[PATCH] D11149: Fix memcheck interval ends for pointers with negative strides

silviu.baranga at arm.com silviu.baranga at arm.com
Mon Jul 13 08:47:47 PDT 2015


sbaranga created this revision.
sbaranga added a reviewer: anemet.
sbaranga added a subscriber: llvm-commits.

The checking pointer grouping algorithm assumes that the
starts/ends of the pointers are well formed (start <= end).

The runtime memory checking algorithm also assumes this by doing:

 start0 < end1 && start1 < end0

to detect conflicts. This check only works if start0 <= end0 and
start1 <= end1.

This change correctly orders the interval ends by either checking
the stride (if it is constant) or by using min/max SCEV expressions.

http://reviews.llvm.org/D11149

Files:
  lib/Analysis/LoopAccessAnalysis.cpp
  test/Analysis/LoopAccessAnalysis/reverse-memcheck-bounds.ll

Index: lib/Analysis/LoopAccessAnalysis.cpp
===================================================================
--- lib/Analysis/LoopAccessAnalysis.cpp
+++ lib/Analysis/LoopAccessAnalysis.cpp
@@ -127,9 +127,28 @@
   const SCEVAddRecExpr *AR = dyn_cast<SCEVAddRecExpr>(Sc);
   assert(AR && "Invalid addrec expression");
   const SCEV *Ex = SE->getBackedgeTakenCount(Lp);
+
+  const SCEV *ScStart = AR->getStart();
   const SCEV *ScEnd = AR->evaluateAtIteration(Ex, *SE);
+  const SCEV *Step = AR->getStepRecurrence(*SE);
+
+  // For expressions with negative step, the upper bound is ScStart and the
+  // lower bound is ScEnd.
+  if (const SCEVConstant *CStep = dyn_cast<const SCEVConstant>(Step)) {
+    if (CStep->getValue()->isNegative())
+      std::swap(ScStart, ScEnd);
+  } else {
+    // Fallback case: the step is not constant, but the we can still
+    // get the upper and lower bounds of the interval by using min/max
+    // expressions.
+    const SCEV *Min = SE->getUMinExpr(ScStart, ScEnd);
+    const SCEV *Max = SE->getUMaxExpr(ScStart, ScEnd);
+    ScStart = Min;
+    ScEnd = Max;
+  }
+
   Pointers.push_back(Ptr);
-  Starts.push_back(AR->getStart());
+  Starts.push_back(ScStart);
   Ends.push_back(ScEnd);
   IsWritePtr.push_back(WritePtr);
   DependencySetId.push_back(DepSetId);
Index: test/Analysis/LoopAccessAnalysis/reverse-memcheck-bounds.ll
===================================================================
--- /dev/null
+++ test/Analysis/LoopAccessAnalysis/reverse-memcheck-bounds.ll
@@ -0,0 +1,47 @@
+; RUN: opt -loop-accesses -analyze < %s | FileCheck %s
+
+; The runtime memory check code and the access grouping
+; algorithm both assume that the start and end values
+; for an access range are ordered (start <= stop).
+; When generating checks for accesses with negative stride
+; we need to take this into account and swap the interval
+; ends.
+;
+;   for (i = 0; i < 10000; i++) {
+;     B[i] = A[15000 - i] * 3;
+;   }
+
+target datalayout = "e-m:e-i64:64-i128:128-n32:64-S128"
+target triple = "aarch64--linux-gnueabi"
+
+; CHECK: (Low: (20000 + %a) High: (60000 + %a))
+
+ at B = common global i32* null, align 8
+ at A = common global i32* null, align 8
+
+define void @f() {
+entry:
+  %a = load i32*, i32** @A, align 8
+  %b = load i32*, i32** @B, align 8
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %idx = phi i64 [ 0, %entry ], [ %add, %for.body ]
+  %negidx = sub i64 15000, %idx
+
+  %arrayidxA0 = getelementptr inbounds i32, i32* %a, i64 %negidx
+  %loadA0 = load i32, i32* %arrayidxA0, align 2
+
+  %res = mul i32 %loadA0, 3
+
+  %add = add nuw nsw i64 %idx, 1
+
+  %arrayidxB = getelementptr inbounds i32, i32* %b, i64 %idx
+  store i32 %res, i32* %arrayidxB, align 2
+
+  %exitcond = icmp eq i64 %idx, 10000
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body
+  ret void
+}


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D11149.29580.patch
Type: text/x-patch
Size: 2958 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20150713/f9c076c1/attachment.bin>


More information about the llvm-commits mailing list