Skip to content

[SCEV] Support non-constant step in howFarToZero #94411

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Jun 5, 2024
Merged

[SCEV] Support non-constant step in howFarToZero #94411

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
fcbdf94
[SCEV] Support scaled vscale as step in howFarToZero
preames Jun 4, 2024
ee61cc2
Address review comment
preames Jun 5, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
38 changes: 27 additions & 11 deletions llvm/lib/Analysis/ScalarEvolution.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10422,6 +10422,17 @@ SolveQuadraticAddRecRange(const SCEVAddRecExpr *AddRec,
return TruncIfPossible(MinOptional(SL.first, SU.first), BitWidth);
}

/// Return true if this is (C * vscale). This is the canonical form
/// of the step for loops vectorized with scalable vectors.
static bool matchScaledVScale(const SCEV *S) {
const SCEVMulExpr *Mul = dyn_cast<SCEVMulExpr>(S);
if (!Mul || Mul->getNumOperands() != 2)
return false;

return isa<SCEVConstant>(Mul->getOperand(0)) &&
Mul->getOperand(1)->getSCEVType() == scVScale;
}

ScalarEvolution::ExitLimit ScalarEvolution::howFarToZero(const SCEV *V,
const Loop *L,
bool ControlsOnlyExit,
Expand Down Expand Up @@ -10483,29 +10494,32 @@ ScalarEvolution::ExitLimit ScalarEvolution::howFarToZero(const SCEV *V,
// Get the initial value for the loop.
const SCEV *Start = getSCEVAtScope(AddRec->getStart(), L->getParentLoop());
const SCEV *Step = getSCEVAtScope(AddRec->getOperand(1), L->getParentLoop());

// For now we handle only constant steps.
//
// TODO: Handle a nonconstant Step given AddRec<NUW>. If the
// AddRec is NUW, then (in an unsigned sense) it cannot be counting up to wrap
// to 0, it must be counting down to equal 0. Consequently, N = Start / -Step.
// We have not yet seen any such cases.
const SCEVConstant *StepC = dyn_cast<SCEVConstant>(Step);
if (!StepC || StepC->getValue()->isZero())
// The code below is correct for arbitrary non-constant steps, but we won't be
// able to prove useful properties given the lack of use dependenct reasoning
// here. To avoid spending compile time for no value, bail early unless the
// step is a "useful" non-constant value.
if (!StepC && !matchScaledVScale(Step))
return getCouldNotCompute();

if (!isLoopInvariant(Step, L) || !isKnownNonZero(Step))
return getCouldNotCompute();

// For positive steps (counting up until unsigned overflow):
// N = -Start/Step (as unsigned)
// For negative steps (counting down to zero):
// N = Start/-Step
// First compute the unsigned distance from zero in the direction of Step.
bool CountDown = StepC->getAPInt().isNegative();
const SCEV *Distance = CountDown ? Start : getNegativeSCEV(Start);
bool CountDown = isKnownNegative(Step);
if (!CountDown && !isKnownNonNegative(Step))
return getCouldNotCompute();

const SCEV *Distance = CountDown ? Start : getNegativeSCEV(Start);
// Handle unitary steps, which cannot wraparound.
// 1*N = -Start; -1*N = Start (mod 2^BW), so:
// N = Distance (as unsigned)
if (StepC->getValue()->isOne() || StepC->getValue()->isMinusOne()) {
if (StepC &&
(StepC->getValue()->isOne() || StepC->getValue()->isMinusOne())) {
APInt MaxBECount = getUnsignedRangeMax(applyLoopGuards(Distance, L));
MaxBECount = APIntOps::umin(MaxBECount, getUnsignedRangeMax(Distance));

Expand Down Expand Up @@ -10550,6 +10564,8 @@ ScalarEvolution::ExitLimit ScalarEvolution::howFarToZero(const SCEV *V,
}

// Solve the general equation.
if (!StepC)
return getCouldNotCompute();
const SCEV *E = SolveLinEquationWithOverflow(StepC->getAPInt(),
getNegativeSCEV(Start), *this);

Expand Down
11 changes: 6 additions & 5 deletions llvm/test/Analysis/ScalarEvolution/scalable-vector.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -91,13 +91,14 @@ define void @vscale_step_ne_tripcount(i64 %N) vscale_range(2, 1024) {
; CHECK-NEXT: %n.vec = sub i64 %n.rnd.up, %n.mod.vf
; CHECK-NEXT: --> (4 * vscale * ((-1 + (4 * vscale)<nuw><nsw> + %N) /u (4 * vscale)<nuw><nsw>)) U: [0,-3) S: [-9223372036854775808,9223372036854775805)
; CHECK-NEXT: %index = phi i64 [ 0, %vector.ph ], [ %index.next, %vector.body ]
; CHECK-NEXT: --> {0,+,(4 * vscale)<nuw><nsw>}<nuw><%vector.body> U: [0,-3) S: [-9223372036854775808,9223372036854775805) Exits: <<Unknown>> LoopDispositions: { %vector.body: Computable }
; CHECK-NEXT: --> {0,+,(4 * vscale)<nuw><nsw>}<nuw><%vector.body> U: [0,-3) S: [-9223372036854775808,9223372036854775805) Exits: (4 * vscale * ((-1 * vscale * (4 + (-4 * ((-1 + (4 * vscale)<nuw><nsw> + %N) /u (4 * vscale)<nuw><nsw>))<nsw>)<nsw>) /u (4 * vscale)<nuw><nsw>)) LoopDispositions: { %vector.body: Computable }
; CHECK-NEXT: %index.next = add nuw i64 %index, %2
; CHECK-NEXT: --> {(4 * vscale)<nuw><nsw>,+,(4 * vscale)<nuw><nsw>}<nuw><%vector.body> U: [8,-3) S: [-9223372036854775808,9223372036854775805) Exits: <<Unknown>> LoopDispositions: { %vector.body: Computable }
; CHECK-NEXT: --> {(4 * vscale)<nuw><nsw>,+,(4 * vscale)<nuw><nsw>}<nuw><%vector.body> U: [8,-3) S: [-9223372036854775808,9223372036854775805) Exits: (vscale * (4 + (4 * ((-1 * vscale * (4 + (-4 * ((-1 + (4 * vscale)<nuw><nsw> + %N) /u (4 * vscale)<nuw><nsw>))<nsw>)<nsw>) /u (4 * vscale)<nuw><nsw>))<nuw><nsw>)<nuw>) LoopDispositions: { %vector.body: Computable }
; CHECK-NEXT: Determining loop execution counts for: @vscale_step_ne_tripcount
; CHECK-NEXT: Loop %vector.body: Unpredictable backedge-taken count.
; CHECK-NEXT: Loop %vector.body: Unpredictable constant max backedge-taken count.
; CHECK-NEXT: Loop %vector.body: Unpredictable symbolic max backedge-taken count.
; CHECK-NEXT: Loop %vector.body: backedge-taken count is ((-1 * vscale * (4 + (-4 * ((-1 + (4 * vscale)<nuw><nsw> + %N) /u (4 * vscale)<nuw><nsw>))<nsw>)<nsw>) /u (4 * vscale)<nuw><nsw>)
; CHECK-NEXT: Loop %vector.body: constant max backedge-taken count is i64 2305843009213693951
; CHECK-NEXT: Loop %vector.body: symbolic max backedge-taken count is ((-1 * vscale * (4 + (-4 * ((-1 + (4 * vscale)<nuw><nsw> + %N) /u (4 * vscale)<nuw><nsw>))<nsw>)<nsw>) /u (4 * vscale)<nuw><nsw>)
; CHECK-NEXT: Loop %vector.body: Trip multiple is 1
;
entry:
%0 = sub i64 -1, %N
Expand Down