BasicAA contains knowledge of certain intrinsics, such as memcpy and memset,
and uses that information to form more-accurate answers to CallSite vs. Loc
ModRef queries. Unfortunately, it did not use this information when answering
CallSite vs. CallSite queries.
Generically, when an intrinsic takes one or more pointers and the intrinsic is
marked only to read/write from its arguments, the offset/size is unknown. As a
result, the generic code that answers CallSite vs. CallSite (and CallSite vs.
Loc) queries in AA uses UnknownSize when forming Locs from an intrinsic's
arguments. While BasicAA's CallSite vs. Loc override could use more-accurate
size information for some intrinsics, it did not do the same for CallSite vs.
CallSite queries.
This change refactors the intrinsic-specific logic in BasicAA into a generic AA
query function: getArgLocation, which is overridden by BasicAA to supply the
intrinsic-specific knowledge, and used by AA's generic implementation. This
allows the intrinsic-specific knowledge to be used by both CallSite vs. Loc and
CallSite vs. CallSite queries, and simplifies the BasicAA implementation.
Currently, only one function, Mac's memset_pattern16, is handled by BasicAA
(all the rest are intrinsics). As a side-effect of this refactoring, BasicAA's
getModRefBehavior override now also returns OnlyAccessesArgumentPointees for
this function (which is an improvement).
llvm-svn: 212572
This patch:
1) Improves the cost model for x86 alternate shuffles (originally
added at revision 211339);
2) Teaches the Cost Model Analysis pass how to analyze alternate shuffles.
Alternate shuffles are a special kind of blend; on x86, we can often
easily lowered alternate shuffled into single blend
instruction (depending on the subtarget features).
The existing cost model didn't take into account subtarget features.
Also, it had a couple of "dead" entries for vector types that are never
legal (example: on x86 types v2i32 and v2f32 are not legal; those are
always either promoted or widened to 128-bit vector types).
The new x86 cost model takes into account what target features we have
before returning the shuffle cost (i.e. the number of instructions
after the blend is lowered/expanded).
This patch also teaches the Cost Model Analysis how to identify and analyze
alternate shuffles (i.e. 'SK_Alternate' shufflevector instructions):
- added function 'isAlternateVectorMask';
- added some logic to check if an instruction is a alternate shuffle and, in
case, call the target specific TTI to get the corresponding shuffle cost;
- added a test to verify the cost model analysis on alternate shuffles.
llvm-svn: 212296
Before, we where looking at the size of the pointer type that specifies the
location from which to load the element. This did not make any sense at all.
This change fixes a bug in the delinearization where we failed to delinerize
certain load instructions.
llvm-svn: 210435
The delinearization is needed only to remove the non linearity induced by
expressions involving multiplications of parameters and induction variables.
There is no problem in dealing with constant times parameters, or constant times
an induction variable.
For this reason, the current patch discards all constant terms and multipliers
before running the delinearization algorithm on the terms. The only thing
remaining in the term expressions are parameters and multiply expressions of
parameters: these simplified term expressions are passed to the array shape
recognizer that will not recognize constant dimensions anymore: these will be
recognized as different strides in parametric subscripts.
The only important special case of a constant dimension is the size of elements.
Instead of relying on the delinearization to infer the size of an element,
compute the element size from the base address type. This is a much more precise
way of computing the element size than before, as we would have mixed together
the size of an element with the strides of the innermost dimension.
llvm-svn: 209691
This commit starts with a "git mv ARM64 AArch64" and continues out
from there, renaming the C++ classes, intrinsics, and other
target-local objects for consistency.
"ARM64" test directories are also moved, and tests that began their
life in ARM64 use an arm64 triple, those from AArch64 use an aarch64
triple. Both should be equivalent though.
This finishes the AArch64 merge, and everyone should feel free to
continue committing as normal now.
llvm-svn: 209577
This is a follow-up to r209358: PR19799: Indvars miscompile due to an
incorrect max backedge taken count from SCEV.
That fix was incomplete as pointed out by Arnold and Michael Z. The
code was also too confusing. It needed a careful rewrite with more
unit tests. This version will also happen to optimize more cases.
<rdar://17005101> PR19799: Indvars miscompile...
llvm-svn: 209545
This has to do with the trip count computation for loops with multiple
exits, which is quite subtle. Most passes just ask for a single trip
count number, so we must be conservative assuming any exit could be
taken. Normally, we rely on the "exact" trip count, which was
correctly given as "unknown". However, SCEV also gives a "max"
back-edge taken count. The loops max BE taken count is conservatively
a maximum over the max of each exit's non-exiting iterations
count. Note that some exit tests can be skipped so the max loop
back-edge taken count can actually exceed the max non-exiting
iterations for some exits. However, when we know the loop *latch*
cannot be skipped, we can directly use its max taken count
disregarding other exits. I previously took the minimum here without
checking whether the other exit could be skipped. The correct, and
simpler thing to do here is just to directly use the loop latch's max
non-exiting iterations as the loops max back-edge count.
In the problematic test case, the first loop exit had a max of zero
non-exiting iterations, but could be skipped. The loop latch was known
not to be skipped but had max of one non-exiting iteration. We
incorrectly claimed the loop back-edge could be taken zero times, when
it is actually taken one time.
Fixes Loop %for.body.i: <multiple exits> Unpredictable backedge-taken count.
Loop %for.body.i: max backedge-taken count is 1.
llvm-svn: 209358
Tested by comparing make check VERBOSE=1 before and after to make sure
no tests are missed. (VERBOSE=1 prints the list of tests.)
Only one test :( remains where .cpp is required:
tools/llvm-cov/range_based_for.cpp:// RUN: llvm-cov range_based_for.cpp | FileCheck %s --check-prefix=STDOUT
The topic was discussed in this thread:
http://lists.cs.uiuc.edu/pipermail/llvm-commits/Week-of-Mon-20140428/214905.html
llvm-svn: 208621
To compute the dimensions of the array in a unique way, we split the
delinearization analysis in three steps:
- find parametric terms in all memory access functions
- compute the array dimensions from the set of terms
- compute the delinearized access functions for each dimension
The first step is executed on all the memory access functions such that we
gather all the patterns in which an array is accessed. The second step reduces
all this information in a unique description of the sizes of the array. The
third step is delinearizing each memory access function following the common
description of the shape of the array computed in step 2.
This rewrite of the delinearization pass also solves a problem we had with the
previous implementation: because the previous algorithm was by induction on the
structure of the SCEV, it would not correctly recognize the shape of the array
when the memory access was not following the nesting of the loops: for example,
see polly/test/ScopInfo/multidim_only_ivs_3d_reverse.ll
; void foo(long n, long m, long o, double A[n][m][o]) {
;
; for (long i = 0; i < n; i++)
; for (long j = 0; j < m; j++)
; for (long k = 0; k < o; k++)
; A[i][k][j] = 1.0;
Starting with this patch we no longer delinearize access functions that do not
contain parameters, for example in test/Analysis/DependenceAnalysis/GCD.ll
;; for (long int i = 0; i < 100; i++)
;; for (long int j = 0; j < 100; j++) {
;; A[2*i - 4*j] = i;
;; *B++ = A[6*i + 8*j];
these accesses will not be delinearized as the upper bound of the loops are
constants, and their access functions do not contain SCEVUnknown parameters.
llvm-svn: 208232
This reverts commit r207287, reapplying r207286.
I'm hoping that declaring an explicit struct and instantiating
`addBlockEdges()` directly works around the GCC crash from r207286.
This is a lot more boilerplate, though.
llvm-svn: 207438
This reverts commit r207286. It causes an ICE on the
cmake-llvm-x86_64-linux buildbot [1]:
llvm/lib/Analysis/BlockFrequencyInfo.cpp: In lambda function:
llvm/lib/Analysis/BlockFrequencyInfo.cpp:182:1: internal compiler error: in get_expr_operands, at tree-ssa-operands.c:1035
[1]: http://bb.pgr.jp/builders/cmake-llvm-x86_64-linux/builds/12093/steps/build_llvm/logs/stdio
llvm-svn: 207287
Previously, irreducible backedges were ignored. With this commit,
irreducible SCCs are discovered on the fly, and modelled as loops with
multiple headers.
This approximation specifies the headers of irreducible sub-SCCs as its
entry blocks and all nodes that are targets of a backedge within it
(excluding backedges within true sub-loops). Block frequency
calculations act as if we insert a new block that intercepts all the
edges to the headers. All backedges and entries to the irreducible SCC
point to this imaginary block. This imaginary block has an edge (with
even probability) to each header block.
The result is now reasonable enough that I've added a number of
testcases for irreducible control flow. I've outlined in
`BlockFrequencyInfoImpl.h` ways to improve the approximation.
<rdar://problem/14292693>
llvm-svn: 207286
Remove the concepts of "forward" and "general" mass distributions, which
was wrong. The split might have made sense in an early version of the
algorithm, but it's definitely wrong now.
<rdar://problem/14292693>
llvm-svn: 207195
Strip irreducible testcases to pure control flow. The function calls
made the branch weights more believable but cluttered it up a lot.
There isn't going to be any constant analysis here, so just use dumb
branch logic to clarify the important parts.
<rdar://problem/14292693>
llvm-svn: 207192
The branch that skips irreducible backedges was only active when
propagating mass at the top-level. In particular, when propagating mass
through a loop recognized by `LoopInfo` with irreducible control flow
inside, irreducible backedges would not be skipped.
Not sure where that idea came from, but the result was that mass was
lost until after loop exit. Added a testcase that covers this case.
llvm-svn: 206860
This reverts commit r206707, reapplying r206704. The preceding commit
to CalcSpillWeights should have sorted out the failing buildbots.
<rdar://problem/14292693>
llvm-svn: 206766
This reverts commit r206677, reapplying my BlockFrequencyInfo rewrite.
I've done a careful audit, added some asserts, and fixed a couple of
bugs (unfortunately, they were in unlikely code paths). There's a small
chance that this will appease the failing bots [1][2]. (If so, great!)
If not, I have a follow-up commit ready that will temporarily add
-debug-only=block-freq to the two failing tests, allowing me to compare
the code path between what the failing bots and what my machines (and
the rest of the bots) are doing. Once I've triggered those builds, I'll
revert both commits so the bots go green again.
[1]: http://bb.pgr.jp/builders/ninja-x64-msvc-RA-centos6/builds/1816
[2]: http://llvm-amd64.freebsd.your.org/b/builders/clang-i386-freebsd/builds/18445
<rdar://problem/14292693>
llvm-svn: 206704
This reverts commit r206666, as planned.
Still stumped on why the bots are failing. Sanitizer bots haven't
turned anything up. If anyone can help me debug either of the failures
(referenced in r206666) I'll owe them a beer. (In the meantime, I'll be
auditing my patch for undefined behaviour.)
llvm-svn: 206677
This reverts commit r206628, reapplying r206622 (and r206626).
Two tests are failing only on buildbots [1][2]: i.e., I can't reproduce
on Darwin, and Chandler can't reproduce on Linux. Asan and valgrind
don't tell us anything, but we're hoping the msan bot will catch it.
So, I'm applying this again to get more feedback from the bots. I'll
leave it in long enough to trigger builds in at least the sanitizer
buildbots (it was failing for reasons unrelated to my commit last time
it was in), and hopefully a few others.... and then I expect to revert a
third time.
[1]: http://bb.pgr.jp/builders/ninja-x64-msvc-RA-centos6/builds/1816
[2]: http://llvm-amd64.freebsd.your.org/b/builders/clang-i386-freebsd/builds/18445
llvm-svn: 206666
This reverts commit r206622 and the MSVC fixup in r206626.
Apparently the remotely failing tests are still failing, despite my
attempt to fix the nondeterminism in r206621.
llvm-svn: 206628
This reverts commit r206556, effectively reapplying commit r206548 and
its fixups in r206549 and r206550.
In an intervening commit I've added target triples to the tests that
were failing remotely [1] (but passing locally). I'm hoping the mystery
is solved? I'll revert this again if the tests are still failing
remotely.
[1]: http://bb.pgr.jp/builders/ninja-x64-msvc-RA-centos6/builds/1816
llvm-svn: 206622
LazyCallGraph. This is the start of the whole point of this different
abstraction, but it is just the initial bits. Here is a run-down of
what's going on here. I'm planning to incorporate some (or all) of this
into comments going forward, hopefully with better editing and wording.
=]
The crux of the problem with the traditional way of building SCCs is
that they are ephemeral. The new pass manager however really needs the
ability to associate analysis passes and results of analysis passes with
SCCs in order to expose these analysis passes to the SCC passes. Making
this work is kind-of the whole point of the new pass manager. =]
So, when we're building SCCs for the call graph, we actually want to
build persistent nodes that stick around and can be reasoned about
later. We'd also like the ability to walk the SCC graph in more complex
ways than just the traditional postorder traversal of the current CGSCC
walk. That means that in addition to being persistent, the SCCs need to
be connected into a useful graph structure.
However, we still want the SCCs to be formed lazily where possible.
These constraints are quite hard to satisfy with the SCC iterator. Also,
using that would bypass our ability to actually add data to the nodes of
the call graph to facilite implementing the Tarjan walk. So I've
re-implemented things in a more direct and embedded way. This
immediately makes it easy to get the persistence and connectivity
correct, and it also allows leveraging the existing nodes to simplify
the algorithm. I've worked somewhat to make this implementation more
closely follow the traditional paper's nomenclature and strategy,
although it is still a bit obtuse because it isn't recursive, using
an explicit stack and a tail call instead, and it is interruptable,
resuming each time we need another SCC.
The other tricky bit here, and what actually took almost all the time
and trials and errors I spent building this, is exactly *what* graph
structure to build for the SCCs. The naive thing to build is the call
graph in its newly acyclic form. I wrote about 4 versions of this which
did precisely this. Inevitably, when I experimented with them across
various use cases, they became incredibly awkward. It was all
implementable, but it felt like a complete wrong fit. Square peg, round
hole. There were two overriding aspects that pushed me in a different
direction:
1) We want to discover the SCC graph in a postorder fashion. That means
the root node will be the *last* node we find. Using the call-SCC DAG
as the graph structure of the SCCs results in an orphaned graph until
we discover a root.
2) We will eventually want to walk the SCC graph in parallel, exploring
distinct sub-graphs independently, and synchronizing at merge points.
This again is not helped by the call-SCC DAG structure.
The structure which, quite surprisingly, ended up being completely
natural to use is the *inverse* of the call-SCC DAG. We add the leaf
SCCs to the graph as "roots", and have edges to the caller SCCs. Once
I switched to building this structure, everything just fell into place
elegantly.
Aside from general cleanups (there are FIXMEs and too few comments
overall) that are still needed, the other missing piece of this is
support for iterating across levels of the SCC graph. These will become
useful for implementing #2, but they aren't an immediate priority.
Once SCCs are in good shape, I'll be working on adding mutation support
for incremental updates and adding the pass manager that this analysis
enables.
llvm-svn: 206581
Rewrite the shared implementation of BlockFrequencyInfo and
MachineBlockFrequencyInfo entirely.
The old implementation had a fundamental flaw: precision losses from
nested loops (or very wide branches) compounded past loop exits (and
convergence points).
The @nested_loops testcase at the end of
test/Analysis/BlockFrequencyAnalysis/basic.ll is motivating. This
function has three nested loops, with branch weights in the loop headers
of 1:4000 (exit:continue). The old analysis gives non-sensical results:
Printing analysis 'Block Frequency Analysis' for function 'nested_loops':
---- Block Freqs ----
entry = 1.0
for.cond1.preheader = 1.00103
for.cond4.preheader = 5.5222
for.body6 = 18095.19995
for.inc8 = 4.52264
for.inc11 = 0.00109
for.end13 = 0.0
The new analysis gives correct results:
Printing analysis 'Block Frequency Analysis' for function 'nested_loops':
block-frequency-info: nested_loops
- entry: float = 1.0, int = 8
- for.cond1.preheader: float = 4001.0, int = 32007
- for.cond4.preheader: float = 16008001.0, int = 128064007
- for.body6: float = 64048012001.0, int = 512384096007
- for.inc8: float = 16008001.0, int = 128064007
- for.inc11: float = 4001.0, int = 32007
- for.end13: float = 1.0, int = 8
Most importantly, the frequency leaving each loop matches the frequency
entering it.
The new algorithm leverages BlockMass and PositiveFloat to maintain
precision, separates "probability mass distribution" from "loop
scaling", and uses dithering to eliminate probability mass loss. I have
unit tests for these types out of tree, but it was decided in the review
to make the classes private to BlockFrequencyInfoImpl, and try to shrink
them (or remove them entirely) in follow-up commits.
The new algorithm should generally have a complexity advantage over the
old. The previous algorithm was quadratic in the worst case. The new
algorithm is still worst-case quadratic in the presence of irreducible
control flow, but it's linear without it.
The key difference between the old algorithm and the new is that control
flow within a loop is evaluated separately from control flow outside,
limiting propagation of precision problems and allowing loop scale to be
calculated independently of mass distribution. Loops are visited
bottom-up, their loop scales are calculated, and they are replaced by
pseudo-nodes. Mass is then distributed through the function, which is
now a DAG. Finally, loops are revisited top-down to multiply through
the loop scales and the masses distributed to pseudo nodes.
There are some remaining flaws.
- Irreducible control flow isn't modelled correctly. LoopInfo and
MachineLoopInfo ignore irreducible edges, so this algorithm will
fail to scale accordingly. There's a note in the class
documentation about how to get closer. See also the comments in
test/Analysis/BlockFrequencyInfo/irreducible.ll.
- Loop scale is limited to 4096 per loop (2^12) to avoid exhausting
the 64-bit integer precision used downstream.
- The "bias" calculation proposed on llvmdev is *not* incorporated
here. This will be added in a follow-up commit, once comments from
this review have been handled.
llvm-svn: 206548
Previously, BranchProbabilityInfo::calcLoopBranchHeuristics would determine the weights of basic blocks inside loops even when it didn't have enough information to estimate the branch probabilities correctly. This patch fixes the function to exit early if it doesn't see any exit edges or back edges and let the later heuristics determine the weights.
This fixes PR18705 and <rdar://problem/15991090>.
Differential Revision: http://reviews.llvm.org/D3363
llvm-svn: 206194
BasicTTI::getMemoryOpCost must explicitly check for non-simple types; setting
AllowUnknown=true with TLI->getSimpleValueType is not sufficient because, for
example, non-power-of-two vector types return non-simple EVTs (not MVT::Other).
llvm-svn: 206150
This provides more realistic costs for the insert/extractelement instructions
(which are load/store pairs), accounts for the cheap unaligned Altivec load
sequence, and for unaligned VSX load/stores.
Bad news:
MultiSource/Applications/sgefa/sgefa - 35% slowdown (this will require more investigation)
SingleSource/Benchmarks/McGill/queens - 20% slowdown (we no longer vectorize this, but it was a constant store that was scalarized)
MultiSource/Benchmarks/FreeBench/pcompress2/pcompress2 - 2% slowdown
Good news:
SingleSource/Benchmarks/Shootout/ary3 - 54% speedup
SingleSource/Benchmarks/Shootout-C++/ary - 40% speedup
MultiSource/Benchmarks/Ptrdist/ks/ks - 35% speedup
MultiSource/Benchmarks/FreeBench/neural/neural - 30% speedup
MultiSource/Benchmarks/TSVC/Symbolics-flt/Symbolics-flt - 20% speedup
Unfortunately, estimating the costs of the stack-based scalarization sequences
is hard, and adjusting these costs is like a game of whac-a-mole :( I'll
revisit this again after we have better codegen for vector extloads and
truncstores and unaligned load/stores.
llvm-svn: 205658
When a vector type legalizes to a larger vector type, and the target does not
support the associated extending load (or truncating store), then legalization
will scalarize the load (or store) resulting in an associated scalarization
cost. BasicTTI::getMemoryOpCost needs to account for this.
Between this, and r205487, PowerPC on the P7 with VSX enabled shows:
MultiSource/Benchmarks/PAQ8p/paq8p: 43% speedup
SingleSource/Benchmarks/BenchmarkGame/puzzle: 51% speedup
SingleSource/UnitTests/Vectorizer/gcc-loops 28% speedup
(some of these are new; some of these, such as PAQ8p, just reverse regressions
that VSX support would trigger)
llvm-svn: 205495
For an cast (extension, etc.), the currently logic predicts a low cost if the
associated operation (keyed on the destination type) is legal (or promoted).
This is not true when the number of values required to legalize the type is
changing. For example, <8 x i16> being sign extended by <8 x i32> is not
generically cheap on PPC with VSX, even though sign extension to v4i32 is
legal, because two output v4i32 values are required compared to the single
v8i16 input value, and without custom logic in the target, this conversion will
scalarize.
llvm-svn: 205487
This adds a second implementation of the AArch64 architecture to LLVM,
accessible in parallel via the "arm64" triple. The plan over the
coming weeks & months is to merge the two into a single backend,
during which time thorough code review should naturally occur.
Everything will be easier with the target in-tree though, hence this
commit.
llvm-svn: 205090
This commit consist of two parts.
The first part fix the PR15967. The wrong conclusion was made when the MaxLookup
limit was reached. The fix introduce a out parameter (MaxLookupReached) to
DecomposeGEPExpression that the function aliasGEP can act upon.
The second part is introducing the constant MaxLookupSearchDepth to make sure
that DecomposeGEPExpression and GetUnderlyingObject use the same search depth.
This is a small cleanup to clarify the original algorithm.
Patch by Karl-Johan Karlsson!
llvm-svn: 204859
If we have a loop of the form
for (unsigned n = 0; n != (k & -32); n += 32) {}
then we know that n is always divisible by 32 and the loop must
terminate. Even if we have a condition where the loop counter will
overflow it'll always hold this invariant.
PR19183. Our loop vectorizer creates this pattern and it's also
occasionally formed by loop counters derived from pointers.
llvm-svn: 204728
On ELF and COFF an alias is just another name for a position in the file.
There is no way to refer to a position in another file, so an alias to
undefined is meaningless.
MachO currently doesn't support aliases. The spec has a N_INDR, which when
implemented will have a different set of restrictions. Adding support for
it shouldn't be harder than any other IR extension.
For now, having the IR represent what is actually possible with current
tools makes it easier to fix the design of GlobalAlias.
llvm-svn: 203705
the legalization cost must be included to get an accurate
estimation of the total cost of the scalarized vector.
The inaccurate cost triggered unprofitable SLP vectorization on
32-bit X86.
Summary:
Include legalization overhead when computing scalarization cost
Reviewers: hfinkel, nadav
CC: chandlerc, rnk, llvm-commits
Differential Revision: http://llvm-reviews.chandlerc.com/D2992
llvm-svn: 203509
'OK_NonUniformConstValue' to identify operands which are constants but
not constant splats.
The cost model now allows returning 'OK_NonUniformConstValue'
for non splat operands that are instances of ConstantVector or
ConstantDataVector.
With this change, targets are now able to compute different costs
for instructions with non-uniform constant operands.
For example, On X86 the cost of a vector shift may vary depending on whether
the second operand is a uniform or non-uniform constant.
This patch applies the following changes:
- The cost model computation now takes into account non-uniform constants;
- The cost of vector shift instructions has been improved in
X86TargetTransformInfo analysis pass;
- BBVectorize, SLPVectorizer and LoopVectorize now know how to distinguish
between non-uniform and uniform constant operands.
Added a new test to verify that the output of opt
'-cost-model -analyze' is valid in the following configurations: SSE2,
SSE4.1, AVX, AVX2.
llvm-svn: 201272
The most important part of this is probably adding any cost at all for
operations like zext <8 x i8> to <8 x i32>. Before they were being
recorded as extremely costly (24, I believe) which made LLVM fall back
on a 4-wide vectorisation of a loop.
It also rebalances the values for sext, zext and trunc. Lacking any
other sane metric that might work across CPU microarchitectures I went
for instructions. This seems to be in reasonable accord with the rest
of the table (sitofp, ...) though no doubt at least one value is
sub-optimal for some bizarre reason.
Finally, separate AVX and AVX2 values are provided where appropriate.
The CodeGen is quite different in many cases.
rdar://problem/15981990
llvm-svn: 200928
The primary motivation for this pass is to separate the call graph
analysis used by the new pass manager's CGSCC pass management from the
existing call graph analysis pass. That analysis pass is (somewhat
unfortunately) over-constrained by the existing CallGraphSCCPassManager
requirements. Those requirements make it *really* hard to cleanly layer
the needed functionality for the new pass manager on top of the existing
analysis.
However, there are also a bunch of things that the pass manager would
specifically benefit from doing differently from the existing call graph
analysis, and this new implementation tries to address several of them:
- Be lazy about scanning function definitions. The existing pass eagerly
scans the entire module to build the initial graph. This new pass is
significantly more lazy, and I plan to push this even further to
maximize locality during CGSCC walks.
- Don't use a single synthetic node to partition functions with an
indirect call from functions whose address is taken. This node creates
a huge choke-point which would preclude good parallelization across
the fanout of the SCC graph when we got to the point of looking at
such changes to LLVM.
- Use a memory dense and lightweight representation of the call graph
rather than value handles and tracking call instructions. This will
require explicit update calls instead of some updates working
transparently, but should end up being significantly more efficient.
The explicit update calls ended up being needed in many cases for the
existing call graph so we don't really lose anything.
- Doesn't explicitly model SCCs and thus doesn't provide an "identity"
for an SCC which is stable across updates. This is essential for the
new pass manager to work correctly.
- Only form the graph necessary for traversing all of the functions in
an SCC friendly order. This is a much simpler graph structure and
should be more memory dense. It does limit the ways in which it is
appropriate to use this analysis. I wish I had a better name than
"call graph". I've commented extensively this aspect.
This is still very much a WIP, in fact it is really just the initial
bits. But it is about the fourth version of the initial bits that I've
implemented with each of the others running into really frustrating
problms. This looks like it will actually work and I'd like to split the
actual complexity across commits for the sake of my reviewers. =] The
rest of the implementation along with lots of wiring will follow
somewhat more rapidly now that there is a good path forward.
Naturally, this doesn't impact any of the existing optimizer. This code
is specific to the new pass manager.
A bunch of thanks are deserved for the various folks that have helped
with the design of this, especially Nick Lewycky who actually sat with
me to go through the fundamentals of the final version here.
llvm-svn: 200903
Unfortunately, this in turn led to some lower quality SCEVs due to some different paths through expression simplification, so add getUDivExactExpr and use it. This fixes all instances of the problems that I found, but we can make that function smarter as necessary.
Merge test "xor-and.ll" into "and-xor.ll" since I needed to update it anyways. Test 'nsw-offset.ll' analyzes a little deeper, %n now gets a scev in terms of %no instead of a SCEVUnknown.
llvm-svn: 200203
cycles
This allows the value equality check to work even if we don't have a dominator
tree. Also add some more comments.
I was worried about compile time impacts and did not implement reachability but
used the dominance check in the initial patch. The trade-off was that the
dominator tree was required.
The llvm utility function isPotentiallyReachable cuts off the recursive search
after 32 visits. Testing did not show any compile time regressions showing my
worries unjustfied.
No compile time or performance regressions at O3 -flto -mavx on test-suite +
externals.
Addresses review comments from r198290.
llvm-svn: 198400
When there are cycles in the value graph we have to be careful interpreting
"Value*" identity as "value" equivalence. We interpret the value of a phi node
as the value of its operands.
When we check for value equivalence now we make sure that the "Value*" dominates
all cycles (phis).
%0 = phi [%noaliasval, %addr2]
%l = load %ptr
%addr1 = gep @a, 0, %l
%addr2 = gep @a, 0, (%l + 1)
store %ptr ...
Before this patch we would return NoAlias for (%0, %addr1) which is wrong
because the value of the load is from different iterations of the loop.
Tested on x86_64 -mavx at O3 and O3 -flto with no performance or compile time
regressions.
PR18068
radar://15653794
llvm-svn: 198290
The tests just hit this with a different sized
address space since I haven't figured out how
to use this to break it.
I thought I committed this a long time ago,
and I'm not sure why missing this hasn't caused
any problems.
llvm-svn: 194903
Patch by Michele Scandale!
Rewrite of the functions used to compute the backedge taken count of a
loop on LT and GT comparisons.
I decided to split the handling of LT and GT cases becasue the trick
"a > b == -a < -b" in some cases prevents the trip count computation
due to the multiplication by -1 on the two operands of the
comparison. This issue comes from the conservative computation of
value range of SCEVs: taking the negative SCEV of an expression that
have a small positive range (e.g. [0,31]), we would have a SCEV with a
fullset as value range.
Indeed, in the new rewritten function I tried to better handle the
maximum backedge taken count computation when MAX/MIN expression are
used to handle the cases where no entry guard is found.
Some test have been modified in order to check the new value correctly
(I manually check them and reasoning on possible overflow the new
values seem correct).
I finally added a new test case related to the multiplication by -1
issue on GT comparisons.
llvm-svn: 194116
This adds another heuristic to BPI, similar to the existing heuristic that
considers (x == 0) unlikely to be true. As suggested in the PACT'98 paper by
Deitrich, Cheng, and Hwu, -1 is often used to indicate an invalid index, and
equality comparisons with -1 are also unlikely to succeed. Local
experimentation supports this hypothesis: This yields a 1-2% speedup in the
test-suite sqlite benchmark on the PPC A2 core, with no significant
regressions.
llvm-svn: 193855
We can't do this for the general case as saying a GEP with a negative index
doesn't have unsigned wrap isn't valid for negative indices.
%gep = getelementptr inbounds i32* %p, i64 -1
But an inbounds GEP cannot run past the end of address space. So we check for
the very common case of a positive index and make GEPs derived from that NUW.
Together with Andy's recent non-unit stride work this lets us analyze loops
like
void foo3(int *a, int *b) {
for (; a < b; a++) {}
}
PR12375, PR12376.
Differential Revision: http://llvm-reviews.chandlerc.com/D2033
llvm-svn: 193514
Major steps include:
1). introduces a not-addr-taken bit-field in GlobalVariable
2). GlobalOpt pass sets "not-address-taken" if it proves a global varirable
dosen't have its address taken.
3). AA use this info for disambiguation.
llvm-svn: 193251
We can have a struct type with a single field and the field does not start
with 0. In that case, we should correctly update the offset.
llvm-svn: 193137
The test before wasn't successfully testing this
since it was missing the datalayout piece to change
the size of the second address space.
llvm-svn: 193102
SCEV currently fails to compute loop counts for nonunit stride
loops. This comes up frequently. It prevents loop optimization and
forces vectorization to insert extra loop checks.
For example:
void foo(int n, int *x) {
for (int i = 0; i < n; i += 3) {
x[i] = i;
x[i+1] = i+1;
x[i+2] = i+2;
}
}
We need to properly handle the case in which limit > INT_MAX-stride. In
the above case: n > INT_MAX-3. In this case the loop counter will step
beyond the limit and overflow at the same time. However, knowing that
signed integer overlow in undefined, we can assume the loop test
behavior is arbitrary after overflow. This obeys both C undefined
behavior rules, and the more strict LLVM poison value rules.
I'm finally fixing this in response to Hal Finkel's persistence.
The most probable reason that we never optimized this before is that
we were being careful to handle case where the developer expected a
side-effect free infinite loop relying on overflow:
for (int i = 0; i < n; i += s) {
++j;
}
return j;
If INT_MAX+1 is a multiple of s and n > INT_MAX-s, then we might
expect an infinite loop. However there are plenty of ways to achieve
this effect without relying on undefined behavior of signed overflow.
llvm-svn: 193015
infrastructure.
This was essentially work toward PGO based on a design that had several
flaws, partially dating from a time when LLVM had a different
architecture, and with an effort to modernize it abandoned without being
completed. Since then, it has bitrotted for several years further. The
result is nearly unusable, and isn't helping any of the modern PGO
efforts. Instead, it is getting in the way, adding confusion about PGO
in LLVM and distracting everyone with maintenance on essentially dead
code. Removing it paves the way for modern efforts around PGO.
Among other effects, this removes the last of the runtime libraries from
LLVM. Those are being developed in the separate 'compiler-rt' project
now, with somewhat different licensing specifically more approriate for
runtimes.
llvm-svn: 191835
Remove the command line argument "struct-path-tbaa" since we should not depend
on command line argument to decide which format the IR file is using. Instead,
we check the first operand of the tbaa tag node, if it is a MDNode, we treat
it as struct-path aware TBAA format, otherwise, we treat it as scalar TBAA
format.
When clang starts to use struct-path aware TBAA format no matter whether
struct-path-tbaa is no, and we can auto-upgrade existing bc files, the support
for scalar TBAA format can be dropped.
Existing testing cases are updated to use the struct-path aware TBAA format.
llvm-svn: 191538
Upcoming SLP vectorization improvements will want to be able to estimate costs
of horizontal reductions. Add infrastructure to support this.
We model reductions as a series of (shufflevector,add) tuples ultimately
followed by an extractelement. For example, for an add-reduction of <4 x float>
we could generate the following sequence:
(v0, v1, v2, v3)
\ \ / /
\ \ /
+ +
(v0+v2, v1+v3, undef, undef)
\ /
((v0+v2) + (v1+v3), undef, undef)
%rdx.shuf = shufflevector <4 x float> %rdx, <4 x float> undef,
<4 x i32> <i32 2, i32 3, i32 undef, i32 undef>
%bin.rdx = fadd <4 x float> %rdx, %rdx.shuf
%rdx.shuf7 = shufflevector <4 x float> %bin.rdx, <4 x float> undef,
<4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
%bin.rdx8 = fadd <4 x float> %bin.rdx, %rdx.shuf7
%r = extractelement <4 x float> %bin.rdx8, i32 0
This commit adds a cost model interface "getReductionCost(Opcode, Ty, Pairwise)"
that will allow clients to ask for the cost of such a reduction (as backends
might generate more efficient code than the cost of the individual instructions
summed up). This interface is excercised by the CostModel analysis pass which
looks for reduction patterns like the one above - starting at extractelements -
and if it sees a matching sequence will call the cost model interface.
We will also support a second form of pairwise reduction that is well supported
on common architectures (haddps, vpadd, faddp).
(v0, v1, v2, v3)
\ / \ /
(v0+v1, v2+v3, undef, undef)
\ /
((v0+v1)+(v2+v3), undef, undef, undef)
%rdx.shuf.0.0 = shufflevector <4 x float> %rdx, <4 x float> undef,
<4 x i32> <i32 0, i32 2 , i32 undef, i32 undef>
%rdx.shuf.0.1 = shufflevector <4 x float> %rdx, <4 x float> undef,
<4 x i32> <i32 1, i32 3, i32 undef, i32 undef>
%bin.rdx.0 = fadd <4 x float> %rdx.shuf.0.0, %rdx.shuf.0.1
%rdx.shuf.1.0 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef,
<4 x i32> <i32 0, i32 undef, i32 undef, i32 undef>
%rdx.shuf.1.1 = shufflevector <4 x float> %bin.rdx.0, <4 x float> undef,
<4 x i32> <i32 1, i32 undef, i32 undef, i32 undef>
%bin.rdx.1 = fadd <4 x float> %rdx.shuf.1.0, %rdx.shuf.1.1
%r = extractelement <4 x float> %bin.rdx.1, i32 0
llvm-svn: 190876
- Instead of setting the suffixes in a bunch of places, just set one master
list in the top-level config. We now only modify the suffix list in a few
suites that have one particular unique suffix (.ml, .mc, .yaml, .td, .py).
- Aside from removing the need for a bunch of lit.local.cfg files, this enables
4 tests that were inadvertently being skipped (one in
Transforms/BranchFolding, a .s file each in DebugInfo/AArch64 and
CodeGen/PowerPC, and one in CodeGen/SI which is now failing and has been
XFAILED).
- This commit also fixes a bunch of config files to use config.root instead of
older copy-pasted code.
llvm-svn: 188513
This patch fixes the multiple breakages on ARM test-suite after the SLP
vectorizer was introduced by default on O3. The problem was an illegal
vector type on ARMTTI::getCmpSelInstrCost() <3 x i1> which is not simple.
The guard protects this code from breaking (cause of the problems) but
doesn't fix the issue that is generating the odd vector in the first
place, which also needs to be investigated.
llvm-svn: 187658
This fixes an oversight that Intrinsic::nearbyint was not being mapped to
ISD::FNEARBYINT (thus fixing the over-optimistic cost we were assigning to
nearbyint calls for some targets).
llvm-svn: 185783
This is easier to read than the internal fixed-point representation.
If anybody knows the correct algorithm for converting fixed-point
numbers to base 10, feel free to fix it.
llvm-svn: 184881
This is a band-aid to fix the most severe regressions we're seeing from basing
spill decisions on block frequencies, until we have a better solution.
llvm-svn: 184835
Zero is used by BlockFrequencyInfo as a special "don't know" value. It also
causes a sink for frequencies as you can't ever get off a zero frequency with
more multiplies.
This recovers a 10% regression on MultiSource/Benchmarks/7zip. A zero frequency
was propagated into an inner loop causing excessive spilling.
PR16402.
llvm-svn: 184584
Other than recognizing the attribute, the patch does little else.
It changes the branch probability analyzer so that edges into
blocks postdominated by a cold function are given low weight.
Added analysis and code generation tests. Added documentation for the
new attribute.
llvm-svn: 182638
We switch the order of offset and field type to make TBAAStructType node
(name, parent node, offset) similar to scalar TBAA node (name, parent node).
TypeIsImmutable is added to TBAAStructTag node.
llvm-svn: 180654
Rather than just splitting the input type and hoping for the best, apply
a bit more cleverness. Just splitting the types until the source is
legal often leads to an illegal result time, which is then widened and a
scalarization step is introduced which leads to truly horrible code
generation. With the loop vectorizer, these sorts of operations are much
more common, and so it's worth extra effort to do them well.
Add a legalization hook for the operands of a TRUNCATE node, which will
be encountered after the result type has been legalized, but if the
operand type is still illegal. If simple splitting of both types
ends up with the result type of each half still being legal, just
do that (v16i16 -> v16i8 on ARM, for example). If, however, that would
result in an illegal result type (v8i32 -> v8i8 on ARM, for example),
we can get more clever with power-two vectors. Specifically,
split the input type, but also widen the result element size, then
concatenate the halves and truncate again. For example on ARM,
To perform a "%res = v8i8 trunc v8i32 %in" we transform to:
%inlo = v4i32 extract_subvector %in, 0
%inhi = v4i32 extract_subvector %in, 4
%lo16 = v4i16 trunc v4i32 %inlo
%hi16 = v4i16 trunc v4i32 %inhi
%in16 = v8i16 concat_vectors v4i16 %lo16, v4i16 %hi16
%res = v8i8 trunc v8i16 %in16
This allows instruction selection to generate three VMOVN instructions
instead of a sequences of moves, stores and loads.
Update the ARMTargetTransformInfo to take this improved legalization
into account.
Consider the simplified IR:
define <16 x i8> @test1(<16 x i32>* %ap) {
%a = load <16 x i32>* %ap
%tmp = trunc <16 x i32> %a to <16 x i8>
ret <16 x i8> %tmp
}
define <8 x i8> @test2(<8 x i32>* %ap) {
%a = load <8 x i32>* %ap
%tmp = trunc <8 x i32> %a to <8 x i8>
ret <8 x i8> %tmp
}
Previously, we would generate the truly hideous:
.syntax unified
.section __TEXT,__text,regular,pure_instructions
.globl _test1
.align 2
_test1: @ @test1
@ BB#0:
push {r7}
mov r7, sp
sub sp, sp, #20
bic sp, sp, #7
add r1, r0, #48
add r2, r0, #32
vld1.64 {d24, d25}, [r0:128]
vld1.64 {d16, d17}, [r1:128]
vld1.64 {d18, d19}, [r2:128]
add r1, r0, #16
vmovn.i32 d22, q8
vld1.64 {d16, d17}, [r1:128]
vmovn.i32 d20, q9
vmovn.i32 d18, q12
vmov.u16 r0, d22[3]
strb r0, [sp, #15]
vmov.u16 r0, d22[2]
strb r0, [sp, #14]
vmov.u16 r0, d22[1]
strb r0, [sp, #13]
vmov.u16 r0, d22[0]
vmovn.i32 d16, q8
strb r0, [sp, #12]
vmov.u16 r0, d20[3]
strb r0, [sp, #11]
vmov.u16 r0, d20[2]
strb r0, [sp, #10]
vmov.u16 r0, d20[1]
strb r0, [sp, #9]
vmov.u16 r0, d20[0]
strb r0, [sp, #8]
vmov.u16 r0, d18[3]
strb r0, [sp, #3]
vmov.u16 r0, d18[2]
strb r0, [sp, #2]
vmov.u16 r0, d18[1]
strb r0, [sp, #1]
vmov.u16 r0, d18[0]
strb r0, [sp]
vmov.u16 r0, d16[3]
strb r0, [sp, #7]
vmov.u16 r0, d16[2]
strb r0, [sp, #6]
vmov.u16 r0, d16[1]
strb r0, [sp, #5]
vmov.u16 r0, d16[0]
strb r0, [sp, #4]
vldmia sp, {d16, d17}
vmov r0, r1, d16
vmov r2, r3, d17
mov sp, r7
pop {r7}
bx lr
.globl _test2
.align 2
_test2: @ @test2
@ BB#0:
push {r7}
mov r7, sp
sub sp, sp, #12
bic sp, sp, #7
vld1.64 {d16, d17}, [r0:128]
add r0, r0, #16
vld1.64 {d20, d21}, [r0:128]
vmovn.i32 d18, q8
vmov.u16 r0, d18[3]
vmovn.i32 d16, q10
strb r0, [sp, #3]
vmov.u16 r0, d18[2]
strb r0, [sp, #2]
vmov.u16 r0, d18[1]
strb r0, [sp, #1]
vmov.u16 r0, d18[0]
strb r0, [sp]
vmov.u16 r0, d16[3]
strb r0, [sp, #7]
vmov.u16 r0, d16[2]
strb r0, [sp, #6]
vmov.u16 r0, d16[1]
strb r0, [sp, #5]
vmov.u16 r0, d16[0]
strb r0, [sp, #4]
ldm sp, {r0, r1}
mov sp, r7
pop {r7}
bx lr
Now, however, we generate the much more straightforward:
.syntax unified
.section __TEXT,__text,regular,pure_instructions
.globl _test1
.align 2
_test1: @ @test1
@ BB#0:
add r1, r0, #48
add r2, r0, #32
vld1.64 {d20, d21}, [r0:128]
vld1.64 {d16, d17}, [r1:128]
add r1, r0, #16
vld1.64 {d18, d19}, [r2:128]
vld1.64 {d22, d23}, [r1:128]
vmovn.i32 d17, q8
vmovn.i32 d16, q9
vmovn.i32 d18, q10
vmovn.i32 d19, q11
vmovn.i16 d17, q8
vmovn.i16 d16, q9
vmov r0, r1, d16
vmov r2, r3, d17
bx lr
.globl _test2
.align 2
_test2: @ @test2
@ BB#0:
vld1.64 {d16, d17}, [r0:128]
add r0, r0, #16
vld1.64 {d18, d19}, [r0:128]
vmovn.i32 d16, q8
vmovn.i32 d17, q9
vmovn.i16 d16, q8
vmov r0, r1, d16
bx lr
llvm-svn: 179989
The costs are overfitted so that I can still use the legalization factor.
For example the following kernel has about half the throughput vectorized than
unvectorized when compiled with SSE2. Before this patch we would vectorize it.
unsigned short A[1024];
double B[1024];
void f() {
int i;
for (i = 0; i < 1024; ++i) {
B[i] = (double) A[i];
}
}
radar://13599001
llvm-svn: 179033
The code in getTypeConversion attempts to promote the element vector type
before it trys to split or widen the vector.
After it failed finding a legal vector type by promoting it would continue using
the promoted vector element type. Thereby missing legal splitted vector types.
For example the type v32i32 that has a legal split of 4 x v3i32 on x86/sse2
would be transformed to: v32i256 and from there on successively split to:
v16i256, v8i256, v1i256 and then finally ends up as an i64 type.
By resetting the vector element type to the original vector element type that
existed before the promotion the code will attempt to split the vector type to
smaller vector widths of the same type.
llvm-svn: 178999
SSE2 has efficient support for shifts by a scalar. My previous change of making
shifts expensive did not take this into account marking all shifts as expensive.
This would prevent vectorization from happening where it is actually beneficial.
With this change we differentiate between shifts of constants and other shifts.
radar://13576547
llvm-svn: 178808
The default logic does not correctly identify costs of casts because they are
marked as custom on x86.
For some cases, where the shift amount is a scalar we would be able to generate
better code. Unfortunately, when this is the case the value (the splat) will get
hoisted out of the loop, thereby making it invisible to ISel.
radar://13130673
radar://13537826
llvm-svn: 178703
Fixes PR15570: SEGV: SCEV back-edge info invalid after dead code removal.
Indvars creates a SCEV expression for the loop's back edge taken
count, then determines that the comparison is always true and
removes it.
When loop-unroll asks for the expression, it contains a NULL
SCEVUnknkown (as a CallbackVH).
forgetMemoizedResults should invalidate the loop back edges expression.
llvm-svn: 177986
Add "evaluate-tbaa" to print alias queries of loads/stores. Alias queries
between pointers do not include TBAA tags.
Add testing case for "placement new". TBAA currently says NoAlias.
llvm-svn: 177772
- After moving logic recognizing vector shift with scalar amount from
DAG combining into DAG lowering, we declare to customize all vector
shifts even vector shift on AVX is legal. As a result, the cost model
needs special tuning to identify these legal cases.
llvm-svn: 177586
The ARM backend currently has poor codegen for long sext/zext
operations, such as v8i8 -> v8i32. This patch addresses this
by performing a custom expansion in ARMISelLowering. It also
adds/changes the cost of such lowering in ARMTTI.
This partially addresses PR14867.
Patch by Pete Couperus
llvm-svn: 177380
The default logic marks them as too expensive.
For example, before this patch we estimated:
cost of 16 for instruction: %r = uitofp <4 x i16> %v0 to <4 x float>
While this translates to:
vmovl.u16 q8, d16
vcvt.f32.u32 q8, q8
All other costs are left to the values assigned by the fallback logic. Theses
costs are mostly reasonable in the sense that they get progressively more
expensive as the instruction sequences emitted get longer.
radar://13445992
llvm-svn: 177334
Fix cost of some "cheap" cast instructions. Before this patch we used to
estimate for example:
cost of 16 for instruction: %r = fptoui <4 x float> %v0 to <4 x i16>
While we would emit:
vcvt.s32.f32 q8, q8
vmovn.i32 d16, q8
vuzp.8 d16, d17
All other costs are left to the values assigned by the fallback logic. Theses
costs are mostly reasonable in the sense that they get progressively more
expensive as the instruction sequences emitted get longer.
radar://13434072
llvm-svn: 177333
I was too pessimistic in r177105. Vector selects that fit into a legal register
type lower just fine. I was mislead by the code fragment that I was using. The
stores/loads that I saw in those cases came from lowering the conditional off
an address.
Changing the code fragment to:
%T0_3 = type <8 x i18>
%T1_3 = type <8 x i1>
define void @func_blend3(%T0_3* %loadaddr, %T0_3* %loadaddr2,
%T1_3* %blend, %T0_3* %storeaddr) {
%v0 = load %T0_3* %loadaddr
%v1 = load %T0_3* %loadaddr2
==> FROM:
;%c = load %T1_3* %blend
==> TO:
%c = icmp slt %T0_3 %v0, %v1
==> USE:
%r = select %T1_3 %c, %T0_3 %v0, %T0_3 %v1
store %T0_3 %r, %T0_3* %storeaddr
ret void
}
revealed this mistake.
radar://13403975
llvm-svn: 177170
By terrible I mean we store/load from the stack.
This matters on PAQp8 in _Z5trainPsS_ii (which is inlined into Mixer::update)
where we decide to vectorize a loop with a VF of 8 resulting in a 25%
degradation on a cortex-a8.
LV: Found an estimated cost of 2 for VF 8 For instruction: icmp slt i32
LV: Found an estimated cost of 2 for VF 8 For instruction: select i1, i32, i32
The bug that tracks the CodeGen part is PR14868.
radar://13403975
llvm-svn: 177105
Increase the cost of v8/v16-i8 to v8/v16-i32 casts and truncates as the backend
currently lowers those using stack accesses.
This was responsible for a significant degradation on
MultiSource/Benchmarks/Trimaran/enc-pc1/enc-pc1
where we vectorize one loop to a vector factor of 16. After this patch we select
a vector factor of 4 which will generate reasonable code.
unsigned char cle[32];
void test(short c) {
unsigned short compte;
for (compte = 0; compte <= 31; compte++) {
cle[compte] = cle[compte] ^ c;
}
}
radar://13220512
llvm-svn: 176898
Summary:
Statistics are still available in Release+Asserts (any +Asserts builds),
and stats can also be turned on with LLVM_ENABLE_STATS.
Move some of the FastISel stats that were moved under DEBUG()
back out of DEBUG(), since stats are disabled across the board now.
Many tests depend on grepping "-stats" output. Move those into
a orig_dir/Stats/. so that they can be marked as unsupported
when building without statistics.
Differential Revision: http://llvm-reviews.chandlerc.com/D486
llvm-svn: 176733
The "invariant.load" metadata indicates the memory unit being accessed is immutable.
A load annotated with this metadata can be moved across any store.
As I am not sure if it is legal to move such loads across barrier/fence, this
change dose not allow such transformation.
rdar://11311484
Thank Arnold for code review.
llvm-svn: 176562
This matters for example in following matrix multiply:
int **mmult(int rows, int cols, int **m1, int **m2, int **m3) {
int i, j, k, val;
for (i=0; i<rows; i++) {
for (j=0; j<cols; j++) {
val = 0;
for (k=0; k<cols; k++) {
val += m1[i][k] * m2[k][j];
}
m3[i][j] = val;
}
}
return(m3);
}
Taken from the test-suite benchmark Shootout.
We estimate the cost of the multiply to be 2 while we generate 9 instructions
for it and end up being quite a bit slower than the scalar version (48% on my
machine).
Also, properly differentiate between avx1 and avx2. On avx-1 we still split the
vector into 2 128bits and handle the subvector muls like above with 9
instructions.
Only on avx-2 will we have a cost of 9 for v4i64.
I changed the test case in test/Transforms/LoopVectorize/X86/avx1.ll to use an
add instead of a mul because with a mul we now no longer vectorize. I did
verify that the mul would be indeed more expensive when vectorized with 3
kernels:
for (i ...)
r += a[i] * 3;
for (i ...)
m1[i] = m1[i] * 3; // This matches the test case in avx1.ll
and a matrix multiply.
In each case the vectorized version was considerably slower.
radar://13304919
llvm-svn: 176403
We make the cost for calling libm functions extremely high as emitting the
calls is expensive and causes spills (on x86) so performance suffers. We still
vectorize important calls like ceilf and friends on SSE4.1. and fabs.
Differential Revision: http://llvm-reviews.chandlerc.com/D466
llvm-svn: 176287
Listing all of the attributes for the callee of a call/invoke instruction is way
too much and makes the IR unreadable. Use references to attributes instead.
llvm-svn: 175877
sext <4 x i1> to <4 x i64>
sext <4 x i8> to <4 x i64>
sext <4 x i16> to <4 x i64>
I'm running Combine on SIGN_EXTEND_IN_REG and revert SEXT patterns:
(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) -> (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
The sext_in_reg (v4i32 x) may be lowered to shl+sar operations.
The "sar" does not exist on 64-bit operation, so lowering sext_in_reg (v4i64 x) has no vector solution.
I also added a cost of this operations to the AVX costs table.
llvm-svn: 175619
Profiling tests *do* need a JIT. They'll pass if a cross-compiler targetting
AArch64 by default has been built, but fail if a native AArch64 compiler has
been build. Therefore XFAIL is inappropriate and we mark them unsupported.
ExecutionEngine tests are JIT by definition, they should also be unsupported.
Transforms/LICM only uses the interpreter to check the output is still sane
after optimisation. It can be switched to use an interpreter.
llvm-svn: 175433
Thanks to help from Nadav and Hal, I have a more reasonable (and even
correct!) approach. This specifically penalizes the insertelement
and extractelement operations for the performance hit that will occur
on PowerPC processors.
llvm-svn: 174725
Adds a function to target transform info to query for the cost of address
computation. The cost model analysis pass now also queries this interface.
The code in LoopVectorize adds the cost of address computation as part of the
memory instruction cost calculation. Only there, we know whether the instruction
will be scalarized or not.
Increase the penality for inserting in to D registers on swift. This becomes
necessary because we now always assume that address computation has a cost and
three is a closer value to the architecture.
radar://13097204
llvm-svn: 174713
Swift has a renaming dependency if we load into D subregisters. We don't have a
way of distinguishing between insertelement operations of values from loads and
other values. Therefore, we are pessimistic for now (The performance problem
showed up in example 14 of gcc-loops).
radar://13096933
llvm-svn: 174300
This provides a place to add customized operation cost information and
control some other target-specific IR-level transformations.
The only non-trivial logic in this checkin assigns a higher cost to
unaligned loads and stores (covered by the included test case).
llvm-svn: 173520
Nick Lewycky