At first, I tried to thread the x86 needle and use a target hook (isVectorShiftByScalarCheap())
to disable the transform only for non-splat pow-of-2 constants, but not AVX2, but only some
element types, but...it's difficult.
Here we just avoid the loop with the x86 vector transform that conflicts with the general DAG
combine and preserve all of the existing behavior AFAICT otherwise.
Some tests that will probably fail if someone does try to restrict this in a more targeted way
for x86-only may be found in:
test/CodeGen/X86/combine-mul.ll
test/CodeGen/X86/vector-mul.ll
test/CodeGen/X86/widen_arith-5.ll
This should prevent the infinite looping seen with:
https://bugs.llvm.org/show_bug.cgi?id=35579
Differential Revision: https://reviews.llvm.org/D41040
llvm-svn: 320374
This commit is the first part of https://reviews.llvm.org/D40348.
In order to allow target combines to be performed on newly combined
indexed loads, add them back to the worklist. The remainder of the
above patch will be committed in subsequent revisions and will use
this. Test cases will be included with those follow-up commits.
llvm-svn: 320365
This is a preparatory step for D34515.
This change:
- makes nodes ISD::ADDCARRY and ISD::SUBCARRY legal for i32
- lowering is done by first converting the boolean value into the carry flag
using (_, C) ← (ARMISD::ADDC R, -1) and converted back to an integer value
using (R, _) ← (ARMISD::ADDE 0, 0, C). An ARMISD::ADDE between the two
operations does the actual addition.
- for subtraction, given that ISD::SUBCARRY second result is actually a
borrow, we need to invert the value of the second operand and result before
and after using ARMISD::SUBE. We need to invert the carry result of
ARMISD::SUBE to preserve the semantics.
- given that the generic combiner may lower ISD::ADDCARRY and
ISD::SUBCARRYinto ISD::UADDO and ISD::USUBO we need to update their lowering
as well otherwise i64 operations now would require branches. This implies
updating the corresponding test for unsigned.
- add new combiner to remove the redundant conversions from/to carry flags
to/from boolean values (ARMISD::ADDC (ARMISD::ADDE 0, 0, C), -1) → C
- fixes PR34045
- fixes PR34564
- fixes PR35103
Differential Revision: https://reviews.llvm.org/D35192
llvm-svn: 320355
I noticed this pattern in D38316 / D38388. We failed to combine a shuffle that is either
repeating a scalar insertion at the same position in a vector or translated to a different
element index.
Like the earlier patch, this could be an instcombine too, but since we opted to make this
a DAG transform earlier, I've made this one a DAG patch too.
We do not need any legality checking because the new insert is identical to the existing
insert except that it may have a different constant insertion operand.
The constant insertion test in test/CodeGen/X86/vector-shuffle-combining.ll was the
motivation for D38756.
Differential Revision: https://reviews.llvm.org/D40209
llvm-svn: 320050
Search from AND nodes to find whether they can be propagated back to
loads, so that the AND and load can be combined into a narrow load.
We search through OR, XOR and other AND nodes and all bar one of the
leaves are required to be loads or constants. The exception node then
needs to be masked off meaning that the 'and' isn't removed, but the
loads(s) are narrowed still.
Differential Revision: https://reviews.llvm.org/D39604
llvm-svn: 319773
Summary:
Found out, at code inspection, that there was a fault in
DAGCombiner::CombineConsecutiveLoads for big-endian targets.
A BUILD_PAIR is always having the least significant bits of
the composite value in element 0. So when we are doing the checks
for consecutive loads, for big endian targets, we should check
if the load to elt 1 is at the lower address and the load
to elt 0 is at the higher address.
Normally this bug only resulted in missed oppurtunities for
doing the load combine. I guess that in some rare situation it
could lead to faulty combines, but I've not seen that happen.
Note that this patch actually will trigger load combine for
some big endian regression tests.
One example is test/CodeGen/PowerPC/anon_aggr.ll where we now get
t76: i64,ch = load<LD8[FixedStack-9]
instead of
t37: i32,ch = load<LD4[FixedStack-10]>
t35: i32,ch = load<LD4[FixedStack-9]>
t41: i64 = build_pair t37, t35
before legalization. Then the legalization will split the LD8
into two loads, so the end result is the same. That should
verify that the transfomation is correct now.
Reviewers: niravd, hfinkel
Reviewed By: niravd
Subscribers: nemanjai, llvm-commits
Differential Revision: https://reviews.llvm.org/D40444
llvm-svn: 319771
Pull the checks upon the load out from ReduceLoadWidth into their own
function.
Differential Revision: https://reviews.llvm.org/D40833
llvm-svn: 319766
MatchRotate assumes the types of the types of LHS and RHS are equal,
which is always the case then they come from an OR node, but here
we're getting them from two different TRUNC nodes, so we have to check
the types.
llvm-svn: 319695
If the truncation has been pushed past the or-node, look through it and
truncate afterwards.
Differential revision: https://reviews.llvm.org/D40792
llvm-svn: 319692
Both LoadedVT and NarrowLoad are passed as references and neither
of them are used by any of its callers.
Differential Revision: https://reviews.llvm.org/D40713
llvm-svn: 319645
visitAND attempts to narrow the width of extending loads that are
then masked off. ReduceLoadWidth already exists for a similar purpose
and handles shifts, so I've moved the code to handle AND nodes there.
Differential Revision: https://reviews.llvm.org/D39595
llvm-svn: 319421
This is needed for cases when the memory access is not as big as the width of
the data type. For instance, storing i1 (1 bit) would be done in a byte (8
bits).
Using 'BitSize >> 3' (or '/ 8') would e.g. give the memory access of an i1 a
size of 0, which for instance makes alias analysis return NoAlias even when
it shouldn't.
There are no tests as this was done as a follow-up to the bugfix for the case
where this was discovered (r318824). This handles more similar cases.
Review: Björn Petterson
https://reviews.llvm.org/D40339
llvm-svn: 319173
Unoptimized IR can have linear sequences of stores to an array, where the
initial GEP for the first store is formed from the pointer to the array, and the
GEP for each store after the first is formed from the previous GEP with some
offset in an inductive fashion.
The (large) resulting DAG when analyzed by DAGCombine undergoes an excessive
number of combines as each store node is examined every time its' offset node
is combined with any child of the offset. One of the transformations is
findBetterNeighborChains which assists MergeConsecutiveStores. The former
relies on repeated chain walking to do its' work, however MergeConsecutiveStores
is disabled at O0 which makes the transformation redundant.
Any optimization level other than O0 would invoke InstCombine which would
resolve the chain of GEPs into flat base + offset GEP for each store which
does not exhibit the repeated examination of each store to the array.
Disabling this optimization fixes an excessive compile time issue (30~ minutes
for the test case provided) at O0.
Reviewers: niravd, craig.topper, t.p.northover
Differential Revision: https://reviews.llvm.org/D40193
llvm-svn: 319142
With AVX512 vXi1 types are legal so we shouldn't be extending them.
This change is similar to existing code in the zext(setcc) combine.
llvm-svn: 319120
Since i1 is a legal type, this:
NumBytes = Op1->getMemoryVT().getSizeInBits() >> 3;
is wrong and should be instead
NumBytes = Op0->getMemoryVT().getStoreSize();
There seems to be more places where this should be fixed outside DAGCombiner.
Review: Hal Finkel
https://bugs.llvm.org/show_bug.cgi?id=35366
llvm-svn: 318824
All these headers already depend on CodeGen headers so moving them into
CodeGen fixes the layering (since CodeGen depends on Target, not the
other way around).
llvm-svn: 318490
Change the calculation for the desired ValueType for non-sign
extending loads, as in those cases we don't care about the
higher bits. This creates a smaller ExtVT and allows for such
combinations as:
(srl (zextload i16, [addr]), 8) -> (zextload i8, [addr + 1])
Differential Revision: https://reviews.llvm.org/D40034
llvm-svn: 318390
In function DAGCombiner::visitSIGN_EXTEND_INREG, sext can be combined with extload even if sextload is not supported by target, then
if sext is the only user of extload, there is no big difference, no harm no benefit.
if extload has more than one user, the combined sextload may block extload from combining with other zext, causes extra zext instructions generated. As demonstrated by the attached test case.
This patch add the constraint that when sextload is not supported by target, sext can only be combined with extload if it is the only user of extload.
Differential Revision: https://reviews.llvm.org/D39108
llvm-svn: 316802
combineShuffleOfScalars is very conservative about shuffled BUILD_VECTORs that can be combined together.
This patch adds one additional case - if both BUILD_VECTORs represent splats of the same scalar value but with different UNDEF elements, then we should create a single splat BUILD_VECTOR, sharing only the UNDEF elements defined by the shuffle mask.
Differential Revision: https://reviews.llvm.org/D38696
llvm-svn: 316331
This is only currently used for mad/fma transforms.
This is the only case where it should be used for AMDGPU,
so add an opcode to be sure.
llvm-svn: 315740
Eg:
insert v4i32 V, (v2i16 X), 2 --> shuffle v8i16 V', X', {0,1,2,3,8,9,6,7}
This is a generalization of the IR fold in D38316 to handle insertion into a non-undef vector.
We may want to abandon that one if we can't find value in squashing the more specific pattern sooner.
We're using the existing legal shuffle target hook to avoid AVX512 horror with vXi1 shuffles.
There may be room for improvement in the shuffle lowering here, but that would be follow-up work.
Differential Revision: https://reviews.llvm.org/D38388
llvm-svn: 315460
Summary:
See https://llvm.org/PR33743 for more details
It seems that for non-power of 2 vector sizes, the algorithm can produce
non-matching sizes for input and result causing an assert.
This usually isn't a problem as the isAnyExtend check will weed these out, but
in some cases (most often with lots of undefined values for the mask indices) it
can pass this check for non power of 2 vectors.
Adding in an extra check that ensures that bit size will match for the result
and input (as required)
Subscribers: nhaehnle
Differential Revision: https://reviews.llvm.org/D35241
llvm-svn: 315307
Without this, we could end up trying to get the Nth (0-indexed) element
from a subvector of size N.
Differential Revision: https://reviews.llvm.org/D37880
llvm-svn: 314380
At least for the 64-bit and less case, we should be able to determine if we even have a mask without counting any bits. This also removes the need to explicitly check for 0 active bits, isMask will return false for 0.
llvm-svn: 313908
If we have an AssertZext of a truncated value that has already been AssertZext'ed,
we can assert on the wider source op to improve the zext-y knowledge:
assert (trunc (assert X, i8) to iN), i1 --> trunc (assert X, i1) to iN
This moves a fold from being Mips-specific to general combining, and x86 shows
improvements.
Differential Revision: https://reviews.llvm.org/D37017
llvm-svn: 313577
rL310710 allowed store merging to occur after legalization to catch stores that are created late,
but this exposes a logic hole seen in PR34217:
https://bugs.llvm.org/show_bug.cgi?id=34217
We will miss merging stores if the target lowers vector extracts into target-specific operations.
This patch allows store merging to occur both before and after legalization if the target chooses
to get maximum merging.
I don't think the potential regressions in the other tests are relevant. The tests are for
correctness of weird IR constructs rather than perf tests, and I think those are still correct.
Differential Revision: https://reviews.llvm.org/D37987
llvm-svn: 313564
We already have a combine for this pattern when the input to shl is add, so we just need to enable the transformation when the input is or.
Original patch by @tstellar
Differential Revision: https://reviews.llvm.org/D19325
llvm-svn: 313251
Fixes some combine issues for AMDGPU where we weren't
getting the many extract_vector_elt combines expected
in a future patch.
This should really be checking isOperationLegalOrCustom on
the extract. That improves a number of x86 lit tests, but
a few get stuck in an infinite loop from one place
where a similar looking extract is created. I have a
different workaround in the backend for that which
keeps many of those improvements, but also adds a few
regressions.
llvm-svn: 312730
The function combineShuffleToVectorExtend in DAGCombine might generate an illegal typed node after "legalize types" phase, causing assertion on non-simple type to fail afterwards.
Adding a type check in case the combine is running after the type legalize pass.
Differential Revision: https://reviews.llvm.org/D37330
llvm-svn: 312438
