[llvm] 1b87882 - [LoopPredication] Rework assumes of widened conditions
Aleksandr Popov via llvm-commits
llvm-commits at lists.llvm.org
Fri Aug 18 05:36:54 PDT 2023
Author: Aleksandr Popov
Date: 2023-08-18T14:35:46+02:00
New Revision: 1b87882228b0371212753781a363d4addd1bafa7
URL: https://github.com/llvm/llvm-project/commit/1b87882228b0371212753781a363d4addd1bafa7
DIFF: https://github.com/llvm/llvm-project/commit/1b87882228b0371212753781a363d4addd1bafa7.diff
LOG: [LoopPredication] Rework assumes of widened conditions
Currently after widening br(WC && (c1 && c2)) we insert assume of
(c1 && c2) which is joined to WC by And operation.
But we are going to support more flexible form of widenable branches
where WC could be placed arbitrary in the expression tree, e.g:
br(c1 && (c2 && WC)).
In that case we won't have (c1 && c2) in the IR. So we need to add
explicit (c1 && c2) and then create an assumption of it.
Reviewed By: anna
Differential Revision: https://reviews.llvm.org/D157502
Added:
Modified:
llvm/lib/Transforms/Scalar/LoopPredication.cpp
llvm/test/Transforms/LoopPredication/basic_widenable_branch_guards.ll
Removed:
################################################################################
diff --git a/llvm/lib/Transforms/Scalar/LoopPredication.cpp b/llvm/lib/Transforms/Scalar/LoopPredication.cpp
index 79a3da9020e2e2..3491362dcd9502 100644
--- a/llvm/lib/Transforms/Scalar/LoopPredication.cpp
+++ b/llvm/lib/Transforms/Scalar/LoopPredication.cpp
@@ -307,8 +307,9 @@ class LoopPredication {
widenICmpRangeCheckDecrementingLoop(LoopICmp LatchCheck, LoopICmp RangeCheck,
SCEVExpander &Expander,
Instruction *Guard);
- unsigned widenChecks(SmallVectorImpl<Value *> &Checks, SCEVExpander &Expander,
- Instruction *Guard);
+ void widenChecks(SmallVectorImpl<Value *> &Checks,
+ SmallVectorImpl<Value *> &WidenedChecks,
+ SCEVExpander &Expander, Instruction *Guard);
bool widenGuardConditions(IntrinsicInst *II, SCEVExpander &Expander);
bool widenWidenableBranchGuardConditions(BranchInst *Guard, SCEVExpander &Expander);
// If the loop always exits through another block in the loop, we should not
@@ -754,17 +755,15 @@ LoopPredication::widenICmpRangeCheck(ICmpInst *ICI, SCEVExpander &Expander,
}
}
-unsigned LoopPredication::widenChecks(SmallVectorImpl<Value *> &Checks,
- SCEVExpander &Expander,
- Instruction *Guard) {
- unsigned NumWidened = 0;
+void LoopPredication::widenChecks(SmallVectorImpl<Value *> &Checks,
+ SmallVectorImpl<Value *> &WidenedChecks,
+ SCEVExpander &Expander, Instruction *Guard) {
for (auto &Check : Checks)
if (ICmpInst *ICI = dyn_cast<ICmpInst>(Check))
if (auto NewRangeCheck = widenICmpRangeCheck(ICI, Expander, Guard)) {
- NumWidened++;
+ WidenedChecks.push_back(Check);
Check = *NewRangeCheck;
}
- return NumWidened;
}
bool LoopPredication::widenGuardConditions(IntrinsicInst *Guard,
@@ -774,12 +773,13 @@ bool LoopPredication::widenGuardConditions(IntrinsicInst *Guard,
TotalConsidered++;
SmallVector<Value *, 4> Checks;
+ SmallVector<Value *> WidenedChecks;
parseWidenableGuard(Guard, Checks);
- unsigned NumWidened = widenChecks(Checks, Expander, Guard);
- if (NumWidened == 0)
+ widenChecks(Checks, WidenedChecks, Expander, Guard);
+ if (WidenedChecks.empty())
return false;
- TotalWidened += NumWidened;
+ TotalWidened += WidenedChecks.size();
// Emit the new guard condition
IRBuilder<> Builder(findInsertPt(Guard, Checks));
@@ -792,7 +792,7 @@ bool LoopPredication::widenGuardConditions(IntrinsicInst *Guard,
}
RecursivelyDeleteTriviallyDeadInstructions(OldCond, nullptr /* TLI */, MSSAU);
- LLVM_DEBUG(dbgs() << "Widened checks = " << NumWidened << "\n");
+ LLVM_DEBUG(dbgs() << "Widened checks = " << WidenedChecks.size() << "\n");
return true;
}
@@ -810,15 +810,16 @@ bool LoopPredication::widenWidenableBranchGuardConditions(
TotalConsidered++;
SmallVector<Value *, 4> Checks;
+ SmallVector<Value *> WidenedChecks;
parseWidenableGuard(BI, Checks);
// At the moment, our matching logic for wideable conditions implicitly
// assumes we preserve the form: (br (and Cond, WC())). FIXME
Checks.push_back(WC);
- unsigned NumWidened = widenChecks(Checks, Expander, BI);
- if (NumWidened == 0)
+ widenChecks(Checks, WidenedChecks, Expander, BI);
+ if (WidenedChecks.empty())
return false;
- TotalWidened += NumWidened;
+ TotalWidened += WidenedChecks.size();
// Emit the new guard condition
IRBuilder<> Builder(findInsertPt(BI, Checks));
@@ -830,13 +831,13 @@ bool LoopPredication::widenWidenableBranchGuardConditions(
// If this block has other predecessors, we might not be able to use Cond.
// In this case, create a Phi where every other input is `true` and input
// from guard block is Cond.
- Value *AssumeCond = Cond;
+ Value *AssumeCond = Builder.CreateAnd(WidenedChecks);
if (!IfTrueBB->getUniquePredecessor()) {
auto *GuardBB = BI->getParent();
- auto *PN = Builder.CreatePHI(Cond->getType(), pred_size(IfTrueBB),
+ auto *PN = Builder.CreatePHI(AssumeCond->getType(), pred_size(IfTrueBB),
"assume.cond");
for (auto *Pred : predecessors(IfTrueBB))
- PN->addIncoming(Pred == GuardBB ? Cond : Builder.getTrue(), Pred);
+ PN->addIncoming(Pred == GuardBB ? AssumeCond : Builder.getTrue(), Pred);
AssumeCond = PN;
}
Builder.CreateAssumption(AssumeCond);
@@ -845,7 +846,7 @@ bool LoopPredication::widenWidenableBranchGuardConditions(
assert(isGuardAsWidenableBranch(BI) &&
"Stopped being a guard after transform?");
- LLVM_DEBUG(dbgs() << "Widened checks = " << NumWidened << "\n");
+ LLVM_DEBUG(dbgs() << "Widened checks = " << WidenedChecks.size() << "\n");
return true;
}
diff --git a/llvm/test/Transforms/LoopPredication/basic_widenable_branch_guards.ll b/llvm/test/Transforms/LoopPredication/basic_widenable_branch_guards.ll
index b54e218b0cbd55..cad4fcec5c81ab 100644
--- a/llvm/test/Transforms/LoopPredication/basic_widenable_branch_guards.ll
+++ b/llvm/test/Transforms/LoopPredication/basic_widenable_branch_guards.ll
@@ -1224,7 +1224,6 @@ define i32 @two_range_checks(ptr %array.1, i32 %length.1, ptr %array.2, i32 %len
; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[I_NEXT:%.*]], [[GUARDED]] ], [ 0, [[LOOP_PREHEADER]] ]
; CHECK-NEXT: [[WITHIN_BOUNDS_1:%.*]] = icmp ult i32 [[I]], [[LENGTH_1]]
; CHECK-NEXT: [[WITHIN_BOUNDS_2:%.*]] = icmp ult i32 [[I]], [[LENGTH_2]]
-; CHECK-NEXT: [[WITHIN_BOUNDS:%.*]] = and i1 [[WITHIN_BOUNDS_1]], [[WITHIN_BOUNDS_2]]
; CHECK-NEXT: [[WIDENABLE_COND:%.*]] = call i1 @llvm.experimental.widenable.condition()
; CHECK-NEXT: [[TMP8:%.*]] = and i1 [[TMP3]], [[TMP7]]
; CHECK-NEXT: [[TMP9:%.*]] = and i1 [[TMP8]], [[WIDENABLE_COND]]
@@ -1233,7 +1232,8 @@ define i32 @two_range_checks(ptr %array.1, i32 %length.1, ptr %array.2, i32 %len
; CHECK-NEXT: [[DEOPTCALL:%.*]] = call i32 (...) @llvm.experimental.deoptimize.i32(i32 9) [ "deopt"() ]
; CHECK-NEXT: ret i32 [[DEOPTCALL]]
; CHECK: guarded:
-; CHECK-NEXT: call void @llvm.assume(i1 [[WITHIN_BOUNDS]])
+; CHECK-NEXT: [[TMP10:%.*]] = and i1 [[WITHIN_BOUNDS_2]], [[WITHIN_BOUNDS_1]]
+; CHECK-NEXT: call void @llvm.assume(i1 [[TMP10]])
; CHECK-NEXT: [[I_I64:%.*]] = zext i32 [[I]] to i64
; CHECK-NEXT: [[ARRAY_1_I_PTR:%.*]] = getelementptr inbounds i32, ptr [[ARRAY_1:%.*]], i64 [[I_I64]]
; CHECK-NEXT: [[ARRAY_1_I:%.*]] = load i32, ptr [[ARRAY_1_I_PTR]], align 4
@@ -1314,8 +1314,6 @@ define i32 @three_range_checks(ptr %array.1, i32 %length.1, ptr %array.2, i32 %l
; CHECK-NEXT: [[WITHIN_BOUNDS_1:%.*]] = icmp ult i32 [[I]], [[LENGTH_1]]
; CHECK-NEXT: [[WITHIN_BOUNDS_2:%.*]] = icmp ult i32 [[I]], [[LENGTH_2]]
; CHECK-NEXT: [[WITHIN_BOUNDS_3:%.*]] = icmp ult i32 [[I]], [[LENGTH_3]]
-; CHECK-NEXT: [[WITHIN_BOUNDS_1_AND_2:%.*]] = and i1 [[WITHIN_BOUNDS_1]], [[WITHIN_BOUNDS_2]]
-; CHECK-NEXT: [[WITHIN_BOUNDS:%.*]] = and i1 [[WITHIN_BOUNDS_1_AND_2]], [[WITHIN_BOUNDS_3]]
; CHECK-NEXT: [[WIDENABLE_COND:%.*]] = call i1 @llvm.experimental.widenable.condition()
; CHECK-NEXT: [[TMP12:%.*]] = and i1 [[TMP3]], [[TMP7]]
; CHECK-NEXT: [[TMP13:%.*]] = and i1 [[TMP12]], [[TMP11]]
@@ -1325,7 +1323,9 @@ define i32 @three_range_checks(ptr %array.1, i32 %length.1, ptr %array.2, i32 %l
; CHECK-NEXT: [[DEOPTCALL:%.*]] = call i32 (...) @llvm.experimental.deoptimize.i32(i32 9) [ "deopt"() ]
; CHECK-NEXT: ret i32 [[DEOPTCALL]]
; CHECK: guarded:
-; CHECK-NEXT: call void @llvm.assume(i1 [[WITHIN_BOUNDS]])
+; CHECK-NEXT: [[TMP15:%.*]] = and i1 [[WITHIN_BOUNDS_3]], [[WITHIN_BOUNDS_2]]
+; CHECK-NEXT: [[TMP16:%.*]] = and i1 [[TMP15]], [[WITHIN_BOUNDS_1]]
+; CHECK-NEXT: call void @llvm.assume(i1 [[TMP16]])
; CHECK-NEXT: [[I_I64:%.*]] = zext i32 [[I]] to i64
; CHECK-NEXT: [[ARRAY_1_I_PTR:%.*]] = getelementptr inbounds i32, ptr [[ARRAY_1:%.*]], i64 [[I_I64]]
; CHECK-NEXT: [[ARRAY_1_I:%.*]] = load i32, ptr [[ARRAY_1_I_PTR]], align 4
@@ -1533,7 +1533,6 @@ define i32 @unsigned_loop_0_to_n_unrelated_condition(ptr %array, i32 %length, i3
; CHECK-NEXT: [[I:%.*]] = phi i32 [ [[I_NEXT:%.*]], [[GUARDED]] ], [ 0, [[LOOP_PREHEADER]] ]
; CHECK-NEXT: [[WITHIN_BOUNDS:%.*]] = icmp ult i32 [[I]], [[LENGTH]]
; CHECK-NEXT: [[UNRELATED_COND:%.*]] = icmp ult i32 [[X:%.*]], [[LENGTH]]
-; CHECK-NEXT: [[GUARD_COND:%.*]] = and i1 [[WITHIN_BOUNDS]], [[UNRELATED_COND]]
; CHECK-NEXT: [[WIDENABLE_COND:%.*]] = call i1 @llvm.experimental.widenable.condition()
; CHECK-NEXT: [[TMP4:%.*]] = and i1 [[UNRELATED_COND]], [[TMP3]]
; CHECK-NEXT: [[TMP5:%.*]] = and i1 [[TMP4]], [[WIDENABLE_COND]]
@@ -1542,7 +1541,7 @@ define i32 @unsigned_loop_0_to_n_unrelated_condition(ptr %array, i32 %length, i3
; CHECK-NEXT: [[DEOPTCALL:%.*]] = call i32 (...) @llvm.experimental.deoptimize.i32(i32 9) [ "deopt"() ]
; CHECK-NEXT: ret i32 [[DEOPTCALL]]
; CHECK: guarded:
-; CHECK-NEXT: call void @llvm.assume(i1 [[GUARD_COND]])
+; CHECK-NEXT: call void @llvm.assume(i1 [[WITHIN_BOUNDS]])
; CHECK-NEXT: [[I_I64:%.*]] = zext i32 [[I]] to i64
; CHECK-NEXT: [[ARRAY_I_PTR:%.*]] = getelementptr inbounds i32, ptr [[ARRAY:%.*]], i64 [[I_I64]]
; CHECK-NEXT: [[ARRAY_I:%.*]] = load i32, ptr [[ARRAY_I_PTR]], align 4
More information about the llvm-commits
mailing list