This patch disables memory-instruction vectorization for types that need padding
bytes, e.g., x86_fp80 has 10 bytes store size with 6 bytes padding in darwin on
x86_64. Because the load/store vectorization is performed by the bit casting to
a packed vector, which has incompatible memory layout due to the lack of padding
bytes, the present vectorizer produces inconsistent result for memory
instructions of those types.
This patch checks an equality of the AllocSize of a scalar type and allocated
size for each vector element, to ensure that there is no padding bytes and the
array can be read/written using vector operations.
Patch by Daisuke Takahashi!
Fixes PR15758.
llvm-svn: 180196
even if erroneously annotated with the parallel loop metadata.
Fixes Bug 15794:
"Loop Vectorizer: Crashes with the use of llvm.loop.parallel metadata"
llvm-svn: 180081
Also make some static function class functions to avoid having to mention the
class namespace for enums all the time.
No functionality change intended.
llvm-svn: 179886
A min/max operation is represented by a select(cmp(lt/le/gt/ge, X, Y), X, Y)
sequence in LLVM. If we see such a sequence we can treat it just as any other
commutative binary instruction and reduce it.
This appears to help bzip2 by about 1.5% on an imac12,2.
radar://12960601
llvm-svn: 179773
This commit adds the infrastructure for performing bottom-up SLP vectorization (and other optimizations) on parallel computations.
The infrastructure has three potential users:
1. The loop vectorizer needs to be able to vectorize AOS data structures such as (sum += A[i] + A[i+1]).
2. The BB-vectorizer needs this infrastructure for bottom-up SLP vectorization, because bottom-up vectorization is faster to compute.
3. A loop-roller needs to be able to analyze consecutive chains and roll them into a loop, in order to reduce code size. A loop roller does not need to create vector instructions, and this infrastructure separates the chain analysis from the vectorization.
This patch also includes a simple (100 LOC) bottom up SLP vectorizer that uses the infrastructure, and can vectorize this code:
void SAXPY(int *x, int *y, int a, int i) {
x[i] = a * x[i] + y[i];
x[i+1] = a * x[i+1] + y[i+1];
x[i+2] = a * x[i+2] + y[i+2];
x[i+3] = a * x[i+3] + y[i+3];
}
llvm-svn: 179117
Pass down the fact that an operand is going to be a vector of constants.
This should bring the performance of MultiSource/Benchmarks/PAQ8p/paq8p on x86
back. It had degraded to scalar performance due to my pervious shift cost change
that made all shifts expensive on x86.
radar://13576547
llvm-svn: 178809
We want vectorization to happen at -g. Ignore calls to the dbg.value intrinsic
and don't transfer them to the vectorized code.
radar://13378964
llvm-svn: 176768
The LoopVectorizer often runs multiple times on the same function due to inlining.
When this happens the loop vectorizer often vectorizes the same loops multiple times, increasing code size and adding unneeded branches.
With this patch, the vectorizer during vectorization puts metadata on scalar loops and marks them as 'already vectorized' so that it knows to ignore them when it sees them a second time.
PR14448.
llvm-svn: 176399
This properly asks TargetLibraryInfo if a call is available and if it is, it
can be translated into the corresponding LLVM builtin. We don't vectorize sqrt()
yet because I'm not sure about the semantics for negative numbers. The other
intrinsic should be exact equivalents to the libm functions.
Differential Revision: http://llvm-reviews.chandlerc.com/D465
llvm-svn: 176188
Storing the load/store instructions with the values
and inspect them using Alias Analysis to make sure
they don't alias, since the GEP pointer operand doesn't
take the offset into account.
Trying hard to not add any extra cost to loads and stores
that don't overlap on global values, AA is *only* calculated
if all of the previous attempts failed.
Using biggest vector register size as the stride for the
vectorization access, as we're being conservative and
the cost model (which calculates the real vectorization
factor) is only run after the legalization phase.
We might re-think this relationship in the future, but
for now, I'd rather be safe than sorry.
llvm-svn: 175818
This fixes PR15289. This bug was introduced (recently) in r175215; collecting
all std::vector references for candidate pairs to delete at once is invalid
because subsequent lookups in the owning DenseMap could invalidate the
references.
bugpoint was able to reduce a useful test case. Unfortunately, because whether
or not this asserts depends on memory layout, this test case will sometimes
appear to produce valid output. Nevertheless, running under valgrind will
reveal the error.
llvm-svn: 175397
Several functions and variable names used the term 'tree' to refer
to what is actually a DAG. Correcting this mistake will, hopefully,
prevent confusion in the future.
No functionality change intended.
llvm-svn: 175278
For some basic blocks, it is possible to generate many candidate pairs for
relatively few pairable instructions. When many (tens of thousands) of these pairs
are generated for a single instruction group, the time taken to generate and
rank the different vectorization plans can become quite large. As a result, we now
cap the number of candidate pairs within each instruction group. This is done by
closing out the group once the threshold is reached (set now at 3000 pairs).
Although this will limit the overall compile-time impact, this may not be the best
way to achieve this result. It might be better, for example, to prune excessive
candidate pairs after the fact the prevent the generation of short, but highly-connected
groups. We can experiment with this in the future.
This change reduces the overall compile-time slowdown of the csa.ll test case in
PR15222 to ~5x. If 5x is still considered too large, a lower limit can be
used as the default.
This represents a functionality change, but only for very large inputs
(thus, there is no regression test).
llvm-svn: 175251
All instances of std::multimap have now been replaced by
DenseMap<K, std::vector<V> >, and this yields a speedup of 5% on the
csa.ll test case from PR15222.
No functionality change intended.
llvm-svn: 175216
This is another commit on the road to removing std::multimap from
BBVectorize. This gives an ~1% speedup on the csa.ll test case
in PR15222.
No functionality change intended.
llvm-svn: 175215
When building the pairable-instruction dependency map, don't search
past the last pairable instruction. For large blocks that have been
divided into multiple instruction groups, searching past the last
instruction in each group is very wasteful. This gives a 32% speedup
on the csa.ll test case from PR15222 (when using 50 instructions
in each group).
No functionality change intended.
llvm-svn: 174915
This map is queried only for instructions in pairs of pairable
instructions; so make sure that only pairs of pairable
instructions are added to the map. This gives a 3.5% speedup
on the csa.ll test case from PR15222.
No functionality change intended.
llvm-svn: 174914
This eliminates one more linear search over a range of
std::multimap entries. This gives a 22% speedup on the
csa.ll test case from PR15222.
No functionality change intended.
llvm-svn: 174893
This removes the last of the linear searches over ranges of std::multimap
iterators, giving a 7% speedup on the doduc.bc input from PR15222.
No functionality change intended.
llvm-svn: 174859
Profiling suggests that getInstructionTypes is performance-sensitive,
this cleans up some double-casting in that function in favor of
using dyn_cast.
No functionality change intended.
llvm-svn: 174857
By itself, this does not have much of an effect, but only because in the default
configuration the full cycle checks are used only for small problem sizes.
This is part of a general cleanup of uses of iteration over std::multimap
ranges only for the purpose of checking membership.
No functionality change intended.
llvm-svn: 174856
This is a follow-up to the cost-model change in r174713 which splits
the cost of a memory operation between the address computation and the
actual memory access. In r174713, this cost is always added to the
memory operation cost, and so BBVectorize will do the same.
Currently, this new cost function is used only by ARM, and I don't
have any ARM test cases for BBVectorize. Assistance in generating some
good ARM test cases for BBVectorize would be greatly appreciated!
llvm-svn: 174743
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
We don't want too many classes in a pass and the classes obscure the details. I
was going a little overboard with object modeling here. Replace classes by
generic code that handles both loads and stores.
No functionality change intended.
llvm-svn: 174646
In the loop vectorizer cost model, we used to ignore stores/loads of a pointer
type when computing the widest type within a loop. This meant that if we had
only stores/loads of pointers in a loop we would return a widest type of 8bits
(instead of 32 or 64 bit) and therefore a vector factor that was too big.
Now, if we see a consecutive store/load of pointers we use the size of a pointer
(from data layout).
This problem occured in SingleSource/Benchmarks/Shootout-C++/hash.cpp (reduced
test case is the first test in vector_ptr_load_store.ll).
radar://13139343
llvm-svn: 174377
When flipping the pair of subvectors that form a vector, if the
vector length is 2, we can use the SK_Reverse shuffle kind to get
more-accurate cost information. Also we can use the SK_ExtractSubvector
shuffle kind to get accurate subvector extraction costs.
The current cost model implementations don't yet seem complex enough
for this to make a difference (thus, there are no test cases with this
commit), but it should help in future.
Depending on how the various targets optimize and combine shuffles in
practice, we might be able to get more-accurate costs by combining the
costs of multiple shuffle kinds. For example, the cost of flipping the
subvector pairs could be modeled as two extractions and two subvector
insertions. These changes, however, should probably be motivated
by specific test cases.
llvm-svn: 173621
We ignore the cpu frontend and focus on pipeline utilization. We do this because we
don't have a good way to estimate the loop body size at the IR level.
llvm-svn: 172964
This separates the check for "too few elements to run the vector loop" from the
"memory overlap" check, giving a lot nicer code and allowing to skip the memory
checks when we're not going to execute the vector code anyways. We still leave
the decision of whether to emit the memory checks as branches or setccs, but it
seems to be doing a good job. If ugly code pops up we may want to emit them as
separate blocks too. Small speedup on MultiSource/Benchmarks/MallocBench/espresso.
Most of this is legwork to allow multiple bypass blocks while updating PHIs,
dominators and loop info.
llvm-svn: 172902
We don't have a detailed analysis on which values are vectorized and which stay scalars in the vectorized loop so we use
another method. We look at reduction variables, loads and stores, which are the only ways to get information in and out
of loop iterations. If the data types are extended and truncated then the cost model will catch the cost of the vector
zext/sext/trunc operations.
llvm-svn: 172178
small loops. On small loops post-loop that handles scalars (and runs slower) can take more time to execute than the
rest of the loop. This patch disables widening of loops with a small static trip count.
llvm-svn: 171798
being present. Make a member of one of the helper classes a reference as
part of this.
Reformatting goodness brought to you by clang-format.
llvm-svn: 171726
This makes the loop vectorizer match the pattern followed by roughly all
other passses. =]
Notably, this header file was braken in several regards: it contained
a using namespace directive, global #define's that aren't globaly
appropriate, and global constants defined directly in the header file.
As a side benefit, lots of the types in this file become internal, which
will cause the optimizer to chew on this pass more effectively.
llvm-svn: 171723
This could be simplified further, but Hal has a specific feature for
ignoring TTI, and so I preserved that.
Also, I needed to use it because a number of tests fail when switching
from a null TTI to the NoTTI nonce implementation. That seems suspicious
to me and so may be something that you need to look into Hal. I worked
it by preserving the old behavior for these tests with the flag that
ignores all target info.
llvm-svn: 171722
this patch brought to you by the tool clang-format.
I wanted to fix up the names of constructor parameters because they
followed a bit of an anti-pattern by naming initialisms with CamelCase:
'Tti', 'Se', etc. This appears to have been in an attempt to not overlap
with the names of member variables 'TTI', 'SE', etc. However,
constructor arguments can very safely alias members, and in fact that's
the conventional way to pass in members. I've fixed all of these I saw,
along with making some strang abbreviations such as 'Lp' be simpler 'L',
or 'Lgl' be the word 'Legal'.
However, the code I was touching had indentation and formatting somewhat
all over the map. So I ran clang-format and fixed them.
I also fixed a few other formatting or doxygen formatting issues such as
using ///< on trailing comments so they are associated with the correct
entry.
There is still a lot of room for improvement of the formating and
cleanliness of this code. ;] At least a few parts of the coding
standards or common practices in LLVM's code aren't followed, the enum
naming rules jumped out at me. I may mix some of these while I'm here,
but not all of them.
llvm-svn: 171719
1. Add code to estimate register pressure.
2. Add code to select the unroll factor based on register pressure.
3. Add bits to TargetTransformInfo to provide the number of registers.
llvm-svn: 171469
into their new header subdirectory: include/llvm/IR. This matches the
directory structure of lib, and begins to correct a long standing point
of file layout clutter in LLVM.
There are still more header files to move here, but I wanted to handle
them in separate commits to make tracking what files make sense at each
layer easier.
The only really questionable files here are the target intrinsic
tablegen files. But that's a battle I'd rather not fight today.
I've updated both CMake and Makefile build systems (I think, and my
tests think, but I may have missed something).
I've also re-sorted the includes throughout the project. I'll be
committing updates to Clang, DragonEgg, and Polly momentarily.
llvm-svn: 171366
directly.
This is in preparation for removing the use of the 'Attribute' class as a
collection of attributes. That will shift to the AttributeSet class instead.
llvm-svn: 171253
LCSSA PHIs may have undef values. The vectorizer updates values that are used by outside users such as PHIs.
The bug happened because undefs are not loop values. This patch handles these PHIs.
PR14725
llvm-svn: 171251
For the time being this includes only some dummy test cases. Once the
generic implementation of the intrinsics cost function does something other
than assuming scalarization in all cases, or some target specializes the
interface, some real test cases can be added.
Also, for consistency, I changed the type of IID from unsigned to Intrinsic::ID
in a few other places.
llvm-svn: 171079
memory bound checks. Before the fix we were able to vectorize this loop from
the Livermore Loops benchmark:
for ( k=1 ; k<n ; k++ )
x[k] = x[k-1] + y[k];
llvm-svn: 170811
Before if-conversion we could check if a value is loop invariant
if it was declared inside the basic block. Now that loops have
multiple blocks this check is incorrect.
This fixes External/SPEC/CINT95/099_go/099_go
llvm-svn: 170756
MapVector is a bit heavyweight, but I don't see a simpler way. Also the
InductionList is unlikely to be large. This should help 3-stage selfhost
compares (PR14647).
llvm-svn: 170528
- added function to VectorTargetTransformInfo to query cost of intrinsics
- vectorize trivially vectorizable intrinsic calls such as sin, cos, log, etc.
Reviewed by: Nadav
llvm-svn: 169711
Added the code that actually performs the if-conversion during vectorization.
We can now vectorize this code:
for (int i=0; i<n; ++i) {
unsigned k = 0;
if (a[i] > b[i]) <------ IF inside the loop.
k = k * 5 + 3;
a[i] = k; <---- K is a phi node that becomes vector-select.
}
llvm-svn: 169217
which is the legality of the if-conversion transformation. The next step is to
implement the cost-model for the if-converted code as well as the
vectorization itself.
llvm-svn: 169152
Sooooo many of these had incorrect or strange main module includes.
I have manually inspected all of these, and fixed the main module
include to be the nearest plausible thing I could find. If you own or
care about any of these source files, I encourage you to take some time
and check that these edits were sensible. I can't have broken anything
(I strictly added headers, and reordered them, never removed), but they
may not be the headers you'd really like to identify as containing the
API being implemented.
Many forward declarations and missing includes were added to a header
files to allow them to parse cleanly when included first. The main
module rule does in fact have its merits. =]
llvm-svn: 169131
For now, this uses 8 on-stack elements. I'll need to do some profiling
to see if this is the best number.
Pointed out by Jakob in post-commit review.
llvm-svn: 167966
Iterating over the children of each node in the potential vectorization
plan must happen in a deterministic order (because it affects which children
are erased when two children conflict). There was no need for this data
structure to be a map in the first place, so replacing it with a vector
is a small change.
I believe that this was the last remaining instance if iterating over the
elements of a Dense* container where the iteration order could matter.
There are some remaining iterations over std::*map containers where the order
might matter, but so long as the Value* for instructions in a block increase
with the order of the instructions in the block (or decrease) monotonically,
then this will appear to be deterministic.
llvm-svn: 167942
Don't choose a vectorization plan containing only shuffles and
vector inserts/extracts. Due to inperfections in the cost model,
these can lead to infinite recusion.
llvm-svn: 167811
This fixes another infinite recursion case when using target costs.
We can only replace insert element input chains that are pure (end
with inserting into an undef).
llvm-svn: 167784
The old checking code, which assumed that input shuffles and insert-elements
could always be folded (and thus were free) is too simple.
This can only happen in special circumstances.
Using the simple check caused infinite recursion.
llvm-svn: 167750
The pass would previously assert when trying to compute the cost of
compare instructions with illegal vector types (like struct pointers).
llvm-svn: 167743
This fixes a bug where shuffles were being fused such that the
resulting input types were not legal on the target. This would
occur only when both inputs and dependencies were also foldable
operations (such as other shuffles) and there were other connected
pairs in the same block.
llvm-svn: 167731
When target cost information is available, compute explicit costs of inserting and
extracting values from vectors. At this point, all costs are estimated using the
target information, and the chain-depth heuristic is not needed. As a result, it is now, by
default, disabled when using target costs.
llvm-svn: 167256
When target costs are available, use them to account for the costs of
shuffles on internal edges of the DAG of candidate pairs.
Because the shuffle costs here are currently for only the internal edges,
the current target cost model is trivial, and the chain depth requirement
is still in place, I don't yet have an easy test
case. Nevertheless, by looking at the debug output, it does seem to do the right
think to the effective "size" of each DAG of candidate pairs.
llvm-svn: 167217
BBVectorize would, except for loads and stores, always fuse instructions
so that the first instruction (in the current source order) would always
represent the low part of the input vectors and the second instruction
would always represent the high part. This lead to too many shuffles
being produced because sometimes the opposite order produces fewer of them.
With this change, BBVectorize tracks the kind of pair connections that form
the DAG of candidate pairs, and uses that information to reorder the pairs to
avoid excess shuffles. Using this information, a future commit will be able
to add VTTI-based shuffle costs to the pair selection procedure. Importantly,
the number of remaining shuffles can now be estimated during pair selection.
There are some trivial instruction reorderings in the test cases, and one
simple additional test where we certainly want to do a reordering to
avoid an unnecessary shuffle.
llvm-svn: 167122
Instead of recomputing relative pointer information just prior to fusing,
cache this information (which also needs to be computed during the
candidate-pair selection process). This cuts down on the total number of
SE queries made, and also is a necessary intermediate step on the road toward
including shuffle costs in the pair selection procedure.
No functionality change is intended.
llvm-svn: 167049
Stop propagating the FlipMemInputs variable into the routines that
create the replacement instructions. Instead, just flip the arguments
of those routines. This allows for some associated cleanup (not all
of which is done here). No functionality change is intended.
llvm-svn: 167042
SE was being called during the instruction-fusion process (when the result
is unreliable, and thus ignored). No functionality change is intended.
llvm-svn: 167037
The monolithic interface for instruction costs has been split into
several functions. This is the corresponding change. No functionality
change is intended.
llvm-svn: 166865
Add getCostXXX calls for different families of opcodes, such as casts, arithmetic, cmp, etc.
Port the LoopVectorizer to the new API.
The LoopVectorizer now finds instructions which will remain uniform after vectorization. It uses this information when calculating the cost of these instructions.
llvm-svn: 166836
This is needed so that perl's SHA can be compiled (otherwise
BBVectorize takes far too long to find its fixed point).
I'll try to come up with a reduced test case.
llvm-svn: 166738
This is the first of several steps to incorporate information from the new
TargetTransformInfo infrastructure into BBVectorize. Two things are done here:
1. Target information is used to determine if it is profitable to fuse two
instructions. This means that the cost of the vector operation must not
be more expensive than the cost of the two original operations. Pairs that
are not profitable are no longer considered (because current cost information
is incomplete, for intrinsics for example, equal-cost pairs are still
considered).
2. The 'cost savings' computed for the profitability check are also used to
rank the DAGs that represent the potential vectorization plans. Specifically,
for nodes of non-trivial depth, the cost savings is used as the node
weight.
The next step will be to incorporate the shuffle costs into the DAG weighting;
this will give the edges of the DAG weights as well. Once that is done, when
target information is available, we should be able to dispense with the
depth heuristic.
llvm-svn: 166716
Unreachable blocks can have invalid instructions. For example,
jump threading can produce self-referential instructions in
unreachable blocks. Also, we should not be spending time
optimizing unreachable code. Fixes PR14133.
llvm-svn: 166423
We used a SCEV to detect that A[X] is consecutive. We assumed that X was
the induction variable. But X can be any expression that uses the induction
for example: X = i + 2;
llvm-svn: 166388
This is important for nested-loop reductions such as :
In the innermost loop, the induction variable does not start with zero:
for (i = 0 .. n)
for (j = 0 .. m)
sum += ...
llvm-svn: 166387
If the pointer is consecutive then it is safe to read and write. If the pointer is non-loop-consecutive then
it is unsafe to vectorize it because we may hit an ordering issue.
llvm-svn: 166371
