Summary:
Existing heuristic uses the ratio between the function entry
frequency and the loop invocation frequency to find cold loops. However,
even if the loop executes frequently, if it has a small trip count per
each invocation, vectorization is not beneficial. On the other hand,
even if the loop invocation frequency is much smaller than the function
invocation frequency, if the trip count is high it is still beneficial
to vectorize the loop.
This patch uses estimated trip count computed from the profile metadata
as a primary metric to determine coldness of the loop. If the estimated
trip count cannot be computed, it falls back to the original heuristics.
Reviewers: Ayal, mssimpso, mkuper, danielcdh, wmi, tejohnson
Reviewed By: tejohnson
Subscribers: tejohnson, mzolotukhin, llvm-commits
Differential Revision: https://reviews.llvm.org/D32451
llvm-svn: 305729
The default behavior of -Rpass-analysis=loop-vectorizer is to report only the
first reason encountered for not vectorizing, if one is found, at which time the
vectorizer aborts its handling of the loop. This patch allows multiple reasons
for not vectorizing to be identified and reported, at the potential expense of
additional compile-time, under allowExtraAnalysis which can currently be turned
on by Clang's -fsave-optimization-record and opt's -pass-remarks-missed.
Removed from LoopVectorizationLegality::canVectorize() the redundant checking
and reporting if we CantComputeNumberOfIterations, as LAI::canAnalyzeLoop() also
does that. This redundancy is caught by a lit test once multiple reasons are
reported.
Patch initially developed by Dror Barak.
Differential Revision: https://reviews.llvm.org/D33396
llvm-svn: 303613
Fixes PR31789 - When loop-vectorize tries to use these intrinsics for a
non-default address space pointer we fail with a "Calling a function with a
bad singature!" assertion. This patch solves this by adding the 'vector of
pointers' argument as an overloaded type which will determine the address
space.
Differential revision: https://reviews.llvm.org/D31490
llvm-svn: 302018
This patch reapplies r298620. The original patch was reverted because of two
issues. First, the patch exposed a bug in InstCombine that caused the Chromium
builds to fail (PR32414). This issue was fixed in r299017. Second, the patch
introduced a bug in the vectorizer's scalars analysis that caused test suite
builds to fail on SystemZ. The scalars analysis was too aggressive and marked a
memory instruction scalar, even though it was going to be vectorized. This
issue has been fixed in the current patch and several new test cases for the
scalars analysis have been added.
llvm-svn: 299770
Reason: breaks linking Chromium with LLD + ThinLTO (a pass crashes)
LLVM bug: https://bugs.llvm.org//show_bug.cgi?id=32413
Original change description:
[LV] Vectorize GEPs
This patch adds support for vectorizing GEPs. Previously, we only generated
vector GEPs on-demand when creating gather or scatter operations. All GEPs from
the original loop were scalarized by default, and if a pointer was to be stored
to memory, we would have to build up the pointer vector with insertelement
instructions.
With this patch, we will vectorize all GEPs that haven't already been marked
for scalarization.
The patch refines collectLoopScalars to more exactly identify the scalar GEPs.
The function now more closely resembles collectLoopUniforms. And the patch
moves vector GEP creation out of vectorizeMemoryInstruction and into the main
vectorization loop. The vector GEPs needed for gather and scatter operations
will have already been generated before vectoring the memory accesses.
Original Differential Revision: https://reviews.llvm.org/D30710
llvm-svn: 298735
This patch adds support for vectorizing GEPs. Previously, we only generated
vector GEPs on-demand when creating gather or scatter operations. All GEPs from
the original loop were scalarized by default, and if a pointer was to be stored
to memory, we would have to build up the pointer vector with insertelement
instructions.
With this patch, we will vectorize all GEPs that haven't already been marked
for scalarization.
The patch refines collectLoopScalars to more exactly identify the scalar GEPs.
The function now more closely resembles collectLoopUniforms. And the patch
moves vector GEP creation out of vectorizeMemoryInstruction and into the main
vectorization loop. The vector GEPs needed for gather and scatter operations
will have already been generated before vectoring the memory accesses.
Differential Revision: https://reviews.llvm.org/D30710
llvm-svn: 298620
The practice in LV is that we emit analysis remarks and then finally report
either a missed or applied remark on the final decision whether vectorization
is taking place. On this code path, we were closing with an analysis remark.
llvm-svn: 296578
Making the cost model selecting between Interleave, GatherScatter or Scalar vectorization form of memory instruction.
The right decision should be done for non-consecutive memory access instrcuctions that may have more than one vectorization solution.
This patch includes the following changes:
- Cost Model calculates the cost of Load/Store vector form and choose the better option between Widening, Interleave, GatherScactter and Scalarization. Cost Model keeps the widening decision.
- Arrays of Uniform and Scalar values are moved from Legality to Cost Model.
- Cost Model collects Uniforms and Scalars per VF. The collection is based on CM decision map of Loadis/Stores vectorization form.
- Vectorization of memory instruction is performed according to the CM decision.
Differential Revision: https://reviews.llvm.org/D27919
llvm-svn: 294503
If a memory instruction will be vectorized, but it's pointer operand is
non-consecutive-like, the instruction is a gather or scatter operation. Its
pointer operand will be non-uniform. This should fix PR31671.
Reference: https://llvm.org/bugs/show_bug.cgi?id=31671
Differential Revision: https://reviews.llvm.org/D28819
llvm-svn: 292254
updated instructions:
pmulld, pmullw, pmulhw, mulsd, mulps, mulpd, divss, divps, divsd, divpd, addpd and subpd.
special optimization case which replaces pmulld with pmullw\pmulhw\pshuf seq.
In case if the real operands bitwidth <= 16.
Differential Revision: https://reviews.llvm.org/D28104
llvm-svn: 291657
This code seems to be target dependent which may not be the same for all targets.
Passed the decision whether the given stride is complex or not to the target by sending stride information via SCEV to getAddressComputationCost instead of 'IsComplex'.
Specifically at X86 targets we dont see any significant address computation cost in case of the strided access in general.
Differential Revision: https://reviews.llvm.org/D27518
llvm-svn: 291106
This patch attempts to scalarize the operand expressions of predicated
instructions if they were conditionally executed in the original loop. After
scalarization, the expressions will be sunk inside the blocks created for the
predicated instructions. The transformation essentially performs
un-if-conversion on the operands.
The cost model has been updated to determine if scalarization is profitable. It
compares the cost of a vectorized instruction, assuming it will be
if-converted, to the cost of the scalarized instruction, assuming that the
instructions corresponding to each vector lane will be sunk inside a predicated
block, possibly avoiding execution. If it's more profitable to scalarize the
entire expression tree feeding the predicated instruction, the expression will
be scalarized; otherwise, it will be vectorized. We only consider the cost of
the entire expression to accurately estimate the cost of the required
insertelement and extractelement instructions.
Differential Revision: https://reviews.llvm.org/D26083
llvm-svn: 288909
The register usage algorithm incorrectly treats instructions whose value is
not used within the loop (e.g. those that do not produce a value).
The algorithm first calculates the usages within the loop. It iterates over
the instructions in order, and records at which instruction index each use
ends (in fact, they're actually recorded against the next index, as this is
when we want to delete them from the open intervals).
The algorithm then iterates over the instructions again, adding each
instruction in turn to a list of open intervals. Instructions are then
removed from the list of open intervals when they occur in the list of uses
ended at the current index.
The problem is, instructions which are not used in the loop are skipped.
However, although they aren't used, the last use of a value may have been
recorded against that instruction index. In this case, the use is not deleted
from the open intervals, which may then bump up the estimated register usage.
This patch fixes the issue by simply moving the "is used" check after the loop
which erases the uses at the current index.
Differential Revision: https://reviews.llvm.org/D26554
llvm-svn: 286969
Add explicit v16i16/v32i8 ADD/SUB costs, matching the costs of v4i64/v8i32 - they were missing for some reason.
This has side effects on the LV max bandwidth tests (AVX1 now prefers 128-bit vectors vs AVX2 which still prefers 256-bit)
llvm-svn: 286832
This is PR28376.
Unfortunately given the current structure of optimization diagnostics we
lack the capability to tell whether the user has
passed -Rpass-analysis=loop-vectorize since this is local to the
front-end (BackendConsumer::OptimizationRemarkHandler).
So rather than printing this even if the user has already
passed -Rpass-analysis, this patch just punts and stops recommending
this option. I don't think that getting this right is worth the
complexity.
Differential Revision: https://reviews.llvm.org/D26563
llvm-svn: 286662
possible pointer-wrap-around concerns, in some cases.
Before this patch, collectConstStridedAccesses (part of interleaved-accesses
analysis) called getPtrStride with [Assume=false, ShouldCheckWrap=true] when
examining all candidate pointers. This is too conservative. Instead, this
patch makes collectConstStridedAccesses use an optimistic approach, calling
getPtrStride with [Assume=true, ShouldCheckWrap=false], and then, once the
candidate interleave groups have been formed, revisits the pointer-wrapping
analysis but only where it matters: namely, in groups that have gaps, and where
the gaps are not at the very end of the group (in which case the loop is
peeled). This second time getPtrStride is called with [Assume=false,
ShouldCheckWrap=true], but this could further be improved to using Assume=true,
once we also add the logic to track that we are not going to meet the scev
runtime checks threshold.
Differential Revision: https://reviews.llvm.org/D25276
llvm-svn: 285517
After r279649 when getting a vector value from VectorLoopValueMap, we create an
insertelement sequence on-demand if the value has been scalarized instead of
vectorized. We previously inserted this insertelement sequence before the
value's first vector user. However, this insert location is problematic if that
user is the phi node of a first-order recurrence. With this patch, we move the
insertelement sequence after the last scalar instruction we created when
scalarizing the value. Thus, the value's vector definition in the new loop will
immediately follow its scalar definitions. This should fix PR30183.
Reference: https://llvm.org/bugs/show_bug.cgi?id=30183
llvm-svn: 280001
This patch unifies the data structures we use for mapping instructions from the
original loop to their corresponding instructions in the new loop. Previously,
we maintained two distinct maps for this purpose: WidenMap and ScalarIVMap.
WidenMap maintained the vector values each instruction from the old loop was
represented with, and ScalarIVMap maintained the scalar values each scalarized
induction variable was represented with. With this patch, all values created
for the new loop are maintained in VectorLoopValueMap.
The change allows for several simplifications. Previously, when an instruction
was scalarized, we had to insert the scalar values into vectors in order to
maintain the mapping in WidenMap. Then, if a user of the scalarized value was
also scalar, we had to extract the scalar values from the temporary vector we
created. We now aovid these unnecessary scalar-to-vector-to-scalar conversions.
If a scalarized value is used by a scalar instruction, the scalar value is used
directly. However, if the scalarized value is needed by a vector instruction,
we generate the needed insertelement instructions on-demand.
A common idiom in several locations in the code (including the scalarization
code), is to first get the vector values an instruction from the original loop
maps to, and then extract a particular scalar value. This patch adds
getScalarValue for this purpose along side getVectorValue as an interface into
VectorLoopValueMap. These functions work together to return the requested
values if they're available or to produce them if they're not.
The mapping has also be made less permissive. Entries can be added to
VectorLoopValue map with the new initVector and initScalar functions.
getVectorValue has been modified to return a constant reference to the mapped
entries.
There's no real functional change with this patch; however, in some cases we
will generate slightly different code. For example, instead of an insertelement
sequence following the definition of an instruction, it will now precede the
first use of that instruction. This can be seen in the test case changes.
Differential Revision: https://reviews.llvm.org/D23169
llvm-svn: 279649
Shifts with a uniform but non-constant count were considered very expensive to
vectorize, because the splat of the uniform count and the shift would tend to
appear in different blocks. That made the splat invisible to ISel, and we'd
scalarize the shift at codegen time.
Since r201655, CodeGenPrepare sinks those splats to be next to their use, and we
are able to select the appropriate vector shifts. This updates the cost model to
to take this into account by making shifts by a uniform cheap again.
Differential Revision: https://reviews.llvm.org/D23049
llvm-svn: 277782
Update comment for isOutOfScope and add a testcase for uniform value being used
out of scope.
Differential Revision: https://reviews.llvm.org/D23073
llvm-svn: 277515
This patch enables the vectorizer to generate both scalar and vector versions
of an integer induction variable for a given loop. Previously, we only
generated a scalar induction variable if we knew all its users were going to be
scalar. Otherwise, we generated a vector induction variable. In the case of a
loop with both scalar and vector users of the induction variable, we would
generate the vector induction variable and extract scalar values from it for
the scalar users. With this patch, we now generate both versions of the
induction variable when there are both scalar and vector users and select which
version to use based on whether the user is scalar or vector.
Differential Revision: https://reviews.llvm.org/D22869
llvm-svn: 277474
Allowed loop vectorization with secondary FP IVs. Like this:
float *A;
float x = init;
for (int i=0; i < N; ++i) {
A[i] = x;
x -= fp_inc;
}
The auto-vectorization is possible when the induction binary operator is "fast" or the function has "unsafe" attribute.
Differential Revision: https://reviews.llvm.org/D21330
llvm-svn: 276554
This patch moves the update instruction for vectorized integer induction phi
nodes to the end of the latch block. This ensures consistent placement of all
induction updates across all the kinds of int inductions we create (scalar,
splat vector, or vector phi).
Differential Revision: https://reviews.llvm.org/D22416
llvm-svn: 276339
For instructions in uniform set, they will not have vector versions so
add them to VecValuesToIgnore.
For induction vars, those only used in uniform instructions or consecutive
ptrs instructions have already been added to VecValuesToIgnore above. For
those induction vars which are only used in uniform instructions or
non-consecutive/non-gather scatter ptr instructions, the related phi and
update will also be added into VecValuesToIgnore set.
The change will make the vector RegUsages estimation less conservative.
Differential Revision: https://reviews.llvm.org/D20474
The recommit fixed the testcase global_alias.ll.
llvm-svn: 275936
For instructions in uniform set, they will not have vector versions so
add them to VecValuesToIgnore.
For induction vars, those only used in uniform instructions or consecutive
ptrs instructions have already been added to VecValuesToIgnore above. For
those induction vars which are only used in uniform instructions or
non-consecutive/non-gather scatter ptr instructions, the related phi and
update will also be added into VecValuesToIgnore set.
The change will make the vector RegUsages estimation less conservative.
Differential Revision: https://reviews.llvm.org/D20474
llvm-svn: 275912
Make some AVX and AVX512 cast costs more precise.
Based on part of a patch by Elena Demikhovsky (D15604).
Differential Revision: http://reviews.llvm.org/D22064
llvm-svn: 275106
The cost model should not assume vector casts get completely scalarized, since
on targets that have vector support, the common case is a partial split up to
the legal vector size. So, when a vector cast gets split, the resulting casts
end up legal and cheap.
Instead of pessimistically assuming scalarization, base TTI can use the costs
the concrete TTI provides for the split vector, plus a fudge factor to account
for the cost of the split itself. This fudge factor is currently 1 by default,
except on AMDGPU where inserts and extracts are considered free.
Differential Revision: http://reviews.llvm.org/D21251
llvm-svn: 274642
Except the seed uniform instructions (conditional branch and consecutive ptr
instructions), dependencies to be added into uniform set should only be used
by existing uniform instructions or intructions outside of current loop.
Differential Revision: http://reviews.llvm.org/D21755
llvm-svn: 274262
This is a resubmittion of 263158 change after fixing the existing problem with intrinsics mangling (see LTO and intrinsics mangling llvm-dev thread for details).
This patch fixes the problem which occurs when loop-vectorize tries to use @llvm.masked.load/store intrinsic for a non-default addrspace pointer. It fails with "Calling a function with a bad signature!" assertion in CallInst constructor because it tries to pass a non-default addrspace pointer to the pointer argument which has default addrspace.
The fix is to add pointer type as another overloaded type to @llvm.masked.load/store intrinsics.
Reviewed By: reames
Differential Revision: http://reviews.llvm.org/D17270
llvm-svn: 274043
This is a resubmittion of 263158 change after fixing the existing problem with intrinsics mangling (see LTO and intrinsics mangling llvm-dev thread for details).
This patch fixes the problem which occurs when loop-vectorize tries to use @llvm.masked.load/store intrinsic for a non-default addrspace pointer. It fails with "Calling a function with a bad signature!" assertion in CallInst constructor because it tries to pass a non-default addrspace pointer to the pointer argument which has default addrspace.
The fix is to add pointer type as another overloaded type to @llvm.masked.load/store intrinsics.
Reviewed By: reames
Differential Revision: http://reviews.llvm.org/D17270
llvm-svn: 273892
Previously, whenever we needed a vector IV, we would create it on the fly,
by splatting the scalar IV and adding a step vector. Instead, we can create a
real vector IV. This tends to save a couple of instructions per iteration.
This only changes the behavior for the most basic case - integer primary
IVs with a constant step.
Differential Revision: http://reviews.llvm.org/D20315
llvm-svn: 271410
Getting accurate locations for loops is important, because those locations are
used by the frontend to generate optimization remarks. Currently, optimization
remarks for loops often appear on the wrong line, often the first line of the
loop body instead of the loop itself. This is confusing because that line might
itself be another loop, or might be somewhere else completely if the body was
inlined function call. This happens because of the way we find the loop's
starting location. First, we look for a preheader, and if we find one, and its
terminator has a debug location, then we use that. Otherwise, we look for a
location on an instruction in the loop header.
The fallback heuristic is not bad, but will almost always find the beginning of
the body, and not the loop statement itself. The preheader location search
often fails because there's often not a preheader, and even when there is a
preheader, depending on how it was formed, it sometimes carries the location of
some preceeding code.
I don't see any good theoretical way to fix this problem. On the other hand,
this seems like a straightforward solution: Put the debug location in the
loop's llvm.loop metadata. A companion Clang patch will cause Clang to insert
llvm.loop metadata with appropriate locations when generating debugging
information. With these changes, our loop remarks have much more accurate
locations.
Differential Revision: http://reviews.llvm.org/D19738
llvm-svn: 270771
By making pointer extraction from a vector more expensive in the cost model,
we avoid the vectorization of a loop that is very likely to be memory-bound:
https://llvm.org/bugs/show_bug.cgi?id=27826
There are still bugs related to this, so we may need a more general solution
to avoid vectorizing obviously memory-bound loops when we don't have HW gather
support.
Differential Revision: http://reviews.llvm.org/D20601
llvm-svn: 270729
I really thought we were doing this already, but we were not. Given this input:
void Test(int *res, int *c, int *d, int *p) {
for (int i = 0; i < 16; i++)
res[i] = (p[i] == 0) ? res[i] : res[i] + d[i];
}
we did not vectorize the loop. Even with "assume_safety" the check that we
don't if-convert conditionally-executed loads (to protect against
data-dependent deferenceability) was not elided.
One subtlety: As implemented, it will still prefer to use a masked-load
instrinsic (given target support) over the speculated load. The choice here
seems architecture specific; the best option depends on how expensive the
masked load is compared to a regular load. Ideally, using the masked load still
reduces unnecessary memory traffic, and so should be preferred. If we'd rather
do it the other way, flipping the order of the checks is easy.
The LangRef is updated to make explicit that llvm.mem.parallel_loop_access also
implies that if conversion is okay.
Differential Revision: http://reviews.llvm.org/D19512
llvm-svn: 267514
Currently each Function points to a DISubprogram and DISubprogram has a
scope field. For member functions the scope is a DICompositeType. DIScopes
point to the DICompileUnit to facilitate type uniquing.
Distinct DISubprograms (with isDefinition: true) are not part of the type
hierarchy and cannot be uniqued. This change removes the subprograms
list from DICompileUnit and instead adds a pointer to the owning compile
unit to distinct DISubprograms. This would make it easy for ThinLTO to
strip unneeded DISubprograms and their transitively referenced debug info.
Motivation
----------
Materializing DISubprograms is currently the most expensive operation when
doing a ThinLTO build of clang.
We want the DISubprogram to be stored in a separate Bitcode block (or the
same block as the function body) so we can avoid having to expensively
deserialize all DISubprograms together with the global metadata. If a
function has been inlined into another subprogram we need to store a
reference the block containing the inlined subprogram.
Attached to https://llvm.org/bugs/show_bug.cgi?id=27284 is a python script
that updates LLVM IR testcases to the new format.
http://reviews.llvm.org/D19034
<rdar://problem/25256815>
llvm-svn: 266446
This is a resubmittion of 263158 change.
This patch fixes the problem which occurs when loop-vectorize tries to use @llvm.masked.load/store intrinsic for a non-default addrspace pointer. It fails with "Calling a function with a bad signature!" assertion in CallInst constructor because it tries to pass a non-default addrspace pointer to the pointer argument which has default addrspace.
The fix is to add pointer type as another overloaded type to @llvm.masked.load/store intrinsics.
Reviewed By: reames
Differential Revision: http://reviews.llvm.org/D17270
llvm-svn: 266086
Vectorization cost of uniform load wasn't correctly calculated.
As a result, a simple loop that loads a uniform value wasn't vectorized.
Differential Revision: http://reviews.llvm.org/D18940
llvm-svn: 265901
InstCombine cannot effectively remove redundant assumptions without them
registered in the assumption cache. The vectorizer can create identical
assumptions but doesn't register them with the cache, resulting in
slower compile times because InstCombine tries to reason about a lot
more assumptions.
Fix this by registering the cloned assumptions.
llvm-svn: 265800
This re-commits r265535 which was reverted in r265541 because it
broke the windows bots. The problem was that we had a PointerIntPair
which took a pointer to a struct allocated with new. The problem
was that new doesn't provide sufficient alignment guarantees.
This pattern was already present before r265535 and it just happened
to work. To fix this, we now separate the PointerToIntPair from the
ExitNotTakenInfo struct into a pointer and a bool.
Original commit message:
Summary:
When the backedge taken codition is computed from an icmp, SCEV can
deduce the backedge taken count only if one of the sides of the icmp
is an AddRecExpr. However, due to sign/zero extensions, we sometimes
end up with something that is not an AddRecExpr.
However, we can use SCEV predicates to produce a 'guarded' expression.
This change adds a method to SCEV to get this expression, and the
SCEV predicate associated with it.
In HowManyGreaterThans and HowManyLessThans we will now add a SCEV
predicate associated with the guarded backedge taken count when the
analyzed SCEV expression is not an AddRecExpr. Note that we only do
this as an alternative to returning a 'CouldNotCompute'.
We use new feature in Loop Access Analysis and LoopVectorize to analyze
and transform more loops.
Reviewers: anemet, mzolotukhin, hfinkel, sanjoy
Subscribers: flyingforyou, mcrosier, atrick, mssimpso, sanjoy, mzolotukhin, llvm-commits
Differential Revision: http://reviews.llvm.org/D17201
llvm-svn: 265786
Summary:
When the backedge taken codition is computed from an icmp, SCEV can
deduce the backedge taken count only if one of the sides of the icmp
is an AddRecExpr. However, due to sign/zero extensions, we sometimes
end up with something that is not an AddRecExpr.
However, we can use SCEV predicates to produce a 'guarded' expression.
This change adds a method to SCEV to get this expression, and the
SCEV predicate associated with it.
In HowManyGreaterThans and HowManyLessThans we will now add a SCEV
predicate associated with the guarded backedge taken count when the
analyzed SCEV expression is not an AddRecExpr. Note that we only do
this as an alternative to returning a 'CouldNotCompute'.
We use new feature in Loop Access Analysis and LoopVectorize to analyze
and transform more loops.
Reviewers: anemet, mzolotukhin, hfinkel, sanjoy
Subscribers: flyingforyou, mcrosier, atrick, mssimpso, sanjoy, mzolotukhin, llvm-commits
Differential Revision: http://reviews.llvm.org/D17201
llvm-svn: 265535
To quote the langref "Unlike sqrt in libm, however, llvm.sqrt has
undefined behavior for negative numbers other than -0.0 (which allows
for better optimization, because there is no need to worry about errno
being set). llvm.sqrt(-0.0) is defined to return -0.0 like IEEE sqrt."
This means that it's unsafe to replace sqrt with llvm.sqrt unless the
call is annotated with nnan.
Thanks to Hal Finkel for pointing this out!
llvm-svn: 265521
This change prevents the loop vectorizer from vectorizing when all of the vector
types it generates will be scalarized. I've run into this problem on the PPC's QPX
vector ISA, which only holds floating-point vector types. The loop vectorizer
will, however, happily vectorize loops with purely integer computation. Here's
an example:
LV: The Smallest and Widest types: 32 / 32 bits.
LV: The Widest register is: 256 bits.
LV: Found an estimated cost of 0 for VF 1 For instruction: %indvars.iv25 = phi i64 [ 0, %entry ], [ %indvars.iv.next26, %for.body ]
LV: Found an estimated cost of 0 for VF 1 For instruction: %arrayidx = getelementptr inbounds [1600 x i32], [1600 x i32]* %a, i64 0, i64 %indvars.iv25
LV: Found an estimated cost of 0 for VF 1 For instruction: %2 = trunc i64 %indvars.iv25 to i32
LV: Found an estimated cost of 1 for VF 1 For instruction: store i32 %2, i32* %arrayidx, align 4
LV: Found an estimated cost of 1 for VF 1 For instruction: %indvars.iv.next26 = add nuw nsw i64 %indvars.iv25, 1
LV: Found an estimated cost of 1 for VF 1 For instruction: %exitcond27 = icmp eq i64 %indvars.iv.next26, 1600
LV: Found an estimated cost of 0 for VF 1 For instruction: br i1 %exitcond27, label %for.cond.cleanup, label %for.body
LV: Scalar loop costs: 3.
LV: Found an estimated cost of 0 for VF 2 For instruction: %indvars.iv25 = phi i64 [ 0, %entry ], [ %indvars.iv.next26, %for.body ]
LV: Found an estimated cost of 0 for VF 2 For instruction: %arrayidx = getelementptr inbounds [1600 x i32], [1600 x i32]* %a, i64 0, i64 %indvars.iv25
LV: Found an estimated cost of 0 for VF 2 For instruction: %2 = trunc i64 %indvars.iv25 to i32
LV: Found an estimated cost of 2 for VF 2 For instruction: store i32 %2, i32* %arrayidx, align 4
LV: Found an estimated cost of 1 for VF 2 For instruction: %indvars.iv.next26 = add nuw nsw i64 %indvars.iv25, 1
LV: Found an estimated cost of 1 for VF 2 For instruction: %exitcond27 = icmp eq i64 %indvars.iv.next26, 1600
LV: Found an estimated cost of 0 for VF 2 For instruction: br i1 %exitcond27, label %for.cond.cleanup, label %for.body
LV: Vector loop of width 2 costs: 2.
LV: Found an estimated cost of 0 for VF 4 For instruction: %indvars.iv25 = phi i64 [ 0, %entry ], [ %indvars.iv.next26, %for.body ]
LV: Found an estimated cost of 0 for VF 4 For instruction: %arrayidx = getelementptr inbounds [1600 x i32], [1600 x i32]* %a, i64 0, i64 %indvars.iv25
LV: Found an estimated cost of 0 for VF 4 For instruction: %2 = trunc i64 %indvars.iv25 to i32
LV: Found an estimated cost of 4 for VF 4 For instruction: store i32 %2, i32* %arrayidx, align 4
LV: Found an estimated cost of 1 for VF 4 For instruction: %indvars.iv.next26 = add nuw nsw i64 %indvars.iv25, 1
LV: Found an estimated cost of 1 for VF 4 For instruction: %exitcond27 = icmp eq i64 %indvars.iv.next26, 1600
LV: Found an estimated cost of 0 for VF 4 For instruction: br i1 %exitcond27, label %for.cond.cleanup, label %for.body
LV: Vector loop of width 4 costs: 1.
...
LV: Selecting VF: 8.
LV: The target has 32 registers
LV(REG): Calculating max register usage:
LV(REG): At #0 Interval # 0
LV(REG): At #1 Interval # 1
LV(REG): At #2 Interval # 2
LV(REG): At #4 Interval # 1
LV(REG): At #5 Interval # 1
LV(REG): VF = 8
The problem is that the cost model here is not wrong, exactly. Since all of
these operations are scalarized, their cost (aside from the uniform ones) are
indeed VF*(scalar cost), just as the model suggests. In fact, the larger the VF
picked, the lower the relative overhead from the loop itself (and the
induction-variable update and check), and so in a sense, picking the largest VF
here is the right thing to do.
The problem is that vectorizing like this, where all of the vectors will be
scalarized in the backend, isn't really vectorizing, but rather interleaving.
By itself, this would be okay, but then the vectorizer itself also interleaves,
and that's where the problem manifests itself. There's aren't actually enough
scalar registers to support the normal interleave factor multiplied by a factor
of VF (8 in this example). In other words, the problem with this is that our
register-pressure heuristic does not account for scalarization.
While we might want to improve our register-pressure heuristic, I don't think
this is the right motivating case for that work. Here we have a more-basic
problem: The job of the vectorizer is to vectorize things (interleaving aside),
and if the IR it generates won't generate any actual vector code, then
something is wrong. Thus, if every type looks like it will be scalarized (i.e.
will be split into VF or more parts), then don't consider that VF.
This is not a problem specific to PPC/QPX, however. The problem comes up under
SSE on x86 too, and as such, this change fixes PR26837 too. I've added Sanjay's
reduced test case from PR26837 to this commit.
Differential Revision: http://reviews.llvm.org/D18537
llvm-svn: 264904
This patch fixes the problem which occurs when loop-vectorize tries to use @llvm.masked.load/store intrinsic for a non-default addrspace pointer. It fails with "Calling a function with a bad signature!" assertion in CallInst constructor because it tries to pass a non-default addrspace pointer to the pointer argument which has default addrspace.
The fix is to add pointer type as another overloaded type to @llvm.masked.load/store intrinsics.
Reviewed By: reames
Differential Revision: http://reviews.llvm.org/D17270
llvm-svn: 263158
The irony of this patch is that one CPU that is affected is AMD Jaguar, and Jaguar
has a completely double-pumped AVX implementation. But getting the cost model to
reflect that is a much bigger problem. The small goal here is simply to improve on
the lie that !AVX2 == SandyBridge.
Differential Revision: http://reviews.llvm.org/D18000
llvm-svn: 263069
Loop vectorizer now knows to vectorize GEP and create masked gather and scatter intrinsics for random memory access.
The feature is enabled on AVX-512 target.
Differential Revision: http://reviews.llvm.org/D15690
llvm-svn: 261140
Use AVX1 FP instructions (vmaskmovps/pd) in place of the AVX2 int instructions (vpmaskmovd/q).
Differential Revision: http://reviews.llvm.org/D16528
llvm-svn: 258675
(This is the third attempt to check in this patch, and the first two are r255454
and r255460. The once failed test file reg-usage.ll is now moved to
test/Transform/LoopVectorize/X86 directory with target datalayout and target
triple indicated.)
LoopVectorizationCostModel::calculateRegisterUsage() is used to estimate the
register usage for specific VFs. However, it takes into account many
instructions that won't be vectorized, such as induction variables,
GetElementPtr instruction, etc.. This makes the loop vectorizer too conservative
when choosing VF. In this patch, the induction variables that won't be
vectorized plus GetElementPtr instruction will be added to ValuesToIgnore set
so that their register usage won't be considered any more.
Differential revision: http://reviews.llvm.org/D15177
llvm-svn: 255691
(This is the second attempt to check in this patch: REQUIRES: asserts is added
to reg-usage.ll now.)
LoopVectorizationCostModel::calculateRegisterUsage() is used to estimate the
register usage for specific VFs. However, it takes into account many
instructions that won't be vectorized, such as induction variables,
GetElementPtr instruction, etc.. This makes the loop vectorizer too conservative
when choosing VF. In this patch, the induction variables that won't be
vectorized plus GetElementPtr instruction will be added to ValuesToIgnore set
so that their register usage won't be considered any more.
Differential revision: http://reviews.llvm.org/D15177
llvm-svn: 255460
LoopVectorizationCostModel::calculateRegisterUsage() is used to estimate the
register usage for specific VFs. However, it takes into account many
instructions that won't be vectorized, such as induction variables,
GetElementPtr instruction, etc.. This makes the loop vectorizer too conservative
when choosing VF. In this patch, the induction variables that won't be
vectorized plus GetElementPtr instruction will be added to ValuesToIgnore set
so that their register usage won't be considered any more.
Differential revision: http://reviews.llvm.org/D15177
llvm-svn: 255454
The masked intrinsics support all integer and floating point data types. I added the pointer type to this list.
Added tests for CodeGen and for Loop Vectorizer.
Updated the Language Reference.
Differential Revision: http://reviews.llvm.org/D14150
llvm-svn: 253544
Previously, subprograms contained a metadata reference to the function they
described. Because most clients need to get or set a subprogram for a given
function rather than the other way around, this created unneeded inefficiency.
For example, many passes needed to call the function llvm::makeSubprogramMap()
to build a mapping from functions to subprograms, and the IR linker needed to
fix up function references in a way that caused quadratic complexity in the IR
linking phase of LTO.
This change reverses the direction of the edge by storing the subprogram as
function-level metadata and removing DISubprogram's function field.
Since this is an IR change, a bitcode upgrade has been provided.
Fixes PR23367. An upgrade script for textual IR for out-of-tree clients is
attached to the PR.
Differential Revision: http://reviews.llvm.org/D14265
llvm-svn: 252219
To be able to maximize the bandwidth during vectorization, this patch provides a new flag vectorizer-maximize-bandwidth. When it is turned on, the vectorizer will determine the vectorization factor (VF) using the smallest instead of widest type in the loop. To avoid increasing register pressure too much, estimates of the register usage for different VFs are calculated so that we only choose a VF when its register usage doesn't exceed the number of available registers.
This is the second attempt to submit this patch. The first attempt got a test failure on ARM. This patch is updated to try to fix the failure (more specifically, by handling the case when VF=1).
Differential revision: http://reviews.llvm.org/D8943
llvm-svn: 251850
To be able to maximize the bandwidth during vectorization, this patch provides a new flag vectorizer-maximize-bandwidth. When it is turned on, the vectorizer will determine the vectorization factor (VF) using the smallest instead of widest type in the loop. To avoid increasing register pressure too much, estimates of the register usage for different VFs are calculated so that we only choose a VF when its register usage doesn't exceed the number of available registers.
llvm-svn: 251592
Vectorization of memory instruction (Load/Store) is possible when the pointer is coming from GEP. The GEP analysis allows to estimate the profit.
In some cases we have a "bitcast" between GEP and memory instruction.
I added code that skips the "bitcast".
http://reviews.llvm.org/D13886
llvm-svn: 251291
with the new pass manager, and no longer relying on analysis groups.
This builds essentially a ground-up new AA infrastructure stack for
LLVM. The core ideas are the same that are used throughout the new pass
manager: type erased polymorphism and direct composition. The design is
as follows:
- FunctionAAResults is a type-erasing alias analysis results aggregation
interface to walk a single query across a range of results from
different alias analyses. Currently this is function-specific as we
always assume that aliasing queries are *within* a function.
- AAResultBase is a CRTP utility providing stub implementations of
various parts of the alias analysis result concept, notably in several
cases in terms of other more general parts of the interface. This can
be used to implement only a narrow part of the interface rather than
the entire interface. This isn't really ideal, this logic should be
hoisted into FunctionAAResults as currently it will cause
a significant amount of redundant work, but it faithfully models the
behavior of the prior infrastructure.
- All the alias analysis passes are ported to be wrapper passes for the
legacy PM and new-style analysis passes for the new PM with a shared
result object. In some cases (most notably CFL), this is an extremely
naive approach that we should revisit when we can specialize for the
new pass manager.
- BasicAA has been restructured to reflect that it is much more
fundamentally a function analysis because it uses dominator trees and
loop info that need to be constructed for each function.
All of the references to getting alias analysis results have been
updated to use the new aggregation interface. All the preservation and
other pass management code has been updated accordingly.
The way the FunctionAAResultsWrapperPass works is to detect the
available alias analyses when run, and add them to the results object.
This means that we should be able to continue to respect when various
passes are added to the pipeline, for example adding CFL or adding TBAA
passes should just cause their results to be available and to get folded
into this. The exception to this rule is BasicAA which really needs to
be a function pass due to using dominator trees and loop info. As
a consequence, the FunctionAAResultsWrapperPass directly depends on
BasicAA and always includes it in the aggregation.
This has significant implications for preserving analyses. Generally,
most passes shouldn't bother preserving FunctionAAResultsWrapperPass
because rebuilding the results just updates the set of known AA passes.
The exception to this rule are LoopPass instances which need to preserve
all the function analyses that the loop pass manager will end up
needing. This means preserving both BasicAAWrapperPass and the
aggregating FunctionAAResultsWrapperPass.
Now, when preserving an alias analysis, you do so by directly preserving
that analysis. This is only necessary for non-immutable-pass-provided
alias analyses though, and there are only three of interest: BasicAA,
GlobalsAA (formerly GlobalsModRef), and SCEVAA. Usually BasicAA is
preserved when needed because it (like DominatorTree and LoopInfo) is
marked as a CFG-only pass. I've expanded GlobalsAA into the preserved
set everywhere we previously were preserving all of AliasAnalysis, and
I've added SCEVAA in the intersection of that with where we preserve
SCEV itself.
One significant challenge to all of this is that the CGSCC passes were
actually using the alias analysis implementations by taking advantage of
a pretty amazing set of loop holes in the old pass manager's analysis
management code which allowed analysis groups to slide through in many
cases. Moving away from analysis groups makes this problem much more
obvious. To fix it, I've leveraged the flexibility the design of the new
PM components provides to just directly construct the relevant alias
analyses for the relevant functions in the IPO passes that need them.
This is a bit hacky, but should go away with the new pass manager, and
is already in many ways cleaner than the prior state.
Another significant challenge is that various facilities of the old
alias analysis infrastructure just don't fit any more. The most
significant of these is the alias analysis 'counter' pass. That pass
relied on the ability to snoop on AA queries at different points in the
analysis group chain. Instead, I'm planning to build printing
functionality directly into the aggregation layer. I've not included
that in this patch merely to keep it smaller.
Note that all of this needs a nearly complete rewrite of the AA
documentation. I'm planning to do that, but I'd like to make sure the
new design settles, and to flesh out a bit more of what it looks like in
the new pass manager first.
Differential Revision: http://reviews.llvm.org/D12080
llvm-svn: 247167
As a follow-up to r246098, require `DISubprogram` definitions
(`isDefinition: true`) to be 'distinct'. Specifically, add an assembler
check, a verifier check, and bitcode upgrading logic to combat testcase
bitrot after the `DIBuilder` change.
While working on the testcases, I realized that
test/Linker/subprogram-linkonce-weak-odr.ll isn't relevant anymore. Its
purpose was to check for a corner case in PR22792 where two subprogram
definitions match exactly and share the same metadata node. The new
verifier check, requiring that subprogram definitions are 'distinct',
precludes that possibility.
I updated almost all the IR with the following script:
git grep -l -E -e '= !DISubprogram\(.* isDefinition: true' |
grep -v test/Bitcode |
xargs sed -i '' -e 's/= \(!DISubprogram(.*, isDefinition: true\)/= distinct \1/'
Likely some variant of would work for out-of-tree testcases.
llvm-svn: 246327
This patch changes the analysis diagnostics produced when loops with
floating-point recurrences or memory operations are identified. The new messages
say "cannot prove it is safe to reorder * operations; allow reordering by
specifying #pragma clang loop vectorize(enable)". Depending on the type of
diagnostic the message will include additional options such as ffast-math or
__restrict__.
This patch also allows the vectorize(enable) pragma to override the low pointer
memory check threshold. When the hint is given a higher threshold is used.
See the clang patch for the options produced for each diagnostic.
llvm-svn: 246187
Sometimes interleaving is not beneficial, as determined by the cost-model and sometimes it is disabled by a loop hint (by the user). This patch modifies the diagnostic messages to make it clear why interleaving wasn't done.
llvm-svn: 244485
Taewook Oh