFNEG function.
-Remove the Op0/Op1 peekThroughBitcast calls that seem unnecessary. getTargetConstantBitsFromNode has its own peekThroughBitcast inside. And we have a separate peekThroughBitcast on the return value.
-Add a check of the scalar size after the first peekThroughBitcast to ensure we haven't changed the element size and just did something like f32->i32 or f64->i64.
-Remove an unnecessary check that Op1's type is floating point after the peekThroughBitcast. We're just going to look for a bit pattern from a constant. We don't care about its type.
-Add VT checks on several places that consume the return value of isFNEG. Due to the peekThroughBitcasts inside, the type of the return value isn't guaranteed. So its not safe to use it to build other nodes without ensuring the type matches the type being used to build the node. We might be able to replace these checks with bitcasts instead, but I don't have a test case so a bail out check seemed better for now.
Differential Revision: https://reviews.llvm.org/D63683
llvm-svn: 364206
Ideally we'd be able to represent this truncate as a any_extend to
v16i32 and a truncate, but SelectionDAG doens't know how to not
fold those together.
We have isel patterns to use a vpmovzxwd+vpdmovdb for the truncate,
but we aren't able to simultaneously fold the load and the store
from the isel pattern. By pulling the truncate into the store we
can successfully hide it from the DAG combiner. Then we can isel
pattern match the truncstore and load+any_extend separately.
llvm-svn: 364163
128/256 bit scalar_to_vectors are canonicalized to (insert_subvector undef, (scalar_to_vector), 0). We have isel patterns that try to match this pattern being used by a vzmovl to use a 128-bit instruction and a subreg_to_reg.
This patch detects the insert_subvector undef portion of this and pulls it through the vzmovl, creating a narrower vzmovl and an insert_subvector allzeroes. We can then match the insertsubvector into a subreg_to_reg operation by itself. Then we can fall back on existing (vzmovl (scalar_to_vector)) patterns.
Note, while the scalar_to_vector case is the motivating case I didn't restrict to just that case. I'm also wondering about shrinking any 256/512 vzmovl to an extract_subvector+vzmovl+insert_subvector(allzeros) but I fear that would have bad implications to shuffle combining.
I also think there is more canonicalization we can do with vzmovl with loads or scalar_to_vector with loads to create vzload.
Differential Revision: https://reviews.llvm.org/D63512
llvm-svn: 364095
This should be unreachable, but bugs can make it reachable. This
adds a debug print so we can see the bad node in the output when
the llvm_unreachable triggers.
llvm-svn: 364091
The sat add/sub tests still have unnecessary extract_subvector((vandnps ymm, ymm), 0) uses that should be split to (vandnps (extract_subvector(ymm, 0), extract_subvector(ymm, 0)), but its getting better.
llvm-svn: 364038
This is an exception to the rule that we should prefer xmm ops to ymm ops.
As shown in PR42305:
https://bugs.llvm.org/show_bug.cgi?id=42305
...the store folding opportunity with vextractf128 may result in better
perf by reducing the instruction count.
Differential Revision: https://reviews.llvm.org/D63517
llvm-svn: 363853
We already do this for ZERO_EXTEND/ZERO_EXTEND_VECTOR_INREG - this just extends the pattern matcher to recognize cases where we don't need the zeros in the extension.
llvm-svn: 363841
This allows targets to make more decisions about reserved registers
after isel. For example, now it should be certain there are calls or
stack objects in the frame or not, which could have been introduced by
legalization.
Patch by Matthias Braun
llvm-svn: 363757
FP_ROUND defaults to Legal for all MVT types and nothing changes
the v4f32 entry way from this default. If we needed this line
we'd also need one for v8f32 with AVX512 which we don't have.
llvm-svn: 363719
Part of fixing the X86 regression noted in D63281 - I've split this into X86 and generic parts - the generic commit will be coming shortly and will fix the vector-reduce-mul-widen.ll regression introduced here.
llvm-svn: 363693
If a XMM non-temporal store has less than natural alignment, scalarize the vector - with SSE4A we can stay on the vector and use MOVNTSD(f64), else we must move to GPRs and use MOVNTI(i32/i64).
llvm-svn: 363592
If a YMM/ZMM non-temporal store has less than natural alignment, split the vector - either they will be satisfactorily aligned or will continue to be split until they are XMMs - at which point the legalizer will scalarize it.
llvm-svn: 363582
This is currently only used for ymm->xmm splitting but we shouldn't hardcode the offsets/alignment.
This is necessary for an upcoming patch to split under-aligned non-temporal vector loads.
llvm-svn: 363570
For loads, pre-SSE41 we can't perform NT loads at all, and after that we can only perform vector aligned loads, so if the alignment is less than for a xmm we'll just end up using the regular unaligned vector loads anyway.
First step towards fixing PR42026 - the next step for stores will be to use SSE4A movntsd where possible and to avoid the stack spill on SSE2 targets.
Differential Revision: https://reviews.llvm.org/D63246
llvm-svn: 363564
This is similar logic/motivation to the select splitting in D62969.
In D63233, the pattern changes so that we no longer have an extract_subvector of vselect,
but the operands of the select are still being concatenated.
The closest case is represented in either the first or last test diffs here - we have an
extra instruction, but we converted 3-4 ymm instructions into 4-5 xmm instructions.
I think that's the right trade-off for most AVX1 targets.
In the example based on PR37428:
https://bugs.llvm.org/show_bug.cgi?id=37428
...this makes the loop about 30% faster (tested on Haswell by compiling with -mavx).
Differential Revision: https://reviews.llvm.org/D63364
llvm-svn: 363508
Previously it copied over MachineMemOperands verbatim which caused MOV32rm to have store flags set, and MOV32mr to have load flags set. This fixes some assertions being thrown with EXPENSIVE_CHECKS on.
Committed on behalf of @luke (Luke Lau)
Differential Revision: https://reviews.llvm.org/D62726
llvm-svn: 363268
As discussed on D62910, we need to check whether particular types of memory access are allowed, not just their alignment/address-space.
This NFC patch adds a MachineMemOperand::Flags argument to allowsMemoryAccess and allowsMisalignedMemoryAccesses, and wires up calls to pass the relevant flags to them.
If people are happy with this approach I can then update X86TargetLowering::allowsMisalignedMemoryAccesses to handle misaligned NT load/stores.
Differential Revision: https://reviews.llvm.org/D63075
llvm-svn: 363179
As suggested by @arsenm on D63075 - this adds a TargetLowering::allowsMemoryAccess wrapper that takes a Load/Store node's MachineMemOperand to handle the AddressSpace/Alignment arguments and will also implicitly handle the MachineMemOperand::Flags change in D63075.
llvm-svn: 363048
Summary:
Our default behavior is to use sign_extend for signed comparisons and zero_extend for everything else. But for equality we have the freedom to use either extension. If we can prove the input has been truncated from something with enough sign bits, we can use sign_extend instead and let DAG combine optimize it out. A similar rule is used by type legalization in LegalizeIntegerTypes.
This gets rid of the movzx in PR42189. The immediate will still take 4 bytes instead of the 2 bytes plus 0x66 prefix a cmp di, 32767 would get, but it avoids a length changing prefix.
Reviewers: RKSimon, spatel, xbolva00
Reviewed By: xbolva00
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D63032
llvm-svn: 362920
Summary:
We can only use the memory form of cvtss2sd under optsize due to a partial register update. So previously we were emitting 2 instructions for extload when optimizing for speed. Also due to a late optimization in preprocessiseldag we had to handle (fpextend (loadf32)) under optsize.
This patch forces extload to expand so that it will always be in the (fpextend (loadf32)) form during isel. And when optimizing for speed we can just let each of those pieces select an instruction independently.
Reviewers: spatel, RKSimon
Reviewed By: RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D62710
llvm-svn: 362919
This is a potentially large perf win for AVX1 targets because of the way we
auto-vectorize to 256-bit but then expect the backend to legalize/optimize
for the half-implemented AVX1 ISA.
On the motivating example from PR37428 (even though this patch doesn't solve
the vector shift issue):
https://bugs.llvm.org/show_bug.cgi?id=37428
...there's a 16% speedup when compiling with "-mavx" (perf tested on Haswell)
because we eliminate the remaining 256-bit vblendv ops.
I added comments on a couple of tests that require further work. If we have
256-bit logic ops separating the vselect and extract, we should probably narrow
everything to 128-bit, but that requires a larger pattern match.
Differential Revision: https://reviews.llvm.org/D62969
llvm-svn: 362797
This is intended to enable the use of an immediate blend or
more optimal instruction. But if the passthru is zero we don't
need any additional instructions.
llvm-svn: 362675
As suggested in D62498 - collectConcatOps() matches both
concat_vectors and insert_subvector patterns, and we see
more test improvements by using the more general match.
llvm-svn: 362620
We already handle the case where we combine shuffle(extractsubvector(x),extractsubvector(x)), this relaxes the requirement to permit different sources as long as they have the same value type.
This causes a couple of cases where the VPERMV3 binary shuffles occur at a wider width than before, which I intend to improve in future commits - but as only the subvector's mask indices are defined, these will broadcast so we don't see any increase in constant size.
llvm-svn: 362599
-Use early returns to reduce indentation
-Replace multipe ifs with a switch.
-Replace an assert with an llvm_unreachable default in the switch.
-Check that the FP type we're going to use for the
X86ISD::FAND/FOR/FXOR is legal rather than checking that the
integer type matches the width of a legal scalar fp type. This all
runs after legalization so it shouldn't really matter, but making
sure we're using a valid type in the X86ISD node is really
whats important.
llvm-svn: 362565
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
As discussed on D62777 - we should be able to use this in more SSE41+ cases as well but that requires us to separate it from the OR(AND(),ANDN()) matcher.
llvm-svn: 362504
Move this combine from x86 into generic DAGCombine, which currently only manages cases where the bitcast is between types of the same scalarsize.
Differential Revision: https://reviews.llvm.org/D59188
llvm-svn: 362324
The LoadExt table defaults to all combinations being Legal. For
vector types, only src VTs with an i1 element type were ever changed.
So we don't need to mark them legal manually.
llvm-svn: 362170
We already have good codegen for (vXiY *ext(vXi1 bitcast(iX))) cases, this patch uses it for loads of vXi1 types as well - changing the load into a iX integer load, and bitcasting so that combineToExtendBoolVectorInReg can then use it.
Differential Revision: https://reviews.llvm.org/D62449
llvm-svn: 362081
This patch add the ISD::LRINT and ISD::LLRINT along with new
intrinsics. The changes are straightforward as for other
floating-point rounding functions, with just some adjustments
required to handle the return value being an interger.
The idea is to optimize lrint/llrint generation for AArch64
in a subsequent patch. Current semantic is just route it to libm
symbol.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D62017
llvm-svn: 361875
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
Forking this out of the discussion in D62498
(and assuming that will be committed later, so adding the helper function here).
The LangRef says:
"the backend should never split or merge target-legal volatile load/store instructions."
Differential Revision: https://reviews.llvm.org/D62506
llvm-svn: 361815
We were only testing for direct SETCC results - this allows us to peek through AND/OR/XOR combinations of the comparison results as well.
There's a missing SEXT(PACKSS) fold that I need to investigate for v8i1 cases before I can enable it there as well.
llvm-svn: 361716
If we have a known non-nan operand, place it in the second operand
of fmin/fmax that is returned if either operand is nan.
Differential Revision: https://reviews.llvm.org/D62448
llvm-svn: 361704
This patch adds the overridable TargetLowering::getTargetConstantFromLoad function which allows targets to return any constant value loaded by a LoadSDNode node - only X86 makes use of this so far but everything should be in place for other targets.
computeKnownBits then uses this function to improve codegen, notably vector code after legalization.
A future commit will do the same for ComputeNumSignBits but computeKnownBits sees the bigger benefit.
This required a couple of fixes:
* SimplifyDemandedBits must early-out for getTargetConstantFromLoad cases to prevent infinite loops of constant regeneration (similar to what we already do for BUILD_VECTOR).
* Fix a DAGCombiner::visitTRUNCATE issue as we had trunc(shl(v8i32),v8i16) <-> shl(trunc(v8i16),v8i32) infinite loops after legalization on AVX512 targets.
Differential Revision: https://reviews.llvm.org/D61887
llvm-svn: 361620
Fix for https://bugs.llvm.org/show_bug.cgi?id=41971. Make the
combineVectorSizedSetCCEquality() transform more conservative by
checking that the bitcast to the vector type will be cheap/free
for both operands. I'm considering it cheap if it's a constant,
a load or already a vector. I've dropped the explicit check for
f128 because it should fall out naturally (in the cases where
it'd be detrimental).
Differential Revision: https://reviews.llvm.org/D62220
llvm-svn: 361352
Same as what we do for vector reductions in combineHorizontalPredicateResult, use movmsk+cmp for scalar (and(extract(x,0),extract(x,1)) reduction patterns.
llvm-svn: 361052
Summary:
This refactors four pieces of code that create SDNodes for references to
symbols:
- normal global address lowering (LEA, MOV, etc)
- callee global address lowering (CALL)
- external symbol address lowering (LEA, MOV, etc)
- external symbol address lowering (CALL)
Each of these pieces of code need to:
- classify the reference
- lower the symbol
- emit a RIP wrapper if needed
- emit a load if needed
- add offsets if needed
I think handling them all in one place will make the code easier to
maintain in the future.
Reviewers: craig.topper, RKSimon
Subscribers: hiraditya, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61690
llvm-svn: 360952
This patch add the ISD::LROUND and ISD::LLROUND along with new
intrinsics. The changes are straightforward as for other
floating-point rounding functions, with just some adjustments
required to handle the return value being an interger.
The idea is to optimize lround/llround generation for AArch64
in a subsequent patch. Current semantic is just route it to libm
symbol.
llvm-svn: 360889
They encode the same way, but OR32mi8Locked sets hasUnmodeledSideEffects set
which should be stronger than the mayLoad/mayStore on LOCK_OR32mi8. I think
this makes sense since we are using it as a fence.
This also seems to hide the operation from the speculative load hardening pass
so I've reverted r360511.
llvm-svn: 360747
This was the portion split off D58632 so that it could follow the redzone API cleanup. Note that I changed the offset preferred from -8 to -64. The difference should be very minor, but I thought it might help address one concern which had been previously raised.
Differential Revision: https://reviews.llvm.org/D61862
llvm-svn: 360719
D61068 handled vector shifts, this patch does the same for scalars where there are similar number of pipes for shifts as bit ops - this is true almost entirely for AMD targets where the scalar ALUs are well balanced.
This combine avoids AND immediate mask which usually means we reduce encoding size.
Some tests show use of (slow, scaled) LEA instead of SHL in some cases, but thats due to particular shift immediates - shift+mask generate these just as easily.
Differential Revision: https://reviews.llvm.org/D61830
llvm-svn: 360684
This is a follow on to D58632, with the same logic. Given a memory operation which needs ordering, but doesn't need to modify any particular address, prefer to use a locked stack op over an mfence.
Differential Revision: https://reviews.llvm.org/D61863
llvm-svn: 360649
Returning SDValue() makes the caller think that nothing happened and it will
end up executing the Expand path. This generates extra nodes that will need to
be pruned as dead code.
Returning an ISD::MERGE_VALUES will tell the caller that we'd like to make a
change and it will take care of replacing uses. This will prevent falling into
the Expand path.
llvm-svn: 360627
These are updates to match how isel table would emit a LOCK_OR32mi8 node.
-Use i32 for the immediate zero even though only 8 bits are encoded.
-Use i16 for segment register.
-Use LOCK_OR32mi8 for idempotent atomic operations in 32-bit mode to match
64-bit mode. I'm not sure why OR32mi8Locked and LOCK_OR32mi8 both exist. The
only difference seems to be that OR32mi8Locked is marked as UnmodeledSideEffects=1.
-Emit an extra i32 result for the flags output.
I don't know if the types here really matter just noticed it was inconsistent
with normal behavior.
llvm-svn: 360619
Summary:
X86TargetLowering::LowerAsmOperandForConstraint had better support than
TargetLowering::LowerAsmOperandForConstraint for arbitrary depth
getelementpointers for "i", "n", and "s" extended inline assembly
constraints. Hoist its support from the derived class into the base
class.
Link: https://github.com/ClangBuiltLinux/linux/issues/469
Reviewers: echristo, t.p.northover
Reviewed By: t.p.northover
Subscribers: t.p.northover, E5ten, kees, jyknight, nemanjai, javed.absar, eraman, hiraditya, jsji, llvm-commits, void, craig.topper, nathanchance, srhines
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D61560
llvm-svn: 360604
Fixes the regression noted in D61782 where a VZEXT_MOVL was being inserted because we weren't discriminating between 'zeroable' and 'all undef' for the upper elts.
Differential Revision: https://reviews.llvm.org/D61782
llvm-svn: 360596
Now that we can use HADD/SUB for scalar additions from any pair of extracted elements (D61263), we can relax the one use limit as we will be able to merge multiple uses into using the same HADD/SUB op.
This exposes a couple of missed opportunities in LowerBuildVectorv4x32 which will be committed separately.
Differential Revision: https://reviews.llvm.org/D61782
llvm-svn: 360594
See if we can simplify the demanded vector elts from the extraction before trying to simplify the demanded bits.
This helps us with target shuffles and hops in particular.
llvm-svn: 360535
If we only use the lower xmm of a ymm hop, then extract the xmm's (for free), perform the xmm hop and then insert back into a ymm (for free).
Fixes some of the regressions noted in D61782
llvm-svn: 360435
The current lowering uses an mfence. mfences are substaintially higher latency than the locked operations originally requested, but we do want to avoid contention on the original cache line. As such, use a locked instruction on a cache line assumed to be thread local.
Differential Revision: https://reviews.llvm.org/D58632
llvm-svn: 360393
As reported on PR39920, "slow horizontal ops" targets tend to internally expand to 2*shuffle+add/sub - so if we can reduce 2*shuffle+add/sub to a hadd/sub then we should do it - similar port usage but reduced instruction count.
This works out in most cases, although the "PR22377" regression in vector-shuffle-combining.ll is annoying - going from 2*shuffle+add+shuffle to hadd+2*shuffle - I've opened PR41813 to cover this.
Differential Revision: https://reviews.llvm.org/D61308
llvm-svn: 360360
Summary:
A COFF stub indirects the reference to a symbol through memory. A
.refptr.$sym global variable pointer is created to refer to $sym.
Typically mingw uses these for external global variable declarations,
but we can use them for weak function declarations as well.
Updates the dso_local classification to add a special case for
extern_weak symbols on COFF in both clang and LLVM.
Fixes PR37598
Reviewers: smeenai, mstorsjo
Subscribers: hiraditya, cfe-commits, llvm-commits
Tags: #clang, #llvm
Differential Revision: https://reviews.llvm.org/D61615
llvm-svn: 360207
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
The FR32/FR64/VR128/VR256 register classes don't contain the upper 16 registers. For most cases we use the default implementation which will find any register class that contains the register in question if the VT is legal for the register class. But if the VT is i32 or i64, we won't find a matching register class and will instead up in the code modified in this patch.
If the requested register is x/y/zmm16-31 we weren't returning a register class that contains those registers and will hit an assertion in the caller.
To fix this, I've changed to use the extended register class instead. I don't believe we need a subtarget check to see if avx512 is enabled. The default implementation just pick whatever register class it finds first. I checked and we currently pick FR32X for XMM0 with an f32 type using the default implementation regardless of whether avx512 is enabled. So I assume its it is ok to do the same for i32.
Differential Revision: https://reviews.llvm.org/D61457
llvm-svn: 360102
Summary:
1. Enable infrastructure of AVX512_BF16, which is supported for BFLOAT16 in Cooper Lake;
2. Enable VCVTNE2PS2BF16, VCVTNEPS2BF16 and DPBF16PS instructions, which are Vector Neural Network Instructions supporting BFLOAT16 inputs and conversion instructions from IEEE single precision.
VCVTNE2PS2BF16: Convert Two Packed Single Data to One Packed BF16 Data.
VCVTNEPS2BF16: Convert Packed Single Data to Packed BF16 Data.
VDPBF16PS: Dot Product of BF16 Pairs Accumulated into Packed Single Precision.
For more details about BF16 isa, please refer to the latest ISE document: https://software.intel.com/en-us/download/intel-architecture-instruction-set-extensions-programming-reference
Author: LiuTianle
Reviewers: craig.topper, smaslov, LuoYuanke, wxiao3, annita.zhang, RKSimon, spatel
Reviewed By: craig.topper
Subscribers: kristina, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60550
llvm-svn: 360017
The default impementation in the base class for TargetLowering::getRegForInlineAsmConstraint doesn't work for mask registers when the VT is a scalar type integer types since the only legal mask types are vXi1. So we end up just getting whatever the first register class that contains the register. Currently this appears to be VK1, but its really dependent on the order tablegen outputs the register classes.
Some code in the caller ends up looking up the type for this register class and find v1i1 then generates a copyfromreg from the physical k-register with the v1i1 type. Then it generates an any_extend from v1i1 to the scalar VT which isn't legal. This bad any_extend sticks around until isel where it selects a MOVZX32rr8 with a v1i1 input or maybe a i8 input. Not sure but eventually we pick up a copy from VK1 to GR8 in MachineIR which isn't supported. This leads to a failure in physical register copying.
This patch uses the scalar type to find a VK class of the right size. In the attached test case this will be VK16. This causes a bitcast from vk16 to i16 to be generated instead of an any_extend. This will be properly iseled to a VK16 to GR32 copy and a GR32->GR16 extract_subreg.
Fixes PR41678
Differential Revision: https://reviews.llvm.org/D61453
llvm-svn: 359837
Limiting scalar hadd/hsub generation to the lowest xmm looks to be unnecessary - we will be extracting one upper xmm whatever, and we can remove a shuffle by using the hop which is inline with what shouldUseHorizontalOp expects to happen anyway.
Testing on btver2 (the main target for fast-hops) shows this is beneficial even for float ops where we have a 'shuffle' to extract the float result:
https://godbolt.org/z/0R-U-K
Differential Revision: https://reviews.llvm.org/D61426
llvm-svn: 359786
We already perform horizontal add/sub if we extract from elements 0 and 1, this patch extends it to non-0/1 element extraction indices (as long as they are from the lowest 128-bit vector).
Differential Revision: https://reviews.llvm.org/D61263
llvm-svn: 359707
This is an alternative to D59669 which more aggressively extracts i1 elements from vXi1 bool vectors using a MOVMSK.
Differential Revision: https://reviews.llvm.org/D61189
llvm-svn: 359666
Current LLVM uses pxor+pinsrb on SSE4+ for INSERT_VECTOR_ELT(ZeroVec, 0, Elt) insead of much simpler movd.
INSERT_VECTOR_ELT(ZeroVec, 0, Elt) is idiomatic construct which is used e.g. for _mm_cvtsi32_si128(Elt) and for lowest element initialization in _mm_set_epi32.
So such inefficient lowering leads to significant performance digradations in ceratin cases switching from SSSE3 to SSE4.
https://bugs.llvm.org/show_bug.cgi?id=41512
Here INSERT_VECTOR_ELT(ZeroVec, 0, Elt) is simply converted to SCALAR_TO_VECTOR(Elt) when applicable since latter is closer match to desired behavior and always efficiently lowered to movd and alike.
Committed on behalf of @Serge_Preis (Serge Preis)
Differential Revision: https://reviews.llvm.org/D60852
llvm-svn: 359545
The MachineFunction wasn't used in getOptimalMemOpType, but more importantly,
this allows reuse of findOptimalMemOpLowering that is calling getOptimalMemOpType.
This is the groundwork for the changes in D59766 and D59787, that allows
implementation of TTI::getMemcpyCost.
Differential Revision: https://reviews.llvm.org/D59785
llvm-svn: 359537
Add target shuffle decoding to isHorizontalBinOp as well as ISD::VECTOR_SHUFFLE support.
This does mean we can go through bitcasts so we need to bitcast the extracted args to ensure they are the correct type
Fixes PR39936 and should help with PR39920/PR39921
Differential Revision: https://reviews.llvm.org/D61245
llvm-svn: 359491
Some of the combines might be further improved if we lower more shuffles with X86ISD::VPERMV3 directly, instead of waiting to combine the results.
llvm-svn: 359400
An xor reduction of a bool vector can be optimized to a parity check of the MOVMSK/BITCAST'd integer - if the population count is odd return 1, else return 0.
Differential Revision: https://reviews.llvm.org/D61230
llvm-svn: 359396
As predicate masks are legal on AVX512 targets, we avoid MOVMSK in these cases, but we can just bitcast the bool vector to the integer equivalent directly - avoiding expansion of the reduction to a shuffle pattern.
llvm-svn: 359386
Fixes PR40332 in the limited case where we're selecting between a target shuffle and a zero vector.
We can extend this in the future to handle more opcodes and non-zero selections.
llvm-svn: 359378
Summary: If we have SSE2 we can use a MOVQ to store 64-bits and avoid falling back to a cmpxchg8b loop. If its a seq_cst store we need to insert an mfence after the store.
Reviewers: spatel, RKSimon, reames, jfb, efriedma
Reviewed By: RKSimon
Subscribers: hiraditya, dexonsmith, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60546
llvm-svn: 359368
Create a matchBitOpReduction helper that checks for the pattern with any opcode.
First step towards reusing this code to recognize other scalar reduction patterns.
llvm-svn: 359296
As detailed on PR40758, Bobcat/Jaguar can perform vector immediate shifts on the same pipes as vector ANDs with the same latency - so it doesn't make sense to replace a shl+lshr with a shift+and pair as it requires an additional mask (with the extra constant pool, loading and register pressure costs).
Differential Revision: https://reviews.llvm.org/D61068
llvm-svn: 359293
A small step towards combining shuffles across vector sizes - this recognizes when a shuffle's operands are all extracted from the same larger source and tries to combine to an unary shuffle of that source instead. Fixes one of the test cases from PR34380.
Differential Revision: https://reviews.llvm.org/D60512
llvm-svn: 359292
Truncate the movmsk scalar integer result to the equivalent scalar integer width as before but then bitcast to the requested type.
We still have the issue identified in PR41594 but D61114 should handle this.
llvm-svn: 359176
If the types don't match, we can't just remove the shuffle.
There may be some other opportunity for optimization here,
but this should prevent the crashing seen in:
https://bugs.llvm.org/show_bug.cgi?id=41414
llvm-svn: 359095
Circling back to a leftover bit from PR39859:
https://bugs.llvm.org/show_bug.cgi?id=39859#c1
...we have this counter-intuitive (based on the test diffs) opportunity to use 'psubus'.
This appears to be the better perf option for both Haswell and Jaguar based on llvm-mca.
We already do this transform for the SETULT predicate, so this makes the code more
symmetrical too. If we have pminub/pminuw, we prefer those, so this should not affect
anything but pre-SSE4.1 subtargets.
$ cat before.s
movdqa -16(%rip), %xmm2 ## xmm2 = [32768,32768,32768,32768,32768,32768,32768,32768]
pxor %xmm0, %xmm2
pcmpgtw -32(%rip), %xmm2 ## xmm2 = [255,255,255,255,255,255,255,255]
pand %xmm2, %xmm0
pandn %xmm1, %xmm2
por %xmm2, %xmm0
$ cat after.s
movdqa -16(%rip), %xmm2 ## xmm2 = [256,256,256,256,256,256,256,256]
psubusw %xmm0, %xmm2
pxor %xmm3, %xmm3
pcmpeqw %xmm2, %xmm3
pand %xmm3, %xmm0
pandn %xmm1, %xmm3
por %xmm3, %xmm0
$ llvm-mca before.s -mcpu=haswell
Iterations: 100
Instructions: 600
Total Cycles: 909
Total uOps: 700
Dispatch Width: 4
uOps Per Cycle: 0.77
IPC: 0.66
Block RThroughput: 1.8
$ llvm-mca after.s -mcpu=haswell
Iterations: 100
Instructions: 700
Total Cycles: 409
Total uOps: 700
Dispatch Width: 4
uOps Per Cycle: 1.71
IPC: 1.71
Block RThroughput: 1.8
Differential Revision: https://reviews.llvm.org/D60838
llvm-svn: 358999
This was supposed to be NFC, but the change in SDLoc
definitions causes instruction scheduling changes.
There's nothing x86-specific in this code, and it can
likely be used from DAGCombiner's simplifyVBinOp().
llvm-svn: 358930
Summary:
If you pass two 1024 bit vectors in IR with AVX2 on Windows 64. Both vectors will be split in four 256 bit pieces. The four pieces of the first argument will be passed indirectly using 4 gprs. The second argument will get passed via pointers in memory.
The PartOffsets stored for the second argument are all in terms of its original 1024 bit size. So the PartOffsets for each piece are 32 bytes apart. So if we consider it for copy elision we'll only load an 8 byte pointer, but we'll move the address 32 bytes. The stack object size we create for the first part is probably wrong too.
This issue was encountered by ISPC. I'm working on getting a reduce test case, but wanted to go ahead and get feedback on the fix.
Reviewers: rnk
Reviewed By: rnk
Subscribers: dbabokin, llvm-commits, hiraditya
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60801
llvm-svn: 358817
combineVectorTruncationWithPACKUS is currently splitting the upper bit bit masking into 128-bit subregs and then concatenating them back together.
This was originally done to avoid regressions that caused existing subregs to be concatenated to the larger type just for the AND masking before being extracted again. This was fixed by @spatel (notably rL303997 and rL347356).
This also lets SimplifyDemandedBits do some further improvements before it hits the recursive depth limit.
My only annoyance with this is that we were broadcasting some xmm masks but we seem to have lost them by moving to ymm - but that's a known issue as the logic in lowerBuildVectorAsBroadcast isn't great.
Differential Revision: https://reviews.llvm.org/D60375#inline-539623
llvm-svn: 358692
This replaces the MOVMSK combine introduced at D52121/rL342326
(movmsk (setne (and X, (1 << C)), 0)) -> (movmsk (X << C))
with the more general icmp lowering so it can pick up more cases through bitcasts - notably vXi8 cases which use vXi16 shifts+masks, this patch can remove the mask and use pcmpgtb(0,x) for the sra.
Differential Revision: https://reviews.llvm.org/D60625
llvm-svn: 358651
As discussed on PR41359, this patch renames the pair of shift-mask target feature functions to make their purposes more obvious.
shouldFoldShiftPairToMask -> shouldFoldConstantShiftPairToMask
preferShiftsToClearExtremeBits -> shouldFoldMaskToVariableShiftPair
llvm-svn: 358526
Improves codegen demonstrated by D60512 - instructions represented by X86ISD::PERMV/PERMV3 can never memory fold the operand used for their index register.
This patch updates the 'isUseOfShuffle' helper into the more capable 'isFoldableUseOfShuffle' that recognises that the op is used for a X86ISD::PERMV/PERMV3 index mask and can't be folded - allowing us to use broadcast/subvector-broadcast ops to reduce the size of the mask constant pool data.
Differential Revision: https://reviews.llvm.org/D60562
llvm-svn: 358516
Currently combineHorizontalPredicateResult only handles anyof/allof reduction patterns of legal types, which can be tricky to match as type legalization of bools can introduce bitcasts/truncs/extensions.
This patch extends combineHorizontalPredicateResult to recognise vXi1 bool reductions as well and uses the existing combineBitcastvxi1 helper to create the MOVMSK necessary to then compare the signmask result.
This ensures the accuracy of the reduction costs added in D60403 which assume the MOVMSK generation.
Differential Revision: https://reviews.llvm.org/D60610
llvm-svn: 358286
If the vector setcc has been legalized then we will need to convert a vector boolean of 0 or -1 to a scalar boolean of 0 or 1.
The added test case previously crashed in 32-bit mode by creating a setcc with an i64 condition that type legalization couldn't expand.
llvm-svn: 358218
This patch adds patterns for turning bitcasted atomic load/store into movss/sd.
It also removes the pseudo instructions for atomic RMW fadd. Instead just adding isel patterns for folding an atomic load into addss/sd. And relying on the new movss/sd store pattern to handle the write part.
This also makes the fadd patterns use VEX and EVEX instructions when AVX or AVX512F are enabled.
Differential Revision: https://reviews.llvm.org/D60394
llvm-svn: 358215
With correct test checks this time.
If we have X87, but not SSE2 we can atomicaly load an i64 value into the significand of an 80-bit extended precision x87 register using fild. We can then use a fist instruction to convert it back to an i64 integ
This matches what gcc and icc do for this case and removes an existing FIXME.
llvm-svn: 358214
If we have X87, but not SSE2 we can atomicaly load an i64 value into the significand of an 80-bit extended precision x87 register using fild. We can then use a fist instruction to convert it back to an i64 integer and store it to a stack temporary. From there we can do two 32-bit loads to get the value into integer registers without worrying about atomicness.
This matches what gcc and icc do for this case and removes an existing FIXME.
Differential Revision: https://reviews.llvm.org/D60156
llvm-svn: 358211
Certain optimisations from ConstantHoisting and CGP rely on Selection DAG not
seeing through to the constant in other blocks. Revert this patch while we come
up with a better way to handle that.
I will try to follow this up with some better tests.
llvm-svn: 358113
Returning SDValue() makes the caller think custom lowering was unsuccessful and then it will fall back to trying to expand the original node. This expanded code will end up with no users and end up being pruned later. But it was useless unnecessary work to create it.
Instead return a MERGE_VALUES with all the results so the caller knows something changed. The caller can handle the replacements.
For one of the cases I had to use UNDEF has a dummy value for a result we know is unused. This should get pruned later.
llvm-svn: 357935
I was looking at a potential DAGCombiner fix for 1 of the regressions in D60278, and it caused severe regression test pain because x86 TLI lies about the desirability of 8-bit shift ops.
We've hinted at making all 8-bit ops undesirable for the reason in the code comment:
// TODO: Almost no 8-bit ops are desirable because they have no actual
// size/speed advantages vs. 32-bit ops, but they do have a major
// potential disadvantage by causing partial register stalls.
...but that leads to massive diffs and exposes all kinds of optimization holes itself.
Differential Revision: https://reviews.llvm.org/D60286
llvm-svn: 357912
Previously LowerOperationWrapper took the number of results from the original
node and counted that many results from the new node. This was intended to drop
chain operands from FP_TO_SINT lowering that uses X87 with memory operations to
stack temporaries. The final load had an extra chain output that needs to be
ignored.
Unfortunately, it didn't work with scatter which has 2 result operands, the
mask output which is discarded and a chain output. The chain output is the one
that is needed but it comes second and it would be dropped by the previous
logic here. To workaround this we were doing a ReplaceAllUses in the lowering
code so that the generic legalization code wouldn't see any uses to replace
since it had been given the wrong result/type.
After this change we take the LowerOperation result directly if the original
node has one result. This allows us to directly return the chain from scatter
or the load data from the FP_TO_SINT case. When the original node has multiple
results we'll ensure the returned node has the same number and copy them over.
For cases where the original node has multiple results and the new code for some
reason has even more results, MERGE_VALUES can be used to pass only the needed
results.
llvm-svn: 357887
In the case where we only want the sign bit (e.g. when using PACKSS truncation of comparison results for MOVMSK) then we can just demand the sign bit of the source operands.
This makes use of the fact that PACKSS saturates out of range values to the min/max int values - so the sign bit is always preserved.
Differential Revision: https://reviews.llvm.org/D60333
llvm-svn: 357859
Summary:
This avoids needing an isel pattern for each condition code. And it removes translation switches for converting between Jcc instructions and condition codes.
Now the printer, encoder and disassembler take care of converting the immediate. We use InstAliases to handle the assembly matching. But we print using the asm string in the instruction definition. The instruction itself is marked IsCodeGenOnly=1 to hide it from the assembly parser.
Reviewers: spatel, lebedev.ri, courbet, gchatelet, RKSimon
Reviewed By: RKSimon
Subscribers: MatzeB, qcolombet, eraman, hiraditya, arphaman, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D60228
llvm-svn: 357802
Summary:
Teach SelectionDAG how to compute known bits of ISD::CopyFromReg if
the virtual reg used has one def only.
This can be particularly useful when calling isBaseWithConstantOffset()
with the ISD::CopyFromReg argument, as more optimizations may get enabled
in the result.
Also add a missing truncation on X86, found by testing of this patch.
Change-Id: Id1c9fceec862d118c54a5b53adf72ada5d6daefa
Reviewers: bogner, craig.topper, RKSimon
Reviewed By: RKSimon
Subscribers: lebedev.ri, nemanjai, jvesely, nhaehnle, javed.absar, jsji, jdoerfert, llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D59535
llvm-svn: 357745
We already promote SRL and SHL to i32.
This will introduce sign extends sometimes which might be harder to deal with than the zero we use for promoting SRL. I ran this through some of our internal benchmark lists and didn't see any major regressions.
I think there might be some DAG combine improvement opportunities in the test changes here.
Differential Revision: https://reviews.llvm.org/D60278
llvm-svn: 357743
It unnecessarily breaks previously-working code which used varargs,
but didn't pass any float/double arguments (such as EDK2).
Also revert the fixup on top of that:
Revert [X86] Fix a test from r357317
This reverts r357317 (git commit d413f41de6)
This reverts r357380 (git commit 7af32444b9)
llvm-svn: 357718
These inserters inserted some instructions to zero some registers and copied from virtual registers to physical registers.
This change instead inserts the zeros directly into the DAG at lowering time using new ISD opcodes
that take the extra zeroes as inputs. The zeros will then go through isel on their own to select
the MOV32r0 pseudo. Then we just need to mention the physical registers directly
in the isel patterns and the isel table and InstrEmitter will take care of inserting the necessary
copies to/from physical registers.
llvm-svn: 357659
This custom inserter existed so we could do a weird thing where we pretended that the instructions support
a full address mode instead of taking a pointer in EAX/RAX. I think was largely so we could be pointer
size agnostic in the isel pattern.
To make this work we would then put the address into an LEA into EAX/RAX in front of the instruction after
isel. But the LEA is overkill when we just have a base pointer. So we end up using the LEA as a slower MOV
instruction.
With this change we now just do custom selection during isel instead and just assign the incoming address
of the intrinsic into EAX/RAX based on its size. After the intrinsic is selected, we can let isel take
care of selecting an LEA or other operation to do any address computation needed in this basic block.
I've also split the instruction into a 32-bit mode version and a 64-bit mode version so the implicit
use is properly sized based on the pointer. Without this we get comments in the assembly output about
killing eax and defing rax or vice versa depending on whether we define the instruction to use EAX/RAX.
llvm-svn: 357652
This pattern would show up as a regression if we more
aggressively convert vector FP ops to scalar ops.
There's still a missed optimization for the v4f64 legal
case (AVX) because we create that h-op with an undef operand.
We should probably just duplicate the operands for that
pattern to avoid trouble.
llvm-svn: 357642
One motivation for making this change is that the lack of using movmsk is likely
a main source of perf difference between clang and gcc on the C-Ray benchmark as
shown here:
https://www.phoronix.com/scan.php?page=article&item=gcc-clang-2019&num=5
...but this change alone isn't enough to solve that problem.
The 'all-of' examples show what is likely the worst case trade-off: we end up with
an extra instruction (or 2 if we count the 'xor' register clearing). The 'any-of'
examples look clearly better using movmsk because we've traded 2 vector instructions
for 2 scalar instructions, and movmsk may have better timing than the generic 'movq'.
If we examine the llvm-mca output for these cases, it appears that even though the
'all-of' movmsk variant looks worse on paper, it would perform better on both
Haswell and Jaguar.
$ llvm-mca -mcpu=haswell no_movmsk.s -timeline
Iterations: 100
Instructions: 400
Total Cycles: 504
Total uOps: 400
Dispatch Width: 4
uOps Per Cycle: 0.79
IPC: 0.79
Block RThroughput: 1.0
$ llvm-mca -mcpu=haswell movmsk.s -timeline
Iterations: 100
Instructions: 600
Total Cycles: 358
Total uOps: 600
Dispatch Width: 4
uOps Per Cycle: 1.68
IPC: 1.68
Block RThroughput: 1.5
$ llvm-mca -mcpu=btver2 no_movmsk.s -timeline
Iterations: 100
Instructions: 400
Total Cycles: 407
Total uOps: 400
Dispatch Width: 2
uOps Per Cycle: 0.98
IPC: 0.98
Block RThroughput: 2.0
$ llvm-mca -mcpu=btver2 movmsk.s -timeline
Iterations: 100
Instructions: 600
Total Cycles: 311
Total uOps: 600
Dispatch Width: 2
uOps Per Cycle: 1.93
IPC: 1.93
Block RThroughput: 3.0
Finally, there may be CPUs where movmsk is horribly slow (old AMD small cores?), but if
that's true, then we're also almost certainly making the wrong transform already for
reductions with >2 elements, so that should be fixed independently.
Differential Revision: https://reviews.llvm.org/D59997
llvm-svn: 357367
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
Simon Pilgrim