This commit adds some missing intrinsics into the isAlwaysUniform list
for the AMDGPU backend.
Differential Revision: https://reviews.llvm.org/D56845
llvm-svn: 351562
compiler identification lines in test-cases.
(Doing so only because it's then easier to search for references which
are actually important and need fixing.)
llvm-svn: 351200
This fixes https://bugs.llvm.org/show_bug.cgi?id=40110.
This implements handling of undef operands for integer intrinsics in
ConstantFolding, in particular for the bitcounting intrinsics (ctpop,
cttz, ctlz), the with.overflow intrinsics, the saturating math
intrinsics and the funnel shift intrinsics.
The undef behavior follows what InstSimplify does for the general cas
e of non-constant operands. For the bitcount intrinsics (where
InstSimplify doesn't do undef handling -- there cannot be a combination
of an undef + non-constant operand) I'm using a 0 result if the intrinsic
is defined for zero and undef otherwise.
Differential Revision: https://reviews.llvm.org/D55950
llvm-svn: 350971
The new-pm version of DA is untested. Testing requires a printer, so
add that and use it in the existing DA tests.
Differential Revision: https://reviews.llvm.org/D56386
llvm-svn: 350624
Noticed in D56011 - handle the case that scalar fp ops are quicker on P3 than P4
Add the other costs so that we're not relying on the default "is legal/custom" cost logic.
llvm-svn: 350403
GetPointerBaseWithConstantOffset include this code, where ByteOffset
and GEPOffset are both of type llvm::APInt :
ByteOffset += GEPOffset.getSExtValue();
The problem with this line is that getSExtValue() returns an int64_t, but
the += matches an overload for uint64_t. The problem is that the resulting
APInt is no longer considered to be signed. That in turn causes assertion
failures later on if the relevant pointer type is > 64 bits in width and
the GEPOffset was negative.
Changing it to
ByteOffset += GEPOffset.sextOrTrunc(ByteOffset.getBitWidth());
resolves the issue and explicitly performs the sign-extending
or truncation. Additionally, instead of asserting later if the result
is > 64 bits, it breaks out of the loop in that case.
See also
https://reviews.llvm.org/D24729https://reviews.llvm.org/D24772
This commit must be merged after D38662 in order for the test to pass.
Patch by Michael Ferguson <mpfergu@gmail.com>.
Reviewers: reames, sanjoy, hfinkel
Reviewed By: hfinkel
Differential Revision: https://reviews.llvm.org/D38501
llvm-svn: 350395
Motivated by the discussion in D38499, this patch updates BasicAA to support
arbitrary pointer sizes by switching most remaining non-APInt calculations to
use APInt. The size of these APInts is set to the maximum pointer size (maximum
over all address spaces described by the data layout string).
Most of this translation is straightforward, but this patch contains a fix for
a bug that revealed itself during this translation process. In order for
test/Analysis/BasicAA/gep-and-alias.ll to pass, which is run with 32-bit
pointers, the intermediate calculations must be performed using 64-bit
integers. This is because, as noted in the patch, when GetLinearExpression
decomposes an expression into C1*V+C2, and we then multiply this by Scale, and
distribute, to get (C1*Scale)*V + C2*Scale, it can be the case that, even
through C1*V+C2 does not overflow for relevant values of V, (C2*Scale) can
overflow. If this happens, later logic will draw invalid conclusions from the
(base) offset value. Thus, when initially applying the APInt conversion,
because the maximum pointer size in this test is 32 bits, it started failing.
Suspicious, I created a 64-bit version of this test (included here), and that
failed (miscompiled) on trunk for a similar reason (the multiplication can
overflow).
After fixing this overflow bug, the first test case (at least) in
Analysis/BasicAA/q.bad.ll started failing. This is also a 32-bit test, and was
relying on having 64-bit intermediate values to have BasicAA return an accurate
result. In order to fix this problem, and because I believe that it is not
uncommon to use i64 indexing expressions in 32-bit code (especially portable
code using int64_t), it seems reasonable to always use at least 64-bit
integers. In this way, we won't regress our analysis capabilities (and there's
a command-line option added, so experimenting with this should be easy).
As pointed out by Eli during the review, there are other potential overflow
conditions that this patch does not address. Fixing those is left to follow-up
work.
Patch by me with contributions from Michael Ferguson (mferguson@cray.com).
Differential Revision: https://reviews.llvm.org/D38662
llvm-svn: 350220
The current llvm.mem.parallel_loop_access metadata has a problem in that
it uses LoopIDs. LoopID unfortunately is not loop identifier. It is
neither unique (there's even a regression test assigning the some LoopID
to multiple loops; can otherwise happen if passes such as LoopVersioning
make copies of entire loops) nor persistent (every time a property is
removed/added from a LoopID's MDNode, it will also receive a new LoopID;
this happens e.g. when calling Loop::setLoopAlreadyUnrolled()).
Since most loop transformation passes change the loop attributes (even
if it just to mark that a loop should not be processed again as
llvm.loop.isvectorized does, for the versioned and unversioned loop),
the parallel access information is lost for any subsequent pass.
This patch unlinks LoopIDs and parallel accesses.
llvm.mem.parallel_loop_access metadata on instruction is replaced by
llvm.access.group metadata. llvm.access.group points to a distinct
MDNode with no operands (avoiding the problem to ever need to add/remove
operands), called "access group". Alternatively, it can point to a list
of access groups. The LoopID then has an attribute
llvm.loop.parallel_accesses with all the access groups that are parallel
(no dependencies carries by this loop).
This intentionally avoid any kind of "ID". Loops that are clones/have
their attributes modifies retain the llvm.loop.parallel_accesses
attribute. Access instructions that a cloned point to the same access
group. It is not necessary for each access to have it's own "ID" MDNode,
but those memory access instructions with the same behavior can be
grouped together.
The behavior of llvm.mem.parallel_loop_access is not changed by this
patch, but should be considered deprecated.
Differential Revision: https://reviews.llvm.org/D52116
llvm-svn: 349725
This is split from D55452 with the correct patch this time.
Pairwise reductions require two shuffles on every level but the last. On the last level the two shuffles are <1, u, u, u...> and <0, u, u, u...>, but <0, u, u, u...> will be dropped by InstCombine/DAGCombine as being an identity shuffle.
Differential Revision: https://reviews.llvm.org/D55615
llvm-svn: 349072
Attribute 'dso_local' generated in bitcode from compiling
original C file but isn't needed.
Differential Revision: https://reviews.llvm.org/D55521
llvm-svn: 348835
Summary: The comment says we need 3 extracts and a select at the end. But didn't we just account for the select in the vector cost above. Aren't we just extracting the single element after taking the min/max in the vector register?
Reviewers: RKSimon, spatel, ABataev
Reviewed By: RKSimon
Subscribers: javed.absar, kristof.beyls, llvm-commits
Differential Revision: https://reviews.llvm.org/D55480
llvm-svn: 348739
We were overcounting the number of arithmetic operations needed at each level before we reach a legal type. We were using the full vector type for that level, but we are going to split the input vector at that level in half. So the effective arithmetic operation cost at that level is half the width.
So for example on 8i32 on an sse target. Were were calculating the cost of an 8i32 op which is likely 2 for basic integer. Then after the loop we count 2 more v4i32 ops. For a total arith cost of 4. But if you look at the assembly there would only be 3 arithmetic ops.
There are still more bugs in this code that I'm going to work on next. The non pairwise code shouldn't count extract subvectors in the loop. There are no extracts, the types are split in registers. For pairwise we need to use 2 two src permute shuffles.
Differential Revision: https://reviews.llvm.org/D55397
llvm-svn: 348621
DemandedBits and BDCE currently only support scalar integers. This
patch extends them to also handle vector integer operations. In this
case bits are not tracked for individual vector elements, instead a
bit is demanded if it is demanded for any of the elements. This matches
the behavior of computeKnownBits in ValueTracking and
SimplifyDemandedBits in InstCombine.
Unlike the previous iteration of this patch, getDemandedBits() can now
again be called on arbirary (sized) instructions, even if they don't
have integer or vector of integer type. (For vector types the size of the
returned mask will now be the scalar size in bits though.)
The added LoopVectorize test case shows a case which triggered an
assertion failure with the previous attempt, because getDemandedBits()
was called on a pointer-typed instruction.
Differential Revision: https://reviews.llvm.org/D55297
llvm-svn: 348602
In some cases different alignments for function might be used to save
space e.g. thumb mode with -Oz will try to use 2 byte function
alignment. Similar patch that fixed this in other areas exists here
https://reviews.llvm.org/D46110
This was approved previously https://reviews.llvm.org/D55115 (r348215)
but when committed it caused failures on the sanitizer buildbots when
building llvm with clang (containing this patch). This is now fixed
because I've added a check to see if getting the parent module returns
null if it does then set the alignment to 0.
Differential Revision: https://reviews.llvm.org/D55115
llvm-svn: 348571
DemandedBits and BDCE currently only support scalar integers. This
patch extends them to also handle vector integer operations. In this
case bits are not tracked for individual vector elements, instead a
bit is demanded if it is demanded for any of the elements. This matches
the behavior of computeKnownBits in ValueTracking and
SimplifyDemandedBits in InstCombine.
The getDemandedBits() method can now only be called on instructions that
have integer or vector of integer type. Previously it could be called on
any sized instruction (even if it was not particularly useful). The size
of the return value is now always the scalar size in bits (while
previously it was the type size in bits).
Differential Revision: https://reviews.llvm.org/D55297
llvm-svn: 348549
In some cases different alignments for function might be used to save
space e.g. thumb mode with -Oz will try to use 2 byte function
alignment. Similar patch that fixed this in other areas exists here
https://reviews.llvm.org/D46110
Differential Revision: https://reviews.llvm.org/D55115
llvm-svn: 348215
A loaded value with multiple users compared with 0 will become a load and
test single instruction. The load is not folded in this case (multiple
users), but the compare instruction is eliminated.
This patch returns 0 cost for the icmp in these cases.
Review: Ulrich Weigand
https://reviews.llvm.org/D55111
llvm-svn: 348141
Fix ScalarEvolution/solve-quadratic.ll test to account for __func__
output listing the complete function prototype rather than just its
name, as it does on NetBSD.
Example Linux output:
GetQuadraticEquation: addrec coeff bw: 4
GetQuadraticEquation: equation -2x^2 + -2x + -4, coeff bw: 5, multiplied by 2
Example NetBSD output:
llvm::Optional<std::tuple<llvm::APInt, llvm::APInt, llvm::APInt, llvm::APInt, unsigned int> > GetQuadraticEquation(const llvm::SCEVAddRecExpr*): addrec coeff bw: 4
llvm::Optional<std::tuple<llvm::APInt, llvm::APInt, llvm::APInt, llvm::APInt, unsigned int> > GetQuadraticEquation(const llvm::SCEVAddRecExpr*): equation -2x^2 + -2x + -4, coeff bw: 5, multiplied by 2
Differential Revision: https://reviews.llvm.org/D55162
llvm-svn: 348096
We were adding the entire scalarization extraction cost for reductions, which returns the total cost of extracting every element of a vector type.
For reductions we don't need to do this - we just need to extract the 0'th element after the reduction pattern has completed.
Fixes PR37731
Rebased and reapplied after being reverted in rL347541 due to PR39774 - which was fixed by D54955/rL347759 and D55017/rL347997
Differential Revision: https://reviews.llvm.org/D54585
llvm-svn: 348076
Summary:
This is patch #3 of the new DivergenceAnalysis
<https://lists.llvm.org/pipermail/llvm-dev/2018-May/123606.html>
The GPUDivergenceAnalysis is intended to eventually supersede the existing
LegacyDivergenceAnalysis. The existing LegacyDivergenceAnalysis produces
incorrect results on unstructured Control-Flow Graphs:
<https://bugs.llvm.org/show_bug.cgi?id=37185>
This patch adds the option -use-gpu-divergence-analysis to the
LegacyDivergenceAnalysis to turn it into a transparent wrapper for the
GPUDivergenceAnalysis.
Reviewers: nhaehnle
Reviewed By: nhaehnle
Subscribers: jholewinski, jvesely, jfb, llvm-commits, alex-t, sameerds, arsenm, nhaehnle
Differential Revision: https://reviews.llvm.org/D53493
llvm-svn: 348048
Three minor changes to these extra costs:
* For ICmp instructions, instead of adding 2 all the time for extending each
operand, this is only done if that operand is neither a load or an
immediate.
* The operands extension costs for divides removed, because we now use a high
cost already for the divide (20).
* The costs for lhsr/ashr extra costs removed as this did not seem useful.
Review: Ulrich Weigand
https://reviews.llvm.org/D55053
llvm-svn: 347961
Unlike most cost model functions this code makes a lot of table lookups without using the results from getTypeLegalizationCost. This means 512-bit vectors can be looked up even when the type isn't legal.
This patch adds a check around the two tables that contain 512-bit types to make sure that neither of the types would be split by type legalization. Meaning 512 bit types are illegal. I wanted to write this in a somewhat generic way that uses type legalization query hooks. But if prefered, I can switch to just using is512BitVector and the subtarget feature.
Differential Revision: https://reviews.llvm.org/D54984
llvm-svn: 347786
This fixes some of scalarization costs reported for sext/zext using avx512bw. This does not fix all scalarization costs being reported. Just the worst.
I've restricted this only to combinations of types that are legal with avx512bw like v32i1/v64i1/v32i16/v64i8 and conversions between vXi1 and vXi8/vXi16 with legal vXi8/vXi16 result types.
Differential Revision: https://reviews.llvm.org/D54979
llvm-svn: 347785
Expansion of SIGN_EXTEND_INREG can create a VSRAI instruction. If there is already a VSRAI after it, we should combine them into a larger VSRAI
Differential Revision: https://reviews.llvm.org/D54959
llvm-svn: 347784
CGF/CLGF compares an i64 register with a sign/zero extended loaded i32 value
in memory.
This patch makes such a load considered foldable and so gets a 0 cost.
Review: Ulrich Weigand
https://reviews.llvm.org/D54944
llvm-svn: 347735
AH, SH and MH costs are already covered in the cases where LHS is 32 bits and
RHS is 16 bits of memory sign-extended to i32.
As these instructions are also used when LHS is i16, this patch recognizes
that the loads will get folded then as well.
Review: Ulrich Weigand
https://reviews.llvm.org/D54940
llvm-svn: 347734
Single instructions exist for i8 and i16 comparisons of memory against a
small immediate.
This patch makes sure that if the load in these cases has a single user (the
ICmp), it gets a 0 cost (folded), and also that the ICmp gets a cost of 1.
Review: Ulrich Weigand
https://reviews.llvm.org/D54897
llvm-svn: 347733
Since byte-swapping loads and stores are supported, a 'load -> bswap' or
'bswap -> store' sequence should have the cost of one.
Review: Ulrich Weigand
https://reviews.llvm.org/D54870
llvm-svn: 347732
The check lines marked AVX256 in the zext256/sext256 functions should be closer to the AVX values which would take into account a splitting cost.
llvm-svn: 347722
Our sext/zext cost modeling was somewhat incomplete. And had no coverage for the fact that avx512bw v32i16/v64i8 types return a scalarization cost.
Truncates are a whole different mess because isTruncateFree is returning true for vectors when it shouldn't and that's the fall back for anything not in the tables.
llvm-svn: 347719
Summary:
IPA is implemented as module pass which produce map from Function or Alias to
StackSafetyInfo for a single function.
From prototype by Evgenii Stepanov and Vlad Tsyrklevich.
Reviewers: eugenis, vlad.tsyrklevich, pcc, glider
Subscribers: hiraditya, mgrang, llvm-commits
Differential Revision: https://reviews.llvm.org/D54543
llvm-svn: 347611
Summary:
Analysis produces StackSafetyInfo which contains information with how allocas
and parameters were used in functions.
From prototype by Evgenii Stepanov and Vlad Tsyrklevich.
Reviewers: eugenis, vlad.tsyrklevich, pcc, glider
Subscribers: hiraditya, llvm-commits
Differential Revision: https://reviews.llvm.org/D54504
llvm-svn: 347603
Add support for funnel shifts to the DemandedBits analysis. The
demanded bits of the first two operands can be determined if the
shift amount is constant. The demanded bits of the third operand
(shift amount) can be determined if the bitwidth is a power of two.
This is basically the same functionality as implemented in D54869
and D54478, but for DemandedBits rather than InstCombine.
Differential Revision: https://reviews.llvm.org/D54876
llvm-svn: 347561
This reverts commit r346970.
It was causing PR39774, a crash in slp-vectorizer on a rather simple loop
with just a bunch of 'and's in the body.
llvm-svn: 347541
Implement getIntrinsicInstrCost() and return costs reflecting that bswap can
be done with a vperm per vector register.
Review: Ulrich Weigand
https://reviews.llvm.org/D54789
llvm-svn: 347445
LVI was symbolically executing binary operators only when the RHS was
constant, missing the case where we have a ConstantRange for the RHS,
but not an actual constant. Tested using check-all and by
bootstrapping. Compile time is not impacted measurably.
Differential Revision: https://reviews.llvm.org/D19859
llvm-svn: 347379
Support saturating add/sub in constant folding, based on the APInt methods introduced in D54332.
Patch by: @nikic (Nikita Popov)
Differential Revision: https://reviews.llvm.org/D54531
llvm-svn: 347328
Summary:
Currently, when vectorizing stores to uniform addresses, the only
instance we prevent vectorization is if there are multiple stores to the
same uniform address causing an unsafe dependency.
This patch teaches LAA to avoid vectorizing loops that have an unsafe
cross-iteration dependency between a load and a store to the same uniform address.
Fixes PR39653.
Reviewers: Ayal, efriedma
Subscribers: rkruppe, llvm-commits
Differential Revision: https://reviews.llvm.org/D54538
llvm-svn: 347220
We were adding the entire scalarization extraction cost for reductions, which returns the total cost of extracting every element of a vector type.
For reductions we don't need to do this - we just need to extract the 0'th element after the reduction pattern has completed.
Fixes PR37731
Differential Revision: https://reviews.llvm.org/D54585
llvm-svn: 346970
Add support for the expansion of funnelshift/rotates to getIntrinsicInstrCost.
This also required us to move the X86 fshl/fshr costs to the same place as the rotates to avoid expansion and get correct scalarization vs vectorization costs.
llvm-svn: 346854
When we repeat the 2 shifting operands then this is a bit rotation - annoyingly this has to be done in the other getIntrinsicInstrCost than most intrinsics as we need to check the operands are the same.
llvm-svn: 346688
Improve getCastInstrCost() by respecting the different types of Src and Dst
for vector integer <-> fp conversions.
This means that extracting from integer becomes more expensive (by the
extraction penalty), and the extraction from fp becomes cheaper (no longer
has a false extraction penalty).
Review: Ulrich Weigand
https://reviews.llvm.org/D54423
llvm-svn: 346663
Instead of defaulting to a cost = 1, expand to element extract/insert like we do for other shuffles.
This exposes an issue in LoopVectorize which could call SK_ExtractSubvector with a scalar subvector type.
llvm-svn: 346656
The patch has been reverted because it ended up prohibiting propagation
of a constant to exit value. For such values, we should skip all checks
related to hard uses because propagating a constant is always profitable.
Differential Revision: https://reviews.llvm.org/D53691
llvm-svn: 346397
This reverts commit 2f425e9c7946b9d74e64ebbfa33c1caa36914402.
It seems that the check that we still should do the transform if we
know the result is constant is missing in this code. So the logic that
has been deleted by this change is still sometimes accidentally useful.
I revert the change to see what can be done about it. The motivating
case is the following:
@Y = global [400 x i16] zeroinitializer, align 1
define i16 @foo() {
entry:
br label %for.body
for.body: ; preds = %entry, %for.body
%i = phi i16 [ 0, %entry ], [ %inc, %for.body ]
%arrayidx = getelementptr inbounds [400 x i16], [400 x i16]* @Y, i16 0, i16 %i
store i16 0, i16* %arrayidx, align 1
%inc = add nuw nsw i16 %i, 1
%cmp = icmp ult i16 %inc, 400
br i1 %cmp, label %for.body, label %for.end
for.end: ; preds = %for.body
%inc.lcssa = phi i16 [ %inc, %for.body ]
ret i16 %inc.lcssa
}
We should be able to figure out that the result is constant, but the patch
breaks it.
Differential Revision: https://reviews.llvm.org/D51584
llvm-svn: 346198
Let i8/i16 uint/sint to fp conversions cost 1 if operand is a load.
Since the load already does the extension, there is no extra cost (previously
returned 2).
Review: Ulrich Weigand
https://reviews.llvm.org/D54028
llvm-svn: 346009
Summary:
The hot and cold count thresholds are derived from the summary, but for
debugging purposes it is convenient to provide the actual thresholds.
Reviewers: davidxl
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D54040
llvm-svn: 346005
Scalar i1 to fp conversions are done with a branch sequence, so it should
have a higher cost.
Review: Ulrich Weigand
https://reviews.llvm.org/D53924
llvm-svn: 345818
This factors out a new method getBoolVecToIntConversionCost() containing the
code for vector sext/zext of i1, in order to reuse it for i1 to double vector
conversions.
Review: Ulrich Weigand
https://reviews.llvm.org/D53923
llvm-svn: 345817
When rewriting loop exit values, IndVars considers this transform not profitable if
the loop instruction has a loop user which it believes cannot be optimized away.
In current implementation only calls that immediately use the instruction are considered
as such.
This patch extends the definition of "hard" users to any side-effecting instructions
(which usually cannot be optimized away from the loop) and also allows handling
of not just immediate users, but use chains.
Differentlai Revision: https://reviews.llvm.org/D51584
Reviewed By: etherzhhb
llvm-svn: 345814
When we calculate a product of 2 AddRecs, we end up making quite massive
computations to deduce the operands of resulting AddRec. This process can
be optimized by computing all args of intermediate sum and then calling
`getAddExpr` once rather than calling `getAddExpr` with intermediate
result every time a new argument is computed.
Differential Revision: https://reviews.llvm.org/D53189
Reviewed By: rtereshin
llvm-svn: 345813
Correct costings of SK_ExtractSubvector requires the SubTy argument to indicate the type/size of the extracted subvector.
Unlike the rest of the shuffle kinds this means that the main Ty argument represents the source vector type not the destination!
I've done my best to fix a number of vectorizer uses:
SLP - the reduction epilogue costs should be using a SK_PermuteSingleSrc shuffle as these all occur at the hardware vector width - we're not extracting (illegal) subvector types. This is causing the cost model diffs as SK_ExtractSubvector costs are poorly handled and tend to just return 1 at the moment.
LV - I'm not clear on what the SK_ExtractSubvector should represents for recurrences - I've used a <1 x ?> subvector extraction as that seems to match the VF delta.
Differential Revision: https://reviews.llvm.org/D53573
llvm-svn: 345617
Sub, SDiv and UDiv are not commutative, so only the RHS operand can fold a
load. This patch adds a check for this.
Review: Ulrich Weigand
https://reviews.llvm.org/D53791
llvm-svn: 345596
The SystemZ backend can do arithmetic of memory by loading and then extending
one of the operands. Similarly, a load + truncate can be folded into an
operand.
This patch improves the SystemZ TTI cost function to recognize this.
Review: Ulrich Weigand
https://reviews.llvm.org/D52692
llvm-svn: 345327
Enable the DAG optimization that converts vector div/rem with constants into
multiply+shifts sequences by expanding them early. This is needed since
ISD::SMUL_LOHI is 'Custom' lowered on SystemZ, and will therefore not be
available to BuildSDIV after legalization.
Better cost values for these instructions based on how they will be
implemented (a constant divisor is cheaper).
Review: Ulrich Weigand
https://reviews.llvm.org/D53196
llvm-svn: 345321
Non-uniform division/remainder handling was added back at D49248/D50765 - so share the 'mul+sub' costs that already exist for uniform cases.
llvm-svn: 345164
Simon Pilgrim