Skip to content

Commit

Permalink
[indvars] Always fallback to truncation if AddRec widening fails (#70967
Browse files Browse the repository at this point in the history 
)

The current code structure results in cases where if a) we can't clone
the IV user (because it's not in our whitelist) or b) can't prove the
SCEV expressions are identical, we'd sometimes leave both the original
unwiddened IV and the partially widdened IV in code. Instead, just
truncate thw wide IV to the use - same as what we'd do if we couldn't
find an addrec to start with.

Noticed this while playing with changing how we produce addrecs. The
current structure results in a very tight interlock between SCEVs
internal capabilities and indvars code.
  • Loading branch information
@preames
preames authored Nov 7, 2023
1 parent 24060db commit 551c280
Show file tree
Hide file tree
Showing 3 changed files with 71 additions and 69 deletions.
107 changes: 56 additions & 51 deletions llvm/lib/Transforms/Utils/SimplifyIndVar.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1805,65 +1805,70 @@ Instruction *WidenIV::widenIVUse(WidenIV::NarrowIVDefUse DU, SCEVExpander &Rewri
return nullptr;
}

// Does this user itself evaluate to a recurrence after widening?
WidenedRecTy WideAddRec = getExtendedOperandRecurrence(DU);
if (!WideAddRec.first)
WideAddRec = getWideRecurrence(DU);

assert((WideAddRec.first == nullptr) ==
(WideAddRec.second == ExtendKind::Unknown));
if (!WideAddRec.first) {
// If use is a loop condition, try to promote the condition instead of
// truncating the IV first.
if (widenLoopCompare(DU))
auto tryAddRecExpansion = [&]() -> Instruction* {
// Does this user itself evaluate to a recurrence after widening?
WidenedRecTy WideAddRec = getExtendedOperandRecurrence(DU);
if (!WideAddRec.first)
WideAddRec = getWideRecurrence(DU);
assert((WideAddRec.first == nullptr) ==
(WideAddRec.second == ExtendKind::Unknown));
if (!WideAddRec.first)
return nullptr;

// We are here about to generate a truncate instruction that may hurt
// performance because the scalar evolution expression computed earlier
// in WideAddRec.first does not indicate a polynomial induction expression.
// In that case, look at the operands of the use instruction to determine
// if we can still widen the use instead of truncating its operand.
if (widenWithVariantUse(DU))
// Reuse the IV increment that SCEVExpander created as long as it dominates
// NarrowUse.
Instruction *WideUse = nullptr;
if (WideAddRec.first == WideIncExpr &&
Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
WideUse = WideInc;
else {
WideUse = cloneIVUser(DU, WideAddRec.first);
if (!WideUse)
return nullptr;
}
// Evaluation of WideAddRec ensured that the narrow expression could be
// extended outside the loop without overflow. This suggests that the wide use
// evaluates to the same expression as the extended narrow use, but doesn't
// absolutely guarantee it. Hence the following failsafe check. In rare cases
// where it fails, we simply throw away the newly created wide use.
if (WideAddRec.first != SE->getSCEV(WideUse)) {
LLVM_DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse << ": "
<< *SE->getSCEV(WideUse) << " != " << *WideAddRec.first
<< "\n");
DeadInsts.emplace_back(WideUse);
return nullptr;
};

// This user does not evaluate to a recurrence after widening, so don't
// follow it. Instead insert a Trunc to kill off the original use,
// eventually isolating the original narrow IV so it can be removed.
truncateIVUse(DU, DT, LI);
return nullptr;
}
// if we reached this point then we are going to replace
// DU.NarrowUse with WideUse. Reattach DbgValue then.
replaceAllDbgUsesWith(*DU.NarrowUse, *WideUse, *WideUse, *DT);

// Reuse the IV increment that SCEVExpander created as long as it dominates
// NarrowUse.
Instruction *WideUse = nullptr;
if (WideAddRec.first == WideIncExpr &&
Rewriter.hoistIVInc(WideInc, DU.NarrowUse))
WideUse = WideInc;
else {
WideUse = cloneIVUser(DU, WideAddRec.first);
if (!WideUse)
return nullptr;
}
// Evaluation of WideAddRec ensured that the narrow expression could be
// extended outside the loop without overflow. This suggests that the wide use
// evaluates to the same expression as the extended narrow use, but doesn't
// absolutely guarantee it. Hence the following failsafe check. In rare cases
// where it fails, we simply throw away the newly created wide use.
if (WideAddRec.first != SE->getSCEV(WideUse)) {
LLVM_DEBUG(dbgs() << "Wide use expression mismatch: " << *WideUse << ": "
<< *SE->getSCEV(WideUse) << " != " << *WideAddRec.first
<< "\n");
DeadInsts.emplace_back(WideUse);
ExtendKindMap[DU.NarrowUse] = WideAddRec.second;
// Returning WideUse pushes it on the worklist.
return WideUse;
};

if (auto *I = tryAddRecExpansion())
return I;

// If use is a loop condition, try to promote the condition instead of
// truncating the IV first.
if (widenLoopCompare(DU))
return nullptr;
}

// if we reached this point then we are going to replace
// DU.NarrowUse with WideUse. Reattach DbgValue then.
replaceAllDbgUsesWith(*DU.NarrowUse, *WideUse, *WideUse, *DT);
// We are here about to generate a truncate instruction that may hurt
// performance because the scalar evolution expression computed earlier
// in WideAddRec.first does not indicate a polynomial induction expression.
// In that case, look at the operands of the use instruction to determine
// if we can still widen the use instead of truncating its operand.
if (widenWithVariantUse(DU))
return nullptr;

ExtendKindMap[DU.NarrowUse] = WideAddRec.second;
// Returning WideUse pushes it on the worklist.
return WideUse;
// This user does not evaluate to a recurrence after widening, so don't
// follow it. Instead insert a Trunc to kill off the original use,
// eventually isolating the original narrow IV so it can be removed.
truncateIVUse(DU, DT, LI);
return nullptr;
}

/// Add eligible users of NarrowDef to NarrowIVUsers.
Expand Down
18 changes: 8 additions & 10 deletions llvm/test/Transforms/IndVarSimplify/pr55925.ll
View file
Original file line number Diff line number Diff line change
Expand Up @@ -14,14 +14,13 @@ define void @test(ptr %p) personality ptr undef {
; CHECK-NEXT: br label [[LOOP:%.*]]
; CHECK: loop:
; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[LOOP_LATCH:%.*]] ], [ 0, [[ENTRY:%.*]] ]
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[IV_NEXT:%.*]], [[LOOP_LATCH]] ]
; CHECK-NEXT: [[RES:%.*]] = invoke i32 @foo(i32 returned [[IV]])
; CHECK-NEXT: [[TMP0:%.*]] = trunc i64 [[INDVARS_IV]] to i32
; CHECK-NEXT: [[RES:%.*]] = invoke i32 @foo(i32 returned [[TMP0]])
; CHECK-NEXT: to label [[LOOP_LATCH]] unwind label [[EXIT:%.*]]
; CHECK: loop.latch:
; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 1
; CHECK-NEXT: [[IV_NEXT]] = add nuw i32 [[IV]], 1
; CHECK-NEXT: [[TMP0:%.*]] = trunc i64 [[INDVARS_IV]] to i32
; CHECK-NEXT: [[TMP1:%.*]] = call i32 @foo(i32 [[TMP0]])
; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[INDVARS_IV]] to i32
; CHECK-NEXT: [[TMP2:%.*]] = call i32 @foo(i32 [[TMP1]])
; CHECK-NEXT: br label [[LOOP]]
; CHECK: exit:
; CHECK-NEXT: [[LP:%.*]] = landingpad { ptr, i32 }
Expand Down Expand Up @@ -55,19 +54,18 @@ define void @test_critedge(i1 %c, ptr %p) personality ptr undef {
; CHECK-NEXT: br label [[LOOP:%.*]]
; CHECK: loop:
; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[INDVARS_IV_NEXT:%.*]], [[LOOP_LATCH:%.*]] ], [ 0, [[ENTRY:%.*]] ]
; CHECK-NEXT: [[IV:%.*]] = phi i32 [ 0, [[ENTRY]] ], [ [[IV_NEXT:%.*]], [[LOOP_LATCH]] ]
; CHECK-NEXT: br i1 [[C:%.*]], label [[LOOP_INVOKE:%.*]], label [[LOOP_OTHER:%.*]]
; CHECK: loop.invoke:
; CHECK-NEXT: [[TMP0:%.*]] = trunc i64 [[INDVARS_IV]] to i32
; CHECK-NEXT: [[RES:%.*]] = invoke i32 @foo(i32 returned [[IV]])
; CHECK-NEXT: [[TMP1:%.*]] = trunc i64 [[INDVARS_IV]] to i32
; CHECK-NEXT: [[RES:%.*]] = invoke i32 @foo(i32 returned [[TMP0]])
; CHECK-NEXT: to label [[LOOP_LATCH]] unwind label [[EXIT:%.*]]
; CHECK: loop.other:
; CHECK-NEXT: br label [[LOOP_LATCH]]
; CHECK: loop.latch:
; CHECK-NEXT: [[PHI:%.*]] = phi i32 [ [[TMP0]], [[LOOP_INVOKE]] ], [ 0, [[LOOP_OTHER]] ]
; CHECK-NEXT: [[PHI:%.*]] = phi i32 [ [[TMP1]], [[LOOP_INVOKE]] ], [ 0, [[LOOP_OTHER]] ]
; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add i64 [[INDVARS_IV]], 1
; CHECK-NEXT: [[IV_NEXT]] = add nuw i32 [[IV]], 1
; CHECK-NEXT: [[TMP1:%.*]] = call i32 @foo(i32 [[PHI]])
; CHECK-NEXT: [[TMP2:%.*]] = call i32 @foo(i32 [[PHI]])
; CHECK-NEXT: br label [[LOOP]]
; CHECK: exit:
; CHECK-NEXT: [[LP:%.*]] = landingpad { ptr, i32 }
Expand Down
15 changes: 7 additions & 8 deletions llvm/test/Transforms/IndVarSimplify/widen-nonnegative-countdown.ll
View file
Original file line number Diff line number Diff line change
Expand Up @@ -660,15 +660,17 @@ define void @zext_postinc_offset_constant_minus_one(ptr %A, i32 %start) {
; CHECK-NEXT: [[NONPOS:%.*]] = icmp slt i32 [[START:%.*]], 2
; CHECK-NEXT: br i1 [[NONPOS]], label [[EXIT:%.*]], label [[FOR_BODY_PREHEADER:%.*]]
; CHECK: for.body.preheader:
; CHECK-NEXT: [[TMP0:%.*]] = zext i32 [[START]] to i64
; CHECK-NEXT: br label [[FOR_BODY:%.*]]
; CHECK: for.body:
; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[TMP0]], [[FOR_BODY_PREHEADER]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; CHECK-NEXT: [[J_016_US:%.*]] = phi i32 [ [[INC_US:%.*]], [[FOR_BODY]] ], [ [[START]], [[FOR_BODY_PREHEADER]] ]
; CHECK-NEXT: [[ADD_US:%.*]] = add i32 [[J_016_US]], -1
; CHECK-NEXT: [[IDXPROM_US:%.*]] = zext i32 [[ADD_US]] to i64
; CHECK-NEXT: [[ARRAYIDX_US:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[IDXPROM_US]]
; CHECK-NEXT: [[TMP1:%.*]] = add nsw i64 [[INDVARS_IV]], -1
; CHECK-NEXT: [[ARRAYIDX_US:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP1]]
; CHECK-NEXT: tail call void @use_ptr(ptr [[ARRAYIDX_US]])
; CHECK-NEXT: [[INC_US]] = add nsw i32 [[J_016_US]], -1
; CHECK-NEXT: [[CMP2_US:%.*]] = icmp ugt i32 [[INC_US]], 6
; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nsw i64 [[INDVARS_IV]], -1
; CHECK-NEXT: br i1 [[CMP2_US]], label [[FOR_BODY]], label [[EXIT_LOOPEXIT:%.*]]
; CHECK: exit.loopexit:
; CHECK-NEXT: br label [[EXIT]]
Expand Down Expand Up @@ -704,12 +706,9 @@ define void @zext_preinc_offset_constant_minus_one(ptr %A, i32 %start) {
; CHECK-NEXT: br label [[FOR_BODY:%.*]]
; CHECK: for.body:
; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[TMP0]], [[FOR_BODY_PREHEADER]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; CHECK-NEXT: [[J_016_US:%.*]] = phi i32 [ [[INC_US:%.*]], [[FOR_BODY]] ], [ [[START]], [[FOR_BODY_PREHEADER]] ]
; CHECK-NEXT: [[ADD_US:%.*]] = add i32 [[J_016_US]], -1
; CHECK-NEXT: [[IDXPROM_US:%.*]] = zext i32 [[ADD_US]] to i64
; CHECK-NEXT: [[ARRAYIDX_US:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[IDXPROM_US]]
; CHECK-NEXT: [[TMP1:%.*]] = add nsw i64 [[INDVARS_IV]], -1
; CHECK-NEXT: [[ARRAYIDX_US:%.*]] = getelementptr inbounds i32, ptr [[A:%.*]], i64 [[TMP1]]
; CHECK-NEXT: tail call void @use_ptr(ptr [[ARRAYIDX_US]])
; CHECK-NEXT: [[INC_US]] = add nsw i32 [[J_016_US]], -1
; CHECK-NEXT: [[CMP2_US:%.*]] = icmp ugt i64 [[INDVARS_IV]], 6
; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nsw i64 [[INDVARS_IV]], -1
; CHECK-NEXT: br i1 [[CMP2_US]], label [[FOR_BODY]], label [[EXIT_LOOPEXIT:%.*]]
Expand Down

0 comments on commit 551c280

Please sign in to comment.