The old method used by X86TTI to determine partial-unrolling thresholds was
messy (because it worked by testing target features), and also would not
correctly identify the target CPU if certain target features were disabled.
After some discussions on IRC with Chandler et al., it was decided that the
processor scheduling models were the right containers for this information
(because it is often tied to special uop dispatch-buffer sizes).
This does represent a small functionality change:
- For generic x86-64 (which uses the SB model and, thus, will get some
unrolling).
- For AMD cores (because they still currently use the SB scheduling model)
- For Haswell (based on benchmarking by Louis Gerbarg, it was decided to bump
the default threshold to 50; we're working on a test case for this).
Otherwise, nothing has changed for any other targets. The logic, however, has
been moved into BasicTTI, so other targets may now also opt-in to this
functionality simply by setting LoopMicroOpBufferSize in their processor
model definitions.
llvm-svn: 208289
This patch changes the vectorization remarks to also inform when
vectorization is possible but not beneficial.
Added tests to exercise some loop remarks.
llvm-svn: 207574
Use -stats to see how many loops were analyzed for possible vectorization and how many of them were actually vectorized.
Patch by Zinovy Nis
Differential Revision: http://reviews.llvm.org/D3438
llvm-svn: 206956
For the purpose of calculating the cost of the loop at various vectorization
factors, we need to count dependencies of consecutive pointers as uniforms
(which means that the VF = 1 cost is used for all overall VF values).
For example, the TSVC benchmark function s173 has:
...
%3 = add nsw i64 %indvars.iv, 16000
%arrayidx8 = getelementptr inbounds %struct.GlobalData* @global_data, i64 0, i32 0, i64 %3
...
and we must realize that the add will be a scalar in order to correctly deduce
it to be profitable to vectorize this on PowerPC with VSX enabled. In fact, all
dependencies of a consecutive pointer must be a scalar (uniform), and so we
simply need to add all consecutive pointers to the worklist that currently
detects collects uniforms.
Fixes PR19296.
llvm-svn: 205387
This provides an initial implementation of getUnrollingPreferences for x86.
getUnrollingPreferences is used by the generic (concatenation) unroller, which
is distinct from the unrolling done by the loop vectorizer. Many modern x86
cores have some kind of uop cache and loop-stream detector (LSD) used to
efficiently dispatch small loops, and taking full advantage of this requires
unrolling small loops (small here means 10s of uops).
These caches also have limits on the number of taken branches in the loop, and
so we also cap the loop unrolling factor based on the maximum "depth" of the
loop. This is currently calculated with a partial DFS traversal (partial
because it will stop early if the path length grows too much). This is still an
approximation, and one that is both conservative (because it does not account
for branches eliminated via block placement) and optimistic (because it is only
recording the maximum depth over minimum paths). Nevertheless, because the
loops that fit in these uop caches are so small, it is not clear how much the
details matter.
The original set of patches posted for review produced the following test-suite
performance results (from the TSVC benchmark) at that time:
ControlLoops-dbl - 13% speedup
ControlLoops-flt - 15% speedup
Reductions-dbl - 7.5% speedup
llvm-svn: 205348
The generic (concatenation) loop unroller is currently placed early in the
standard optimization pipeline. This is a good place to perform full unrolling,
but not the right place to perform partial/runtime unrolling. However, most
targets don't enable partial/runtime unrolling, so this never mattered.
However, even some x86 cores benefit from partial/runtime unrolling of very
small loops, and follow-up commits will enable this. First, we need to move
partial/runtime unrolling late in the optimization pipeline (importantly, this
is after SLP and loop vectorization, as vectorization can drastically change
the size of a loop), while keeping the full unrolling where it is now. This
change does just that.
llvm-svn: 205264
There is no direct AVX instruction to convert to unsigned. I have some ideas
how we may be able to do this with three vector instructions but the current
backend just bails on this to get it scalarized.
See the comment why we need to adjust the cost returned by BasicTTI.
The test is a bit roundabout (and checks assembly rather than bit code) because
I'd like it to work even if at some point we could vectorize this conversion.
Fixes <rdar://problem/16371920>
llvm-svn: 205159
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
Before conditional store vectorization/unrolling we had only one
vectorized/unrolled basic block. After adding support for conditional store
vectorization this will not only be one block but multiple basic blocks. The
last block would have the back-edge. I updated the code to use a vector of basic
blocks instead of a single basic block and fixed the users to use the last entry
in this vector. But, I forgot to add the basic blocks to this vector!
Fixes PR18724.
llvm-svn: 201028
unrolling heuristic per default
Benchmarking on x86_64 (thanks Chandler!) and ARM has shown those options speed
up some benchmarks while not causing any interesting regressions.
llvm-svn: 200621
loop vectorizer to not do so when runtime pointer checks are needed and
share code with the new (not yet enabled) load/store saturation runtime
unrolling. Also ensure that we only consider the runtime checks when the
loop hasn't already been vectorized. If it has, the runtime check cost
has already been paid.
I've fleshed out a test case to cover the scalar unrolling as well as
the vector unrolling and comment clearly why we are or aren't following
the pattern.
llvm-svn: 200530
vectorizer, placing it behind an off-by-default flag.
It turns out that block frequency isn't what we want at all, here or
elsewhere. This has been I think a nagging feeling for several of us
working with it, but Arnold has given some really nice simple examples
where the results are so comprehensively wrong that they aren't useful.
I'm planning to email the dev list with a summary of why its not really
useful and a couple of ideas about how to better structure these types
of heuristics.
llvm-svn: 200294
The vectorizer takes a loop like this and widens all instructions except for the
store. The stores are scalarized/unrolled and hidden behind an "if" block.
for (i = 0; i < 128; ++i) {
if (a[i] < 10)
a[i] += val;
}
for (i = 0; i < 128; i+=2) {
v = a[i:i+1];
v0 = (extract v, 0) + 10;
v1 = (extract v, 1) + 10;
if (v0 < 10)
a[i] = v0;
if (v1 < 10)
a[i] = v1;
}
The vectorizer relies on subsequent optimizations to sink instructions into the
conditional block where they are anticipated.
The flag "vectorize-num-stores-pred" controls whether and how many stores to
handle this way. Vectorization of conditional stores is disabled per default for
now.
This patch also adds a change to the heuristic when the flag
"enable-loadstore-runtime-unroll" is enabled (off by default). It unrolls small
loops until load/store ports are saturated. This heuristic uses TTI's
getMaxUnrollFactor as a measure for load/store ports.
I also added a second flag -enable-cond-stores-vec. It will enable vectorization
of conditional stores. But there is no cost model for vectorization of
conditional stores in place yet so this will not do good at the moment.
rdar://15892953
Results for x86-64 -O3 -mavx +/- -mllvm -enable-loadstore-runtime-unroll
-vectorize-num-stores-pred=1 (before the BFI change):
Performance Regressions:
Benchmarks/Ptrdist/yacr2/yacr2 7.35% (maze3() is identical but 10% slower)
Applications/siod/siod 2.18%
Performance improvements:
mesa -4.42%
libquantum -4.15%
With a patch that slightly changes the register heuristics (by subtracting the
induction variable on both sides of the register pressure equation, as the
induction variable is probably not really unrolled):
Performance Regressions:
Benchmarks/Ptrdist/yacr2/yacr2 7.73%
Applications/siod/siod 1.97%
Performance Improvements:
libquantum -13.05% (we now also unroll quantum_toffoli)
mesa -4.27%
llvm-svn: 200270
cold loops as-if they were being optimized for size.
Nothing fancy here. Simply test case included. The nice thing is that we
can now incrementally build on top of this to drive other heuristics.
All of the infrastructure work is done to get the profile information
into this layer.
The remaining work necessary to make this a fully general purpose loop
unroller for very hot loops is to make it a fully general purpose loop
unroller. Things I know of but am not going to have time to benchmark
and fix in the immediate future:
1) Don't disable the entire pass when the target is lacking vector
registers. This really doesn't make any sense any more.
2) Teach the unroller at least and the vectorizer potentially to handle
non-if-converted loops. This is trivial for the unroller but hard for
the vectorizer.
3) Compute the relative hotness of the loop and thread that down to the
various places that make cost tradeoffs (very likely only the
unroller makes sense here, and then only when dealing with loops that
are small enough for unrolling to not completely blow out the LSD).
I'm still dubious how useful hotness information will be. So far, my
experiments show that if we can get the correct logic for determining
when unrolling actually helps performance, the code size impact is
completely unimportant and we can unroll in all cases. But at least
we'll no longer burn code size on cold code.
One somewhat unrelated idea that I've had forever but not had time to
implement: mark all functions which are only reachable via the global
constructors rigging in the module as optsize. This would also decrease
the impact of any more aggressive heuristics here on code size.
llvm-svn: 200219
powers of two. This is essentially always the correct thing given the
impact on alignment, scaling factors that can be used in addressing
modes, etc. Also, fix the management of the unroll vs. small loop cost
to more accurately model things with this world.
Enhance a test case to actually exercise more of the unroll machinery if
using synthetic constants rather than a specific target model. Before
this change, with the added flags this test will unroll 3 times instead
of either 2 or 4 (the two sensible answers).
While I don't expect this to make a huge difference, if there are lots
of loops sitting right on the edge of hitting the 'small unroll' factor,
they might change behavior. However, I've benchmarked moving the small
loop cost up and down in many various ways and by a huge factor (2x)
without seeing more than 0.2% code size growth. Small adjustments such
as the series that led up here have led to about 1% improvement on some
benchmarks, but it is very close to the noise floor so I mostly checked
that nothing regressed. Let me know if you see bad behavior on other
targets but I don't expect this to be a sufficiently dramatic change to
trigger anything.
llvm-svn: 200213
a reduction.
Really. Under certain circumstances (the use list of an instruction has to be
set up right - hence the extra pass in the test case) we would not recognize
when a value in a potential reduction cycle was used multiple times by the
reduction cycle.
Fixes PR18526.
radar://15851149
llvm-svn: 199570
for (i = 0; i < N; ++i)
A[i * Stride1] += B[i * Stride2];
We take loops like this and check that the symbolic strides 'Strided1/2' are one
and drop to the scalar loop if they are not.
This is currently disabled by default and hidden behind the flag
'enable-mem-access-versioning'.
radar://13075509
llvm-svn: 198950
A phi node operand or an instruction operand could be a constant expression that
can trap (division). Check that we don't vectorize such cases.
PR16729
radar://15653590
llvm-svn: 197449
The intended behaviour is to force vectorization on the presence
of the flag (either turn on or off), and to continue the behaviour
as expected in its absence. Tests were added to make sure the all
cases are covered in opt. No tests were added in other tools with
the assumption that they should use the PassManagerBuilder in the
same way.
This patch also removes the outdated -late-vectorize flag, which was
on by default and not helping much.
The pragma metadata is being attached to the same place as other loop
metadata, but nothing forbids one from attaching it to a function
(to enable #pragma optimize) or basic blocks (to hint the basic-block
vectorizers), etc. The logic should be the same all around.
Patches to Clang to produce the metadata will be produced after the
initial implementation is agreed upon and committed. Patches to other
vectorizers (such as SLP and BB) will be added once we're happy with
the pass manager changes.
llvm-svn: 196537
clang enables vectorization at optimization levels > 1 and size level < 2. opt
should behave similarily.
Loop vectorization and SLP vectorization can be disabled with the flags
-disable-(loop/slp)-vectorization.
llvm-svn: 196294
In signed arithmetic we could end up with an i64 trip count for an i32 phi.
Because it is signed arithmetic we know that this is only defined if the i32
does not wrap. It is therefore safe to truncate the i64 trip count to a i32
value.
Fixes PR18049.
llvm-svn: 195787
We are going to drop debug info without a version number or with a different
version number, to make sure we don't crash when we see bitcode files with
different debug info metadata format.
llvm-svn: 195504
We are slicing an array of Value pointers and process those slices in a loop.
The problem is that we might invalidate a later slice by vectorizing a former
slice.
Use a WeakVH to track the pointer. If the pointer is deleted or RAUW'ed we can
tell.
The test case will only fail when running with libgmalloc.
radar://15498655
llvm-svn: 195162
In some case the loop exit count computation can overflow. Extend the type to
prevent most of those cases.
The problem is loops like:
int main ()
{
int a = 1;
char b = 0;
lbl:
a &= 4;
b--;
if (b) goto lbl;
return a;
}
The backedge count is 255. The induction variable type is i8. If we add one to
255 to get the exit count we overflow to zero.
To work around this issue we extend the type of the induction variable to i32 in
the case of i8 and i16.
PR17532
llvm-svn: 195008
When we vectorize a scalar access with no alignment specified, we have to set
the target's abi alignment of the scalar access on the vectorized access.
Using the same alignment of zero would be wrong because most targets will have a
bigger abi alignment for vector types.
This probably fixes PR17878.
llvm-svn: 194876
When the elements are extracted from a select on vectors
or a vector select, do the select on the extracted scalars
from the input if there is only one use.
llvm-svn: 194013
When the loop vectorizer was part of the SCC inliner pass manager gvn would
run after the loop vectorizer followed by instcombine. This way redundancy
(multiple uses) were removed and instcombine could perform scalarization on the
induction variables. Having moved the loop vectorizer to later we no longer run
any form of redundancy elimination before we perform instcombine. This caused
vectorized induction variables to survive that did not before.
On a recent iMac this helps linpack back from 6000Mflops to 7000Mflops.
This should also help lpbench and paq8p.
I ran a Release (without Asserts) build over the test-suite and did not see any
negative impact on compile time.
radar://15339680
llvm-svn: 193891
When a dependence check fails we can still try to vectorize loops with runtime
array bounds checks.
This helps linpack to vectorize a loop in dgefa. And we are back to 2x of the
scalar performance on a corei7-avx.
radar://15339680
llvm-svn: 193853
Hal Finkel