[llvm] r271929 - Reapply [LSR] Create fewer redundant instructions.
Geoff Berry via llvm-commits
llvm-commits at lists.llvm.org
Mon Jun 6 12:10:46 PDT 2016
Author: gberry
Date: Mon Jun 6 14:10:46 2016
New Revision: 271929
URL: http://llvm.org/viewvc/llvm-project?rev=271929&view=rev
Log:
Reapply [LSR] Create fewer redundant instructions.
Summary:
Fix LSRInstance::HoistInsertPosition() to check the original insert
position block first for a canonical insertion point that is dominated
by all inputs. This leads to SCEV being able to reuse more instructions
since it currently tracks the instructions it creates for reuse by
keeping a table of <Value, insert point> pairs.
Originally reviewed in http://reviews.llvm.org/D18001
Reviewers: atrick
Subscribers: llvm-commits, mzolotukhin, mcrosier
Differential Revision: http://reviews.llvm.org/D18480
Added:
llvm/trunk/test/Transforms/LoopStrengthReduce/AArch64/lsr-reuse.ll
llvm/trunk/test/Transforms/LoopStrengthReduce/scev-insertpt-bug.ll
Modified:
llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp
Modified: llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp?rev=271929&r1=271928&r2=271929&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp (original)
+++ llvm/trunk/lib/Transforms/Scalar/LoopStrengthReduce.cpp Mon Jun 6 14:10:46 2016
@@ -4315,28 +4315,10 @@ BasicBlock::iterator
LSRInstance::HoistInsertPosition(BasicBlock::iterator IP,
const SmallVectorImpl<Instruction *> &Inputs)
const {
+ Instruction *Tentative = &*IP;
for (;;) {
- const Loop *IPLoop = LI.getLoopFor(IP->getParent());
- unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
-
- BasicBlock *IDom;
- for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
- if (!Rung) return IP;
- Rung = Rung->getIDom();
- if (!Rung) return IP;
- IDom = Rung->getBlock();
-
- // Don't climb into a loop though.
- const Loop *IDomLoop = LI.getLoopFor(IDom);
- unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
- if (IDomDepth <= IPLoopDepth &&
- (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
- break;
- }
-
bool AllDominate = true;
Instruction *BetterPos = nullptr;
- Instruction *Tentative = IDom->getTerminator();
// Don't bother attempting to insert before a catchswitch, their basic block
// cannot have other non-PHI instructions.
if (isa<CatchSwitchInst>(Tentative))
@@ -4349,7 +4331,7 @@ LSRInstance::HoistInsertPosition(BasicBl
}
// Attempt to find an insert position in the middle of the block,
// instead of at the end, so that it can be used for other expansions.
- if (IDom == Inst->getParent() &&
+ if (Tentative->getParent() == Inst->getParent() &&
(!BetterPos || !DT.dominates(Inst, BetterPos)))
BetterPos = &*std::next(BasicBlock::iterator(Inst));
}
@@ -4359,6 +4341,26 @@ LSRInstance::HoistInsertPosition(BasicBl
IP = BetterPos->getIterator();
else
IP = Tentative->getIterator();
+
+ const Loop *IPLoop = LI.getLoopFor(IP->getParent());
+ unsigned IPLoopDepth = IPLoop ? IPLoop->getLoopDepth() : 0;
+
+ BasicBlock *IDom;
+ for (DomTreeNode *Rung = DT.getNode(IP->getParent()); ; ) {
+ if (!Rung) return IP;
+ Rung = Rung->getIDom();
+ if (!Rung) return IP;
+ IDom = Rung->getBlock();
+
+ // Don't climb into a loop though.
+ const Loop *IDomLoop = LI.getLoopFor(IDom);
+ unsigned IDomDepth = IDomLoop ? IDomLoop->getLoopDepth() : 0;
+ if (IDomDepth <= IPLoopDepth &&
+ (IDomDepth != IPLoopDepth || IDomLoop == IPLoop))
+ break;
+ }
+
+ Tentative = IDom->getTerminator();
}
return IP;
Added: llvm/trunk/test/Transforms/LoopStrengthReduce/AArch64/lsr-reuse.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/LoopStrengthReduce/AArch64/lsr-reuse.ll?rev=271929&view=auto
==============================================================================
--- llvm/trunk/test/Transforms/LoopStrengthReduce/AArch64/lsr-reuse.ll (added)
+++ llvm/trunk/test/Transforms/LoopStrengthReduce/AArch64/lsr-reuse.ll Mon Jun 6 14:10:46 2016
@@ -0,0 +1,34 @@
+; RUN: llc -mtriple=arm64-unknown-unknown -print-lsr-output < %s 2>&1 | FileCheck %s
+
+declare void @foo(i64)
+
+; Verify that redundant adds aren't inserted by LSR.
+; CHECK-LABEL: @bar(
+define void @bar(double* %A) {
+entry:
+ br label %while.cond
+
+while.cond:
+; CHECK-LABEL: while.cond:
+; CHECK: add i64 %lsr.iv, 1
+; CHECK-NOT: add i64 %lsr.iv, 1
+; CHECK-LABEL: land.rhs:
+ %indvars.iv28 = phi i64 [ %indvars.iv.next29, %land.rhs ], [ 50, %entry ]
+ %cmp = icmp sgt i64 %indvars.iv28, 0
+ br i1 %cmp, label %land.rhs, label %while.end
+
+land.rhs:
+ %indvars.iv.next29 = add nsw i64 %indvars.iv28, -1
+ %arrayidx = getelementptr inbounds double, double* %A, i64 %indvars.iv.next29
+ %Aload = load double, double* %arrayidx, align 8
+ %cmp1 = fcmp oeq double %Aload, 0.000000e+00
+ br i1 %cmp1, label %while.cond, label %if.end
+
+while.end:
+ %indvars.iv28.lcssa = phi i64 [ %indvars.iv28, %while.cond ]
+ tail call void @foo(i64 %indvars.iv28.lcssa)
+ br label %if.end
+
+if.end:
+ ret void
+}
Added: llvm/trunk/test/Transforms/LoopStrengthReduce/scev-insertpt-bug.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/LoopStrengthReduce/scev-insertpt-bug.ll?rev=271929&view=auto
==============================================================================
--- llvm/trunk/test/Transforms/LoopStrengthReduce/scev-insertpt-bug.ll (added)
+++ llvm/trunk/test/Transforms/LoopStrengthReduce/scev-insertpt-bug.ll Mon Jun 6 14:10:46 2016
@@ -0,0 +1,47 @@
+; RUN: opt < %s -loop-reduce -S
+
+; Test that SCEV insertpoint's don't get corrupted and cause an
+; invalid instruction to be inserted in a block other than its parent.
+; See http://reviews.llvm.org/D20703 for context.
+define void @test() {
+entry:
+ %bf.load = load i32, i32* null, align 4
+ %bf.clear = lshr i32 %bf.load, 1
+ %div = and i32 %bf.clear, 134217727
+ %sub = add nsw i32 %div, -1
+ %0 = zext i32 %sub to i64
+ br label %while.cond
+
+while.cond: ; preds = %cond.end, %entry
+ %indvars.iv = phi i64 [ %indvars.iv.next, %cond.end ], [ 0, %entry ]
+ %cmp = icmp eq i64 %indvars.iv, %0
+ br i1 %cmp, label %cleanup16, label %while.body
+
+while.body: ; preds = %while.cond
+ %1 = trunc i64 %indvars.iv to i32
+ %mul = shl i32 %1, 1
+ %add = add nuw i32 %mul, 2
+ %cmp3 = icmp ult i32 %add, 0
+ br i1 %cmp3, label %if.end, label %if.then
+
+if.then: ; preds = %while.body
+ unreachable
+
+if.end: ; preds = %while.body
+ br i1 false, label %cond.end, label %cond.true
+
+cond.true: ; preds = %if.end
+ br label %cond.end
+
+cond.end: ; preds = %cond.true, %if.end
+ %add7 = add i32 %1, 1
+ %cmp12 = icmp ugt i32 %add7, %sub
+ %indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
+ br i1 %cmp12, label %if.then13, label %while.cond
+
+if.then13: ; preds = %cond.end
+ unreachable
+
+cleanup16: ; preds = %while.cond
+ ret void
+}
More information about the llvm-commits
mailing list