The existing logic was unsound, in two ways.
First, due to wrapping on the trip count computation, it could compute a value which convert a loop which exiting on iteration 256, to one which exited at 255. (With i8 trip counts.)
Second, it allowed rewriting when the trip count implies wrapping around the alternate IV. As a trivial example, it allowed rewriting an i128 exit test in terms of an i64 IV. This is obviously wrong.
Note that the test change is fairly minimal - i.e. only the targeted test - but that's only because I precommitted a change which switched the test from 32 to 64 bit pointers. For 32 bit point architectures with 32 bit primary inductions, this transform is almost always unsound to perform.
Differential Revision: https://reviews.llvm.org/D146429
Also, add a comment to highlight that the "good" result on this test is accidental, and not based on a principled decision. I matched the original behavior to make this nfc, but selecting the last legal IV is not well motivated here.
Main benefit here is making the logic easier to follow, slightly more efficient, and more in line with LFTR. This is not NFC. There are three semantic changes here.
First, we drop handling for constants on the LHS of the comparison. These are non-canonical, and we're very late in the optimization pipeline here, so there's no point in supporting this. I removed a test which covered this case.
Second, we don't need the almost dead IV to be an addrec. We just need SCEV to be able to compute a trip count for it.
Third, we require a simple IV for the almost dead IV. In theory, this removes cases we could have previously handled, but given a) zero testing and b) multiple known correctness issues, I'm adopting an attidute of narrowing this down to something which works correctly, and *then* expanding.
There were two major problems with the tests.
First, with the pointer size being 32 bit and the original IVs also being 32 bit, almost all of the positive tests were actually unsound. An upcoming change will add the appropriate safety check, but the test diffs are really hard to understand without switching the tests to 64 bit pointers first.
Second, checking debug messages for failures is a major bad practice. This should not have been accepted in review at all. The reason is that it makes the *order* of legality checks visibile and modifying any of them becomes annoying and tedious.
Autogen for naming change, and remove comments about C code inspiration. Multiple of these are out of sync with the actual IR, and these are IR tests anyways.
DFAJumpThreading
JumpThreading
LibCallsShrink
LoopVectorize
SLPVectorizer
DeadStoreElimination
AggressiveDCE
CorrelatedValuePropagation
IndVarSimplify
These are part of the optimization pipeline, of which the legacy version is deprecated and being removed.
Like what has been done in AArch64 (D125335).
We enable this under `-O2` to show the codegen diffs here but we
may only do this under `-O3` like AArch64.
There are two cases that we may produce these eliminable copies:
1. ISel of `FrameIndex`. Like `rvv/fixed-vectors-calling-conv.ll`.
2. Tail duplication. Like `select-optimize-multiple.ll`.
Reviewed By: craig.topper
Differential Revision: https://reviews.llvm.org/D144535
The vsetvli insertion pass can replace it with MU if needed by
a using instruction. The vsetvli insertion pass will not convert
MU to MA so we need to start at MA.
Reviewed By: eopXD
Differential Revision: https://reviews.llvm.org/D143790
Original code will cause crash when the load/store memory type is structure because isIndexedLoadLegal/isIndexedStore doesn't support struct type.
So we limit the load/store memory type to integer.
Origin commit message:
When the latch block is different from header block, IVInc will be expanded in the latch loop. We can't generate the post index load/store this case.
But if the IVInc only used in the loop, actually we still can use the post index load/store because when exit loop we don't care the last IVInc value.
So, trying to hoist IVInc to help backend to generate more post index load/store.
Fix#53625
Reviewed By: eopXD
Differential Revision: https://reviews.llvm.org/D138636
This matches the behavior from a number of other targets, including e.g. X86. This does have the effect of increasing register pressure slightly, but we have a relative abundance of registers in the ISA compared to other targets which use the same heuristic.
The motivation here is that our current cost heuristic treats number of registers as the dominant cost. As a result, an extra use outside of a loop can radically change the LSR result. As an example consider test4 from the recently added test/Transforms/LoopStrengthReduce/RISCV/lsr-cost-compare.ll. Without a use outside the loop (see test3), we convert the IV into a pointer increment. With one, we leave the gep in place.
The pointer increment version both decreases number of instructions in some loops, and creates parallel chains of computation (i.e. decreases critical path depth). Both are generally profitable.
Arguably, we should really be using a more sophisticated model here - such as e.g. using profile information or explicitly modeling parallelism gains. However, as a practical matter starting with the same mild hack that other targets have used seems reasonable.
Differential Revision: https://reviews.llvm.org/D142227
There's no need to require the start value to come directly from the loop predecessor. This was sometimes covering up a latent miscompile in this off-by-default option, but the miscompile needs fixed anyways and the issue has been raised on the original review.
Differential Revision: https://reviews.llvm.org/D142240
IR is now always parsed in opaque pointer mode, unless
-opaque-pointers=0 is explicitly given. There is no automatic
detection of typed pointers anymore.
The -opaque-pointers=0 option is added to any remaining IR tests
that haven't been migrated yet.
Differential Revision: https://reviews.llvm.org/D141912
This patch relaxes the restriction that both reassociate operands must
be in the same block as the root instruction.
The comment indicates that the reason for this restriction was that the
operands not in the same block won't have a depth in the trace.
I believe this is outdated; if the operand is in a different block, it
must dominate the current block (otherwise it would need to be phi),
which in turn means the operand's block must be included in the current
rance, and depths must be available.
There's a test case (no_reassociate_different_block) added in
70520e2f1c5fc4 which shows that we have accurate depths for operands
defined in other blocks.
This allows reassociation of code that computes the final reduction
value after vectorization, among other things.
Reviewed By: dmgreen
Differential Revision: https://reviews.llvm.org/D141302
When the latch block is different from header block, IVInc will be expanded in the latch loop. We can't generate the post index load/store this case.
But if the IVInc only used in the loop, actually we still can use the post index load/store because when exit loop we don't care the last IVInc value.
So, trying to hoist IVInc to help backend to generate more post index load/store.
Fix#53625
Reviewed By: eopXD
Differential Revision: https://reviews.llvm.org/D138636
One of these two changes is exposing (or causing) some more miscompiles.
A reproducer is in progress, so reverting until resolved.
This reverts commit 428f36401b1b695fd501ebfdc8773bed8ced8d4e.
This reverts commit 37b8f09a4b61bf9bf9d0b9017d790c8b82be2e17,
and returns commit 1bd0b82e508d049efdb07f4f8a342f35818df341.
The miscompile was in InstCombine, and it has been addressed.
This tries to approach the problem noted by @arsenm:
terrible codegen for `__builtin_fpclassify()`:
https://godbolt.org/z/388zqdE37
Just because the PHI in the common successor happens to have different
incoming values for these two blocks, doesn't mean we have to give up.
It's quite easy to deal with this, we just need to produce a select:
https://alive2.llvm.org/ce/z/000srb
Now, the cost model for this transform is rather overly strict,
so this will basically never fire. We tally all (over all preds)
the selects needed to the NumBonusInsts
Differential Revision: https://reviews.llvm.org/D139275
This tries to approach the problem noted by @arsenm:
terrible codegen for `__builtin_fpclassify()`:
https://godbolt.org/z/388zqdE37
Just because the PHI in the common successor happens to have different
incoming values for these two blocks, doesn't mean we have to give up.
It's quite easy to deal with this, we just need to produce a select:
https://alive2.llvm.org/ce/z/000srb
Now, the cost model for this transform is rather overly strict,
so this will basically never fire. We tally all (over all preds)
the selects needed to the NumBonusInsts
Differential Revision: https://reviews.llvm.org/D139275
According to report by @JojoR, the assertion error was hit hence we need
to have this check before the actual transformation.
Reviewed By: Meinersbur, #loopoptwg
Differential Revision: https://reviews.llvm.org/D136415
The LSR may suggest less profitable transformation to the loop. This
patch adds check to prevent LSR from generating worse code than what
we already have.
Since LSR affects nearly all targets, the patch is guarded by the
option 'lsr-drop-solution' and default as disable for now.
The next step should be extending an TTI interface to allow target(s)
to enable this enhancememnt.
Debug log is added to remind user of such choice to skip the LSR
solution.
Reviewed By: Meinersbur, #loopoptwg
Differential Revision: https://reviews.llvm.org/D126043
This stops reporting CostPerUse 1 for `R8`-`R15` and `XMM8`-`XMM31`.
This was previously done because instruction encoding require a REX
prefix when using them resulting in longer instruction encodings. I
found that this regresses the quality of the register allocation as the
costs impose an ordering on eviction candidates. I also feel that there
is a bit of an impedance mismatch as the actual costs occure when
encoding instructions using those registers, but the order of VReg
assignments is not primarily ordered by number of Defs+Uses.
I did extensive measurements with the llvm-test-suite wiht SPEC2006 +
SPEC2017 included, internal services showed similar patterns. Generally
there are a log of improvements but also a lot of regression. But on
average the allocation quality seems to improve at a small code size
regression.
Results for measuring static and dynamic instruction counts:
Dynamic Counts (scaled by execution frequency) / Optimization Remarks:
Spills+FoldedSpills -5.6%
Reloads+FoldedReloads -4.2%
Copies -0.1%
Static / LLVM Statistics:
regalloc.NumSpills mean -1.6%, geomean -2.8%
regalloc.NumReloads mean -1.7%, geomean -3.1%
size..text mean +0.4%, geomean +0.4%
Static / LLVM Statistics:
mean -2.2%, geomean -3.1%) regalloc.NumSpills
mean -2.6%, geomean -3.9%) regalloc.NumReloads
mean +0.6%, geomean +0.6%) size..text
Static / LLVM Statistics:
regalloc.NumSpills mean -3.0%
regalloc.NumReloads mean -3.3%
size..text mean +0.3%, geomean +0.3%
Differential Revision: https://reviews.llvm.org/D133902
When the IV is only used by the terminating condition (say IV-A) and the loop
has a predictable back-edge count and we have another IV (say IV-B) that is an
affine add recursion, we will be able to calculate the terminating value of
IV-B in the loop pre-header. This patch adds attempts to replace IV-B as the
new terminating condition and remove IV-A. It is safe to do so since IV-A is
only used as the terminating condition.
This transformation is suitable to be appended after LSR as it may optimize the
loop into the situation mentioned above. The transformation can reduce number of
IV-s in the loop by one.
A cli option `lsr-term-fold` is added and default disabled.
Reviewed By: mcberg2021, craig.topper
Differential Revision: https://reviews.llvm.org/D132443
Changes since initial commit:
* Wrapping a pointer in an SCEV unknown hides the base, and SCEV is only able to compute a subtraction when the bases are known to be equal. This results in a SCEVCouldNotCompute flowing forward and triggering asserts. Test case added in d767b392.
* isLoopInvariant returns true for instructions outside the loop, but not necessarily *above* the loop. Since this code is allowed to visit uses of an IV outside of a loop, we have to make sure the operands of the compare are both invariant and dominating the header. Test case added in 2aed3cdb.
Original commit message follows...
The ICmpZero matching is checking to see if the expression is loop invariant per SCEV and expandable. This allows expressions inside the loop which can be made loop invariant to be seamlessly expanded, but is overly conservative for expressions which already *are* loop invariant.
As a simple justification for why this is correct, consider a loop invariant urem as RHS vs an alternate function with that same urem wrapped inside a helper call. Why would it be legal to match the later, but not the former?
Differential Revision: https://reviews.llvm.org/D129793
The ICmpZero matching is checking to see if the expression is loop invariant per SCEV and expandable. This allows expressions inside the loop which can be made loop invariant to be seamlessly expanded, but is overly conservative for expressions which already *are* loop invariant.
As a simple justification for why this is correct, consider a loop invariant urem as RHS vs an alternate function with that same urem wrapped inside a helper call. Why would it be legal to match the later, but not the former?
Differential Revision: https://reviews.llvm.org/D129793
Following some recent discussions, this changes the representation
of callbrs in IR. The current blockaddress arguments are replaced
with `!` label constraints that refer directly to callbr indirect
destinations:
; Before:
%res = callbr i8* asm "", "=r,r,i"(i8* %x, i8* blockaddress(@test8, %foo))
to label %asm.fallthrough [label %foo]
; After:
%res = callbr i8* asm "", "=r,r,!i"(i8* %x)
to label %asm.fallthrough [label %foo]
The benefit of this is that we can easily update the successors of
a callbr, without having to worry about also updating blockaddress
references. This should allow us to remove some limitations:
* Allow unrolling/peeling/rotation of callbr, or any other
clone-based optimizations
(https://github.com/llvm/llvm-project/issues/41834)
* Allow duplicate successors
(https://github.com/llvm/llvm-project/issues/45248)
This is just the IR representation change though, I will follow up
with patches to remove limtations in various transformation passes
that are no longer needed.
Differential Revision: https://reviews.llvm.org/D129288
Follow up to 3bc09c7da5 - remove a fixme I forgot to remove, and add test cases showing remaining work.
Note that scaled vscales show up in vectorized code from a couple of sources:
* Element types smaller than vector block size (i.e. everything under i64)
* Unrolling
* LMUL > 1
The largest scaling we can currently have is 256 (e8 in every possible vector register). More practically useful scales are in the 2-16 range.
