The expansion of a SCEVUnknown is trivial (it's just the wrapped value).
If we try to reuse an existing value it might be a more complex
expression that simplifies to the SCEVUnknown.
This is inspired by https://github.com/llvm/llvm-project/issues/114879,
because SCEVExpander replacing a constant with a phi node is just silly.
(I don't consider this a fix for that issue though.)
Comments attached to the `ScaledReg` field of `struct Formula` explains
that, `ScaledReg` must be non-null when `Scale` is non-zero.
This fixes up a code path where this invariant is violated. Also, add an
assert to ensure this invariant holds true.
Without this patch, compiler aborts with the attached test case.
Fixes#76504
As of rev ea222be0d, LLVMs assembler will actually try to honour the
"fill value" part of p2align directives. X86 printed these as 0x90, which
isn't actually what it wanted: we want multi-byte nops for .text
padding. Compiling via a textual assembly file produces single-byte
nop padding since ea222be0d but the built-in assembler will produce
multi-byte nops. This divergent behaviour is undesirable.
To fix: don't set the byte padding field for x86, which allows the
assembler to pick multi-byte nops. Test that we get the same multi-byte
padding when compiled via textual assembly or directly to object file.
Added same-align-bytes-with-llasm-llobj.ll to that effect, updated
numerous other tests to not contain check-lines for the explicit padding.
When expanding SCEV adds to geps, transfer the nuw flag to the resulting
gep. (Note that this doesn't apply to IV increment GEPs, which go
through a different code path.)
As pointed out in the review of #102133, SCEVExpander currently
incorrectly reuses GEP instructions that have poison-generating flags
set. Fix this by clearing the flags on the reused instruction.
Motivating example: https://godbolt.org/z/eb97zrxhx
Here we have 2 induction variables in the loop: one is corresponding to
i variable (add rdx, 4), the other - to res (add rax, 2). The second
induction variable can be removed by rewriteLoopExitValues() method
(final value of res at loop exit is unroll_iter * -2); however, this
doesn't happen because we have duplicated LCSSA phi nodes at loop exit:
```
; Preheader:
for.body.preheader.new: ; preds = %for.body.preheader
%unroll_iter = and i64 %N, -4
br label %for.body
; Loop:
for.body: ; preds = %for.body, %for.body.preheader.new
%lsr.iv = phi i64 [ %lsr.iv.next, %for.body ], [ 0, %for.body.preheader.new ]
%i.07 = phi i64 [ 0, %for.body.preheader.new ], [ %inc.3, %for.body ]
%inc.3 = add nuw i64 %i.07, 4
%lsr.iv.next = add nsw i64 %lsr.iv, -2
%niter.ncmp.3.not = icmp eq i64 %unroll_iter, %inc.3
br i1 %niter.ncmp.3.not, label %for.end.loopexit.unr-lcssa.loopexit, label %for.body, !llvm.loop !7
; Exit blocks
for.end.loopexit.unr-lcssa.loopexit: ; preds = %for.body
%inc.3.lcssa = phi i64 [ %inc.3, %for.body ]
%lsr.iv.next.lcssa11 = phi i64 [ %lsr.iv.next, %for.body ]
%lsr.iv.next.lcssa = phi i64 [ %lsr.iv.next, %for.body ]
br label %for.end.loopexit.unr-lcssa
```
rewriteLoopExitValues requires %lsr.iv.next value to have only 2 uses:
one in LCSSA phi node, the other - in induction phi node. Here we have 3
uses of this value because of duplicated lcssa nodes, so the transform
doesn't apply and leads to an extra add operation inside the loop. The
proposed solution is to accumulate inserted instructions that will
require LCSSA form update into SetVector and then call
formLCSSAForInstructions for this SetVector once, so the same
instructions don't process twice.
Reland fixes the issue with preserve-lcssa.ll test: it fails in the situation
when x86_64-unknown-linux-gnu target is unavailable in opt. The changes are
moved into separate duplicated-phis.ll test with explicit x86 target requirement
to fix bots which are not building this target.
Motivating example: https://godbolt.org/z/eb97zrxhx
Here we have 2 induction variables in the loop: one is corresponding to
i variable (add rdx, 4), the other - to res (add rax, 2). The second
induction variable can be removed by rewriteLoopExitValues() method
(final value of res at loop exit is unroll_iter * -2); however, this
doesn't happen because we have duplicated LCSSA phi nodes at loop exit:
```
; Preheader:
for.body.preheader.new: ; preds = %for.body.preheader
%unroll_iter = and i64 %N, -4
br label %for.body
; Loop:
for.body: ; preds = %for.body, %for.body.preheader.new
%lsr.iv = phi i64 [ %lsr.iv.next, %for.body ], [ 0, %for.body.preheader.new ]
%i.07 = phi i64 [ 0, %for.body.preheader.new ], [ %inc.3, %for.body ]
%inc.3 = add nuw i64 %i.07, 4
%lsr.iv.next = add nsw i64 %lsr.iv, -2
%niter.ncmp.3.not = icmp eq i64 %unroll_iter, %inc.3
br i1 %niter.ncmp.3.not, label %for.end.loopexit.unr-lcssa.loopexit, label %for.body, !llvm.loop !7
; Exit blocks
for.end.loopexit.unr-lcssa.loopexit: ; preds = %for.body
%inc.3.lcssa = phi i64 [ %inc.3, %for.body ]
%lsr.iv.next.lcssa11 = phi i64 [ %lsr.iv.next, %for.body ]
%lsr.iv.next.lcssa = phi i64 [ %lsr.iv.next, %for.body ]
br label %for.end.loopexit.unr-lcssa
```
rewriteLoopExitValues requires %lsr.iv.next value to have only 2 uses:
one in LCSSA phi node, the other - in induction phi node. Here we have 3
uses of this value because of duplicated lcssa nodes, so the transform
doesn't apply and leads to an extra add operation inside the loop. The
proposed solution is to accumulate inserted instructions that will
require LCSSA form update into SetVector and then call
formLCSSAForInstructions for this SetVector once, so the same
instructions don't process twice.
This transformation doesn't actually use any of the internal state of
LSR and recomputes all information from SCEV. Splitting it out makes
it easier to test.
Note that long term I would like to write a version of this transform
which *is* integrated with LSR's solver, but if that happens, we'll
just delete the extra pass.
Integration wise, I switched from using TTI to using a pass configuration
variable. This seems slightly more idiomatic, and means we don't run
the extra logic on any target other than RISCV.
Somewhat confusingly a `SCEVMulExpr` is a `SCEVNAryExpr`, so can have
> 2 operands. Previously, the vscale immediate matching did not check
the number of operands of the `SCEVMulExpr`, so would ignore any
operands after the first two.
This led to incorrect codegen (and results) for ArmSME in IREE
(https://github.com/iree-org/iree), which sometimes addresses things
that are a `vscale * vscale` multiple away. The test added with this
change shows an example reduced from IREE. The second write should
be offset from the first `16 * vscale * vscale` (* 4 bytes), however,
previously LSR dropped the second vscale and instead offset the write by
`#4, mul vl`, which is an offset of `16 * vscale` (* 4 bytes).
Currently before forwarding an AVL we do a simple non-exhaustive check
to see if its LiveInterval is extendable. But we also need to check for
this when we're extending an AVL's LiveInterval via merging the
VSETVLIInfos in transferBefore with equally zero AVLs.
Rather than trying to conservatively prevent these cases, this inserts a
copy of the AVL instead if we don't know we'll be able to extend it.
This is likely to be more robust, and even if the extra copy is
undesirable these cases should be rare in practice.
Fix#97510 .
Note that, for the new phi instruction `NewPH`, which replaces the old
phi `PH` and the cast `ShadowUse`, I choose to propagate the debug
location of `PH` to it, because the cast is eliminated according to the
optimization semantics.
This avoids some cases where LSR produces results that lead to very poor
codegen. There's a chance we'll see minor degradations for some inputs
in the case that our metrics say the found solution is worse, but in
reality it's better than the starting point.
Per the review thread, at least one vendor has been enabling this by
defualt for some time and found overall it's an improvement. As such,
we'll enable by default and aim to fix any as-yet-unknown regressions
in-tree.
This fixes a crash found when compiling OpenBLAS with -mllvm
-verify-machineinstrs.
When we "forward" the AVL from the output of a vsetvli, we might have to
extend the LiveInterval of the AVL to where insert the new vsetvli.
Most of the time we are able to extend the LiveInterval because there's
only one val num (definition) for the register. But PHI elimination can
assign multiple values to the same register, in which case we end up
clobbering a different val num when extending:
%x = PseudoVSETVLI %avl, ...
%avl = ADDI ...
%v = PseudoVADD ..., avl=%x
; %avl is forwarded to PseudoVADD:
%x = PseudoVSETVLI %avl, ...
%avl = ADDI ...
%v = PseudoVADD ..., avl=%avl
Here there's no way to extend the %avl from the vsetvli since %avl is
redefined, i.e. we have two val nums.
This fixes it by only forwarding it when we have exactly one val num,
where it should be safe to extend it.
Extends LoopStrengthReduce to recognize immediates multiplied by vscale, and query the current target for whether they are legal offsets for memory operations or adds.
If the AVL in a VSETVLIInfo is the output VL of a vsetvli with the same
VLMAX, we treat it as the AVL of said vsetvli.
This allows us to remove a true dependency as well as treating
VSETVLIInfos as equal in more places and avoid toggles.
We do this in two places, needVSETVLI and computeInfoForInstr. However
we don't do this in computeInfoForInstr's vsetvli equivalent,
getInfoForVSETVLI.
We also have a restriction only in computeInfoForInstr that the AVL
can't be a register as we want to avoid extending live ranges.
This patch does two interlinked things:
1) It adds this AVL "peeking" to getInfoForVSETVLI
2) It relaxes the constraint that the AVL can't be a register in
computeInfoForInstr, since it removes a use of the output VL which can
actually reduce register pressure. E.g. see the diff in
@vector_init_vsetvli_N and @test6
Now that getInfoForVSETVLI and computeInfoForInstr are consistent, we
can remove the check in needVSETVLI.
We also need to update how we update LiveIntervals in insertVSETVLI, as
we can now end up needing to extend the LiveRange of the AVL across
blocks.
If we have a urem expression, emitting it as a urem is significantly
better that letting the fully expansion kick in. We have the risk of a
udiv or mul which could have previously been shared, but loosing that
seems like a reasonable tradeoff for being able to round trip a urem w/o
modification.
This patch makes the final major change of the RemoveDIs project, changing the
default IR output from debug intrinsics to debug records. This is expected to
break a large number of tests: every single one that tests for uses or
declarations of debug intrinsics and does not explicitly disable writing
records.
If this patch has broken your downstream tests (or upstream tests on a
configuration I wasn't able to run):
1. If you need to immediately unblock a build, pass
`--write-experimental-debuginfo=false` to LLVM's option processing for all
failing tests (remember to use `-mllvm` for clang/flang to forward arguments to
LLVM).
2. For most test failures, the changes are trivial and mechanical, enough that
they can be done by script; see the migration guide for a guide on how to do
this: https://llvm.org/docs/RemoveDIsDebugInfo.html#test-updates
3. If any tests fail for reasons other than FileCheck check lines that need
updating, such as assertion failures, that is most likely a real bug with this
patch and should be reported as such.
For more information, see the recent PSA:
https://discourse.llvm.org/t/psa-ir-output-changing-from-debug-intrinsics-to-debug-records/79578
Adds an AArch64-specific version of isLSRCostLess, changing the relative
importance of the various terms from the formulae being evaluated.
This has been split out from my vscale-aware LSR work, see the RFC for
reference:
https://discourse.llvm.org/t/rfc-vscale-aware-loopstrengthreduce/77131
This patch canonicalizes constant expression GEPs to use i8 source
element type, aka ptradd. This is the ConstantFolding equivalent of the
InstCombine canonicalization introduced in #68882.
I believe all our optimizations working on constant expression GEPs
(like GlobalOpt etc) have already been switched to work on offsets, so I
don't expect any significant fallout from this change.
This is part of:
https://discourse.llvm.org/t/rfc-replacing-getelementptr-with-ptradd/68699
It's common to delete some instructions after using SCEVExpander,
while it is still live (but will not be used afterwards). In that
case, the AssertingVH may trigger. Replace it with a PoisoningVH
so that we only detect the case where the SCEVExpander actually is
used in a problematic fashion after the instruction removal.
The alternative would be to add clear() calls to more code paths.
Fixes https://github.com/llvm/llvm-project/issues/83404.
If the phi node found by matchSimpleRecurrence is not from the current
loop, then isAlmostDeadIV panics. With this patch we bail out early.
Signed-off-by: Patrick O'Neill <patrick@rivosinc.com>
---------
Signed-off-by: Patrick O'Neill <patrick@rivosinc.com>
Follow up to https://github.com/llvm/llvm-project/pull/74747
This change extends the previously added fixed expansion threshold by
scaling down the cost allowed for an expansion for a loop with either a
small known trip count or a profile which indicates the trip count is
likely small. The goal here is to improve code generation for a loop
nest where the outer loop has a high trip count, and the inner loop runs
only a handful of iterations.
---------
Co-authored-by: Nikita Popov <github@npopov.com>
This is a follow up to an item I noted in my submission comment for
e947f95. I don't have a real world example where this is triggering
unprofitably, but avoiding the transform when we estimate the loop to be
short running from profiling seems quite reasonable. It's also now come
up as a possibility in a regression twice in two days, so I'd like to
get this in to close out the possibility if nothing else.
The original review dropped the threshold for short trip count loops. I
will return to that in a separate review if this lands.
This patch trivially extends support for DbgValueInst recovery to
DPValues in LoopStrengthReduce; they are handled identically, so this is
mostly done by reusing the DbgValueInst code (using templates or
auto-parameter lambdas to reduce actual code duplication).
Similar to 806761a7629df268c8aed49657aeccffa6bca449.
For IR files without a target triple, -mtriple= specifies the full
target triple while -march= merely sets the architecture part of the
default target triple, leaving a target triple which may not make sense,
e.g. amdgpu-apple-darwin.
Therefore, -march= is error-prone and not recommended for tests without
a target triple. The issue has been benign as we recognize
$unknown-apple-darwin as ELF instead of rejecting it outrightly.
This patch changes AMDGPU tests to not rely on the default
OS/environment components. Tests that need fixes are not changed:
```
LLVM :: CodeGen/AMDGPU/fabs.f64.ll
LLVM :: CodeGen/AMDGPU/fabs.ll
LLVM :: CodeGen/AMDGPU/floor.ll
LLVM :: CodeGen/AMDGPU/fneg-fabs.f64.ll
LLVM :: CodeGen/AMDGPU/fneg-fabs.ll
LLVM :: CodeGen/AMDGPU/r600-infinite-loop-bug-while-reorganizing-vector.ll
LLVM :: CodeGen/AMDGPU/schedule-if-2.ll
```
The existing logic in isKnownNonZero relies on unsigned ranges, which
can be problematic when our range calculation is imprecise. Consider the
following:
%offset.nonzero = or i32 %offset, 1
--> %offset.nonzero U: [1,0) S: [1,0)
%offset.i64 = sext i32 %offset.nonzero to i64
--> (sext i32 %offset.nonzero to i64) U: [-2147483648,2147483648)
S: [-2147483648,2147483648)
Note that the unsigned range for the sext does contain zero in this case
despite the fact that it can never actually be zero.
Instead, we can push the query down one level - relying on the fact that
the sext is an invertible operation and that the result can only be zero
if the input is. We could likely generalize this reasoning for other
invertible operations, but special casing sext seems worthwhile.
This reverts commit fdb87640ee2be63af9b0e0cd943cb13d79686a03, and thus
re-enables terminator folding for RISCV. The reported miscompile has
been fixed in f5dd70c58277d925710e5a7c25c86d7565cc3c6c.
The term folding logic needs to prove that the induction variable does
not cycle through the same set of values so that testing for the value
of the IV on the exiting iteration is guaranteed to trigger only on that
iteration. The prior code checked the no-self-wrap property on the IV,
but this is insufficient as a zero step is trivially no-self-wrap per
SCEV's definition but does repeat the same series of values.
In the current form, this has the effect of basically disabling lsr's
term-folding for all non-constant strides. This is still a net
improvement as we've disabled term-folding entirely, so being able to
enable it for constant strides is still a net improvement.
As future work, there's two SCEV weakness worth investigating.
First sext (or i32 %a, 1) to i64 does not return true for
isKnownNonZero. This is because we check only the unsigned range in that
query. We could either do query pushdown, or check the signed range as
well. I tried the second locally and it has very broad impact - i.e. we
have a bunch of missing optimizations here.
Second, zext (or i32 %a, 1) to i64 as the increment to the IV in
expensive_expand_short_tc causes the addrec to no longer be provably
no-self-wrap. I didn't investigate this so it might be necessary, but
the loop structure is such that I find this result surprising.
