The motivation for this change is a workload generated by the XLA compiler
targeting nvidia GPUs.
This kernel has a few hundred i8 loads and stores. Merging is critical for
performance.
The current LSV doesn't merge these well because it only considers instructions
within a block of 64 loads+stores. This limit is necessary to contain the
O(n^2) behavior of the pass. I'm hesitant to increase the limit, because this
pass is already one of the slowest parts of compiling an XLA program.
So we rewrite basically the whole thing to use a new algorithm. Before, we
compared every load/store to every other to see if they're consecutive. The
insight (from tra@) is that this is redundant. If we know the offset from PtrA
to PtrB, then we don't need to compare PtrC to both of them in order to tell
whether C may be adjacent to A or B.
So that's what we do. When scanning a basic block, we maintain a list of
chains, where we know the offset from every element in the chain to the first
element in the chain. Each instruction gets compared only to the leaders of
all the chains.
In the worst case, this is still O(n^2), because all chains might be of length
1. To prevent compile time blowup, we only consider the 64 most recently used
chains. Thus we do no more comparisons than before, but we have the potential
to make much longer chains.
This rewrite affects many tests. The changes to tests fall into two
categories.
1. The old code had what appears to be a bug when deciding whether a misaligned
vectorized load is fast. Suppose TTI reports that load <i32 x 4> align 4
has relative speed 1, and suppose that load i32 align 4 has relative speed
32.
The intent of the code seems to be that we prefer the scalar load, because
it's faster. But the old code would choose the vectorized load.
accessIsMisaligned would set RelativeSpeed to 0 for the scalar load (and not
even call into TTI to get the relative speed), because the scalar load is
aligned.
After this patch, we will prefer the scalar load if it's faster.
2. This patch changes the logic for how we vectorize. Usually this results in
vectorizing more.
Explanation of changes to tests:
- AMDGPU/adjust-alloca-alignment.ll: #1
- AMDGPU/flat_atomic.ll: #2, we vectorize more.
- AMDGPU/int_sideeffect.ll: #2, there are two possible locations for the call to @foo, and the pass is brittle to this. Before, we'd vectorize in case 1 and not case 2. Now we vectorize in case 2 and not case 1. So we just move the call.
- AMDGPU/adjust-alloca-alignment.ll: #2, we vectorize more
- AMDGPU/insertion-point.ll: #2 we vectorize more
- AMDGPU/merge-stores-private.ll: #1 (undoes changes from git rev 86f9117d476, which appear to have hit the bug from #1)
- AMDGPU/multiple_tails.ll: #1
- AMDGPU/vect-ptr-ptr-size-mismatch.ll: Fix alignment (I think related to #1 above).
- AMDGPU CodeGen: I have difficulty commenting on these changes, but many of them look like #2, we vectorize more.
- NVPTX/4x2xhalf.ll: Fix alignment (I think related to #1 above).
- NVPTX/vectorize_i8.ll: We don't generate <3 x i8> vectors on NVPTX because they're not legal (and eventually get split)
- X86/correct-order.ll: #2, we vectorize more, probably because of changes to the chain-splitting logic.
- X86/subchain-interleaved.ll: #2, we vectorize more
- X86/vector-scalar.ll: #2, we can now vectorize scalar float + <1 x float>
- X86/vectorize-i8-nested-add-inseltpoison.ll: Deleted the nuw test because it was nonsensical. It was doing `add nuw %v0, -1`, but this is equivalent to `add nuw %v0, 0xffff'ffff`, which is equivalent to asserting that %v0 == 0.
- X86/vectorize-i8-nested-add.ll: Same as nested-add-inseltpoison.ll
Differential Revision: https://reviews.llvm.org/D149893
This is a follow-up to b71edfaa4ec3c998aadb35255ce2f60bba2940b0
since I forgot the lit.local.cfg files in that one.
Reformatting is done with `black`.
If you end up having problems merging this commit because you
have made changes to a python file, the best way to handle that
is to run git checkout --ours <yourfile> and then reformat it
with black.
If you run into any problems, post to discourse about it and
we will try to help.
RFC Thread below:
https://discourse.llvm.org/t/rfc-document-and-standardize-python-code-style
Reviewed By: barannikov88, kwk
Differential Revision: https://reviews.llvm.org/D150762
Over the past day or so, i've took a large swing at our tests,
and reduced the number of tests that were still using the old syntax
from ~1800 to just 200.
Left to handle: (as it is seen in this patch)
* Transforms/LSR
* Transforms/CGP
* Transforms/TypePromotion
* Transforms/HardwareLoops
* Analysis/*
* some misc.
I think this is the right point to start actively refusing
to honor the old syntax, except for the old tests,
to prevent the old syntax from creeping back in.
Thus, let's add temporary default-off flag,
and if it is not passed refuse to accept old syntax.
The tests that still need porting are annotated with this flag.
Reviewed By: aeubanks
Differential Revision: https://reviews.llvm.org/D139647
Currently, opaque pointers are supported in two forms: The
-force-opaque-pointers mode, where all pointers are opaque and
typed pointers do not exist. And as a simple ptr type that can
coexist with typed pointers.
This patch removes support for the mixed mode. You either get
typed pointers, or you get opaque pointers, but not both. In the
(current) default mode, using ptr is forbidden. In -opaque-pointers
mode, all pointers are opaque.
The motivation here is that the mixed mode introduces additional
issues that don't exist in fully opaque mode. D105155 is an example
of a design problem. Looking at D109259, it would probably need
additional work to support mixed mode (e.g. to generate GEPs for
typed base but opaque result). Mixed mode will also end up
inserting many casts between i8* and ptr, which would require
significant additional work to consistently avoid.
I don't think the mixed mode is particularly valuable, as it
doesn't align with our end goal. The only thing I've found it to
be moderately useful for is adding some opaque pointer tests in
between typed pointer tests, but I think we can live without that.
Differential Revision: https://reviews.llvm.org/D109290
First we refactor the code which does no wrapping add sequences
match: we need to allow different operand orders for
the key add instructions involved in the match.
Then we use the refactored code trying 4 variants of matching operands.
Originally the code relied on the fact that the matching operands
of the two last add instructions of memory index calculations
had the same LHS argument. But which operand is the same
in the two instructions is actually not essential, so now we allow
that to be any of LHS or RHS of each of the two instructions.
This increases the chances of vectorization to happen.
Reviewed By: volkan
Differential Revision: https://reviews.llvm.org/D103912
This change enables cases for which the index value for the first
load/store instruction in a pair could be a function argument. This
allows using llvm.assume to provide known bits information in such
cases.
Patch by Viacheslav Nikolaev. Thanks!
Differential Revision: https://reviews.llvm.org/D101680
This commit copies existing tests at llvm/Transforms and replaces
'insertelement undef' in those files with 'insertelement poison'.
(see https://reviews.llvm.org/D93586)
Tests listed using this script:
grep -R -E '^[^;]*insertelement <.*> undef,' . | cut -d":" -f1 | uniq |
wc -l
Tests updated:
file_org=llvm/test/Transforms/$1
file=${file_org%.ll}-inseltpoison.ll
cp $file_org $file
sed -i -E 's/^([^;]*)insertelement <(.*)> undef/\1insertelement <\2> poison/g' $file
head -1 $file | grep "Assertions have been autogenerated by utils/update_test_checks.py" -q
if [ "$?" == 1 ]; then
echo "$file : should be manually updated"
# I manually updated the script
exit 1
fi
python3 ./llvm/utils/update_test_checks.py --opt-binary=./build-releaseassert/bin/opt $file
Summary:
Currently when --passes is used, any passes specified via -foo are
ignored. Explicitly bail out when that happens.
This requires changing some tests. Most were straightforward, but
codegenprepare-produced-address-math.ll is tricky. One of its RUNs runs
CodeGenPrepare. I tried porting CodeGenPrepare to the NPM, but ended up
getting stuck when I needed a TargetMachine. NPM doesn't have support
for MachineFunctions yet. So I just deleted that RUN line, since it was
mass-added in https://reviews.llvm.org/D54848 and is likely not that
useful.
Reviewers: echristo, hans
Subscribers: llvm-commits
Tags: #llvm
Differential Revision: https://reviews.llvm.org/D82271
If both OpA and OpB is an add with NSW/NUW and with the same LHS operand,
we can guarantee that the transformation is safe if we can prove that OpA
won't overflow when IdxDiff added to the RHS of OpA.
Review: https://reviews.llvm.org/D79817
As it's causing some bot failures (and per request from kbarton).
This reverts commit r358543/ab70da07286e618016e78247e4a24fcb84077fda.
llvm-svn: 358546
This reverts r319889.
Unfortunately, wrapping flags are not a part of SCEV's identity (they
do not participate in computing a hash value or in equality
comparisons) and in fact they could be assigned after the fact w/o
rebuilding a SCEV.
Grep for const_cast's to see quite a few of examples, apparently all
for AddRec's at the moment.
So, if 2 expressions get built in 2 slightly different ways: one with
flags set in the beginning, the other with the flags attached later
on, we may end up with 2 expressions which are exactly the same but
have their operands swapped in one of the commutative N-ary
expressions, and at least one of them will have "sorted by complexity"
invariant broken.
2 identical SCEV's won't compare equal by pointer comparison as they
are supposed to.
A real-world reproducer is added as a regression test: the issue
described causes 2 identical SCEV expressions to have different order
of operands and therefore compare not equal, which in its turn
prevents LoadStoreVectorizer from vectorizing a pair of consecutive
loads.
On a larger example (the source of the test attached, which is a
bugpoint) I have seen even weirder behavior: adding a constant to an
existing SCEV changes the order of the existing terms, for instance,
getAddExpr(1, ((A * B) + (C * D))) returns (1 + (C * D) + (A * B)).
Differential Revision: https://reviews.llvm.org/D40645
llvm-svn: 340777
if the top level addition in (D + (C-D + x + ...)) could be proven to
not wrap, where the choice of D also maximizes the number of trailing
zeroes of (C-D + x + ...), ensuring homogeneous behaviour of the
transformation and better canonicalization of such expressions.
This enables better canonicalization of expressions like
1 + zext(5 + 20 * %x + 24 * %y) and
zext(6 + 20 * %x + 24 * %y)
which get both transformed to
2 + zext(4 + 20 * %x + 24 * %y)
This pattern is common in address arithmetics and the transformation
makes it easier for passes like LoadStoreVectorizer to prove that 2 or
more memory accesses are consecutive and optimize (vectorize) them.
Reviewed By: mzolotukhin
Differential Revision: https://reviews.llvm.org/D48853
llvm-svn: 337859
The LoadStoreVectorizer thought that <1 x T> and T were the same types
when merging stores, leading to a crash later.
Patch by Erik Hogeman.
Differential Revision: https://reviews.llvm.org/D44014
llvm-svn: 326884
There are cases when we have to merge TBAA access tags with the
same base access type, but different final access types. For
example, accesses to different members of the same structure may
be vectorized into a single load or store instruction. Since we
currently assume that the tags to merge always share the same
final access type, we incorrectly return a tag that describes an
access to one of the original final access types as the generic
tag. This patch fixes that by producing generic tags for the
common type and not the final access types of the original tags.
Resolves:
PR35225: Wrong tbaa metadata after load store vectorizer due to
recent change
https://bugs.llvm.org/show_bug.cgi?id=35225
Differential Revision: https://reviews.llvm.org/D39732
llvm-svn: 317682
Summary:
The code comments indicate that no effort has been spent on
handling load/stores when the size isn't a multiple of the
byte size correctly. However, the code only avoided types
smaller than 8 bits. So for example a load of an i28 could
still be considered as a candidate for vectorization.
This patch adjusts the code to behave according to the code
comment.
The test case used to hit the following assert when
trying to use "cast" an i32 to i28 using CreateBitOrPointerCast:
opt: ../lib/IR/Instructions.cpp:2565: Assertion `castIsValid(op, S, Ty) && "Invalid cast!"' failed.
#0 PrintStackTraceSignalHandler(void*)
#1 SignalHandler(int)
#2 __restore_rt
#3 __GI_raise
#4 __GI_abort
#5 __GI___assert_fail
#6 llvm::CastInst::Create(llvm::Instruction::CastOps, llvm::Value*, llvm::Type*, llvm::Twine const&, llvm::Instruction*)
#7 llvm::IRBuilder<llvm::ConstantFolder, llvm::IRBuilderDefaultInserter>::CreateBitOrPointerCast(llvm::Value*, llvm::Type*, llvm::Twine const&)
#8 (anonymous namespace)::Vectorizer::vectorizeLoadChain(llvm::ArrayRef<llvm::Instruction*>, llvm::SmallPtrSet<llvm::Instruction*, 16u>*)
Reviewers: arsenm
Reviewed By: arsenm
Subscribers: llvm-commits
Differential Revision: https://reviews.llvm.org/D39295
llvm-svn: 316663
Summary:
LSV wants to know the maximum size that can be loaded to a vector register.
On X86, this always matches the maximum register width. Implement this
accordingly and add a test to make sure that LSV can vectorize up to the
maximum permissible width on X86.
Reviewers: delena, arsenm
Reviewed By: arsenm
Subscribers: wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D31504
llvm-svn: 299589
Summary:
LSV was using two vector sets (heads and tails) to track pairs of adjiacent position to vectorize.
A recent optimization is trying to obtain the longest chain to vectorize and assumes the positions
in heads(H) and tails(T) match, which is not the case is there are multiple tails for the same head.
e.g.:
i1: store a[0]
i2: store a[1]
i3: store a[1]
Leads to:
H: i1
T: i2 i3
Instead of:
H: i1 i1
T: i2 i3
So the positions for instructions that follow i3 will have different indexes in H/T.
This patch resolves PR29148.
This issue also surfaced the fact that if the chain is too long, and TLI
returns a "not-fast" answer, the whole chain will be abandoned for
vectorization, even though a smaller one would be beneficial.
Added a testcase and FIXME for this.
Reviewers: tstellarAMD, arsenm, jlebar
Subscribers: mzolotukhin, wdng, llvm-commits
Differential Revision: https://reviews.llvm.org/D24057
llvm-svn: 280179
Summary:
TargetBaseAlign is no longer required since LSV checks if target allows misaligned accesses.
A constant defining a base alignment is still needed for stack accesses where alignment can be adjusted.
Previous patch (D22936) was reverted because tests were failing. This patch also fixes the cause of those failures:
- x86 failing tests either did not have the right target, or the right alignment.
- NVPTX failing tests did not have the right alignment.
- AMDGPU failing test (merge-stores) should allow vectorization with the given alignment but the target info
considers <3xi32> a non-standard type and gives up early. This patch removes the condition and only checks
for a maximum size allowed and relies on the next condition checking for %4 for correctness.
This should be revisited to include 3xi32 as a MVT type (on arsenm's non-immediate todo list).
Note that checking the sizeInBits for a MVT is undefined (leads to an assertion failure),
so we need to create an EVT, hence the interface change in allowsMisaligned to include the Context.
Reviewers: arsenm, jlebar, tstellarAMD
Subscribers: jholewinski, arsenm, mzolotukhin, llvm-commits
Differential Revision: https://reviews.llvm.org/D23068
llvm-svn: 277735
Summary:
This helps keep us honest -- there were a number of ways we could screw
up and still have passed this test.
Reviewers: asbirlea
Subscribers: llvm-commits, arsenm
Differential Revision: https://reviews.llvm.org/D22531
llvm-svn: 276053
Summary:
LSV used to abort vectorizing a chain for interleaved load/store accesses that alias.
Allow a valid prefix of the chain to be vectorized, mark just the prefix and retry vectorizing the remaining chain.
Reviewers: llvm-commits, jlebar, arsenm
Subscribers: mzolotukhin
Differential Revision: http://reviews.llvm.org/D22119
llvm-svn: 275317
Summary:
Aiming to correct the ordering of loads/stores. This patch changes the
insert point for loads to the position of the first load.
It updates the ordering method for loads to insert before, rather than after.
Before this patch the following sequence:
"load a[1], store a[1], store a[0], load a[2]"
Would incorrectly vectorize to "store a[0,1], load a[1,2]".
The correctness check was assuming the insertion point for loads is at
the position of the first load, when in practice it was at the last
load. An alternative fix would have been to invert the correctness check.
The current fix changes insert position but also requires reordering of
instructions before the vectorized load.
Updated testcases to reflect the changes.
Reviewers: tstellarAMD, llvm-commits, jlebar, arsenm
Subscribers: mzolotukhin
Differential Revision: http://reviews.llvm.org/D22071
llvm-svn: 275117
Summary:
GetBoundryInstruction returns the last instruction as the instruction which follows or end(). Otherwise the last instruction in the boundry set is not being tested by isVectorizable().
Partially solve reordering of instructions. More extensive solution to follow.
Reviewers: tstellarAMD, llvm-commits, jlebar
Subscribers: escha, arsenm, mzolotukhin
Differential Revision: http://reviews.llvm.org/D21934
llvm-svn: 274389
