LoadStoreVectorizer hangs on certain examples, when its reorder function
goes into a cycle. Detect this cycle and explicitly forbid it, using an
assert, and document the resulting crash in a test-case under AArch64.
Vectors are always bit-packed and don't respect the elements' alignment
requirements. This is different from arrays. This means offsets of
vector GEPs need to be computed differently than offsets of array GEPs.
This PR fixes many places that rely on an incorrect pattern
that always relies on `DL.getTypeAllocSize(GTI.getIndexedType())`.
We replace these by usages of `GTI.getSequentialElementStride(DL)`,
which is a new helper function added in this PR.
This changes behavior for GEPs into vectors with element types for which
the (bit) size and alloc size is different. This includes two cases:
* Types with a bit size that is not a multiple of a byte, e.g. i1.
GEPs into such vectors are questionable to begin with, as some elements
are not even addressable.
* Overaligned types, e.g. i16 with 32-bit alignment.
Existing tests are unaffected, but a miscompilation of a new test is fixed.
---------
Co-authored-by: Nikita Popov <github@npopov.com>
This reverts commit b5743d4798b250506965e07ebab806a3c2d767cc.
This causes some minor compile-time impact. Revert for now, better
to do the change more gradually.
Swap order for checking address space and the strip pointer cast
when analyzing if a load/store is accessing an alloca. This to
make sure we do the cheaper check first.
This is done as a follow up to D152386.
In commit 2be0abb7fe72ed453 (D149893) the load store vectorized was
reimplemented. One thing that can happen with the new LSV is that
it can increase the align of alloca and global objects. However,
the code comments indicate that the intention only was to increase
alignment of alloca.
Now we will use stripPointerCasts to analyse if the load/store really
is accessing an alloca (same as getOrEnforceKnownAlignment is using).
And then we only try to change the align if we find an alloca
instruction. This way the code will match better with code comments,
and we won't change alignment of non-stack variables to use the
"StackAdjustedAlignment".
Differential Revision: https://reviews.llvm.org/D152386
Previously, getConstantOffset could return an APInt with a different
bitwidth than the input pointers. For example, we might be loading an
opaque 64-bit pointer, but stripAndAccumulateInBoundsConstantOffsets
might give a 32-bit offset.
This was OK in most cases because in gatherChains, we casted the APInt
back to the original ASPtrBits.
But it was not OK when considering selects. We'd call getConstantOffset
twice and compare the resulting APInt's, which might not have the same
bit width.
This fixes that. Now getConstantOffset always returns offsets with the
correct width, so we don't need the hack of casting it in gatherChains,
and it works correctly when we're handling selects.
Differential Revision: https://reviews.llvm.org/D151640
Previously we used the later of GEPA or GEPB. This is hacky because
really we should be using the later of the two load/store instructions
being considered. But also it's flat-out incorrect, because GEPA and
GEPB might be in different BBs, in which case we cannot ask which one
comes last (assertion failure,
https://reviews.llvm.org/D149893#4378332).
Fixed, now we use the correct context instruction.
Differential Revision: https://reviews.llvm.org/D151630
This patch fixes:
llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp:140:20: error:
unused function 'operator<<' [-Werror,-Wunused-function]
llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp:176:6: error:
unused function 'dumpChain' [-Werror,-Wunused-function]
This patch fixes:
llvm/lib/Transforms/Vectorize/LoadStoreVectorizer.cpp:1429:23:
error: comparison of integers of different signs: 'int' and 'const
size_t' (aka 'const unsigned long') [-Werror,-Wsign-compare]
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
Currently we happen to split a chain of 12xi8 accesses into 6xi8 + 6xi8, which
produces rather suboptimal code.
This change attempts to split-off non-multiples of 4bytes at the end and if that
does not work, splits on the smaller power-of-2 boundary.
Differential Revision: https://reviews.llvm.org/D147976
Removed definitions of vectorizeBasicBlock and VectorizeConfig
(possibly a remnant from the BBVectorize pass that was removed
way back in 2017).
Also reduced amount of include dependencies to Transforms/Vectorize.h.
Many uses of getIntPtrType() were using that type to calculate the
neened type for GEP offset arguments. However, some time ago,
DataLayout was extended to support pointers where the size of the
pointer is not equal to the size of the values used to index it.
Much code was already migrated to, for example, use getIndexSizeInBits
instead of getPtrSizeInBits, but some rewrites still used
getIntPtrType() to get the type for GEP offsets.
This commit changes uses of getIntPtrType() to getIndexType() where
they are involved in a GEP-related calculation.
In at least one case (bounds check insertion) this resolves a compiler
crash that the new test added here would previously trigger.
This commit does not impact
- C library-related rewriting (memcpy()), which are operating under
the assumption that intptr_t == size_t. While all the mechanisms for
breaking this assumption now exist, doing so is outside the scope of
this commit.
- Code generation and below. Note that the use of getIntPtrType() in
CodeGenPrepare will be changed in a future commit.
- Usage of getIntPtrType() in any backend
Depends on D143435
Reviewed By: arichardson
Differential Revision: https://reviews.llvm.org/D143437
Compare a relative speed of misaligned accesses before and
after vectorization, not just check the new instruction is
not going to be slower.
Since no target now returns anything but 0 or 1 for Fast
argument of the allowsMisalignedMemoryAccesses this is still NFCI.
The subsequent patch will tune actual vaues of Fast on AMDGPU.
Differential Revision: https://reviews.llvm.org/D124218
A target can return if a misaligned access is 'fast' as defined
by the target or not. In reality there can be different levels
of 'fast' and 'slow'. This patch changes the boolean 'Fast'
argument of the allowsMisalignedMemoryAccesses family of functions
to an unsigned representing its speed.
A target can still define it as it wants and the direct translation
of the current code uses 0 and 1 for current false and true. This
makes the change an NFC.
Subsequent patch will start using an actual value of speed in
the load/store vectorizer to compare if a vectorized access going
to be not just fast, but not slower than before.
Differential Revision: https://reviews.llvm.org/D124217
Most clients only used these methods because they wanted to be able to
extend or truncate to the same bit width (which is a no-op). Now that
the standard zext, sext and trunc allow this, there is no reason to use
the OrSelf versions.
The OrSelf versions additionally have the strange behaviour of allowing
extending to a *smaller* width, or truncating to a *larger* width, which
are also treated as no-ops. A small amount of client code relied on this
(ConstantRange::castOp and MicrosoftCXXNameMangler::mangleNumber) and
needed rewriting.
Differential Revision: https://reviews.llvm.org/D125557
Use shufflevector to do the subvector extracts. This allows a lot more
load merging on AMDGPU and also on NVPTX when <2 x half> is involved.
Differential Revision: https://reviews.llvm.org/D117219
Rather than checking for nounwind in particular, make sure the
instruction is guaranteed to transfer execution, which will also
handle non-willreturn calls correctly.
Fixes https://github.com/llvm/llvm-project/issues/52950.
This patch is changing the InsertElement's placeholder to poison without changing the LSV's behavior.
Regardless of whether `StoreTy` is FixedVectorType or not, the poison value will be overwritten with a different value.
Therefore, whether the InsertElement's placeholder is poison or undef will not affect the result of the program.
Reviewed By: nikic
Differential Revision: https://reviews.llvm.org/D111005
The load store vectorizer currently uses isNoAlias() to determine
whether memory-accessing instructions should prevent vectorization.
However, this only works for loads and stores. Additionally, a
couple of intrinsics like assume are special-cased to be ignored.
Instead use getModRefInfo() to generically determine whether the
instruction accesses/modifies the relevant location. This will
automatically handle all inaccessiblememonly intrinsics correctly
(as well as other calls that don't modref for other reasons).
This requires generalizing the code a bit, as it was previously
only considering loads and stored in particular.
Differential Revision: https://reviews.llvm.org/D109020
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 change introduces a new IR intrinsic named `llvm.pseudoprobe` for pseudo-probe block instrumentation. Please refer to https://reviews.llvm.org/D86193 for the whole story.
A pseudo probe is used to collect the execution count of the block where the probe is instrumented. This requires a pseudo probe to be persisting. The LLVM PGO instrumentation also instruments in similar places by placing a counter in the form of atomic read/write operations or runtime helper calls. While these operations are very persisting or optimization-resilient, in theory we can borrow the atomic read/write implementation from PGO counters and cut it off at the end of compilation with all the atomics converted into binary data. This was our initial design and we’ve seen promising sample correlation quality with it. However, the atomics approach has a couple issues:
1. IR Optimizations are blocked unexpectedly. Those atomic instructions are not going to be physically present in the binary code, but since they are on the IR till very end of compilation, they can still prevent certain IR optimizations and result in lower code quality.
2. The counter atomics may not be fully cleaned up from the code stream eventually.
3. Extra work is needed for re-targeting.
We choose to implement pseudo probes based on a special LLVM intrinsic, which is expected to have most of the semantics that comes with an atomic operation but does not block desired optimizations as much as possible. More specifically the semantics associated with the new intrinsic enforces a pseudo probe to be virtually executed exactly the same number of times before and after an IR optimization. The intrinsic also comes with certain flags that are carefully chosen so that the places they are probing are not going to be messed up by the optimizer while most of the IR optimizations still work. The core flags given to the special intrinsic is `IntrInaccessibleMemOnly`, which means the intrinsic accesses memory and does have a side effect so that it is not removable, but is does not access memory locations that are accessible by any original instructions. This way the intrinsic does not alias with any original instruction and thus it does not block optimizations as much as an atomic operation does. We also assign a function GUID and a block index to an intrinsic so that they are uniquely identified and not merged in order to achieve good correlation quality.
Let's now look at an example. Given the following LLVM IR:
```
define internal void @foo2(i32 %x, void (i32)* %f) !dbg !4 {
bb0:
%cmp = icmp eq i32 %x, 0
br i1 %cmp, label %bb1, label %bb2
bb1:
br label %bb3
bb2:
br label %bb3
bb3:
ret void
}
```
The instrumented IR will look like below. Note that each `llvm.pseudoprobe` intrinsic call represents a pseudo probe at a block, of which the first parameter is the GUID of the probe’s owner function and the second parameter is the probe’s ID.
```
define internal void @foo2(i32 %x, void (i32)* %f) !dbg !4 {
bb0:
%cmp = icmp eq i32 %x, 0
call void @llvm.pseudoprobe(i64 837061429793323041, i64 1)
br i1 %cmp, label %bb1, label %bb2
bb1:
call void @llvm.pseudoprobe(i64 837061429793323041, i64 2)
br label %bb3
bb2:
call void @llvm.pseudoprobe(i64 837061429793323041, i64 3)
br label %bb3
bb3:
call void @llvm.pseudoprobe(i64 837061429793323041, i64 4)
ret void
}
```
Reviewed By: wmi
Differential Revision: https://reviews.llvm.org/D86490
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
Now that load/store alignment is required, we no longer need most
of them. Also switch the getLoadStoreAlignment() helper to return
Align instead of MaybeAlign.
