[ScopInfo] Make memset etc. affine where possible.

[efriedma]

Summary

We don't actually check whether a MemoryAccess is affine in very many places, but one important one is in checks for aliasing.

[efriedma]

Delete outdated comment.

[jdoerfert]

While I agree with the commit message, I am not aware of any real benefit this will provide us given that we only leverage the constant range ScalarEvolution range analysis. If we would use a better one, e.g., along the lines of https://reviews.llvm.org/D20230, I think this patch does make sense.

This opinion was basically written down in the code:

// Check if we have non-affine accesses left, if so bail out as we cannot
// generate a good access range yet.

[efriedma]

We ran into a case similar to the included testcase, where we need to generate a runtime alias check with a memset in order to vectorize a loop. (memset isn't really non-affine, but we currently model it that way.)

[jdoerfert]

In D25706#573332, @efriedma wrote:

We ran into a case similar to the included testcase, where we need to generate a runtime alias check with a memset in order to vectorize a loop. (memset isn't really non-affine, but we currently model it that way.)

Why don't we then treat a memset/memcpy/memmove with affine length as affine access instead?
E.g. replace the "false" in ScopBuilder.cpp line 279 and 300 with "LengthIsAffine"?

[efriedma]

I assumed there was some reason we were treating the memset as non-affine... if there isn't, I'm happy to change it.

[jdoerfert]

You also have to adjust the call location to

buildMemIntrinsicAccessRelation

but I think this is the best way to go.

[efriedma]

Ping.

[efriedma]

Ping.

[jdoerfert]

LGTM.

lib/Analysis/ScopInfo.cpp

for (unsigned u = 0, e = isl_set_n_param(Set); u < e; u++)
         InvolvedParams++;
 
     if (InvolvedParams > RunTimeChecksMaxParameters) {
       isl_set_free(Set);
       return isl_stat_error;
     }
   }
 
+  for (int i = 0, e = isl_set_dim(Set, isl_dim_out); i != e; ++i) {
+    if (!isl_set_dim_has_lower_bound(Set, isl_dim_out, i) ||
+        !isl_set_dim_has_upper_bound(Set, isl_dim_out, i)) {
+      DEBUG(dbgs() << "Non-affine set with unknown bounds\n");
+      isl_set_free(Set);
+      return isl_stat_error;
+    }
+  }
+
   MinPMA = isl_set_lexmin_pw_multi_aff(isl_set_copy(Set));
   MaxPMA = isl_set_lexmax_pw_multi_aff(isl_set_copy(Set));
 
   MinPMA = isl_pw_multi_aff_coalesce(MinPMA);
   MaxPMA = isl_pw_multi_aff_coalesce(MaxPMA);
 
   // Adjust the last dimension of the maximal access by one as we want to
   // enclose the accessed memory region by MinPMA and MaxPMA. The pointer
for (AliasGroupTy &AG : AliasGroups) {
     if (NonReadOnlyBaseValues.empty()) {
       AG.clear();
       continue;
     }
 
     // Check if we have non-affine accesses left, if so bail out as we cannot
     // generate a good access range yet.
     for (auto *MA : AG) {
-      if (!MA->isAffine()) {
-        invalidate(ALIASING, MA->getAccessInstruction()->getDebugLoc());
-        return false;
-      }
       if (auto *BasePtrMA = lookupBasePtrAccess(MA))
         addRequiredInvariantLoad(
             cast<LoadInst>(BasePtrMA->getAccessInstruction()));
     }
     for (auto &ReadOnlyPair : ReadOnlyPairs)
       for (auto *MA : ReadOnlyPair.second) {
-        if (!MA->isAffine()) {
-          invalidate(ALIASING, MA->getAccessInstruction()->getDebugLoc());
-          return false;
-        }
         if (auto *BasePtrMA = lookupBasePtrAccess(MA))
           addRequiredInvariantLoad(
               cast<LoadInst>(BasePtrMA->getAccessInstruction()));
       }
 
     // Calculate minimal and maximal accesses for non read only accesses.
     MinMaxAliasGroups.emplace_back();
     MinMaxVectorPairTy &pair = MinMaxAliasGroups.back();

test/ScopInfo/aliasing_with_non_affine_access.ll

-; RUN: opt %loadPolly -analyze -polly-scops < %s | FileCheck %s
-; RUN: opt %loadPolly -analyze -polly-scops -pass-remarks-analysis="polly-scops" 2>&1 < %s | FileCheck %s --check-prefix=REMARK
+; RUN: opt %loadPolly -analyze -polly-ast -polly-process-unprofitable -polly-allow-nonaffine < %s | FileCheck %s
 ;
-; This test case has a non-affine access (the memset call) that aliases with
-; other accesses. Thus, we bail out.
+; @test1
+; Make sure we generate the correct aliasing check for a fixed-size memset operation.
+; CHECK: if (1 && (&MemRef_0[15] <= &MemRef_1[0] || &MemRef_1[32] <= &MemRef_0[14]))
+
+; @test2
+; Make sure we generate the correct aliasing check for a variable-size memset operation.
+; CHECK: if (1 && (&MemRef_0[15] <= &MemRef_1[0] || &MemRef_1[n] <= &MemRef_0[14]))
 ;
-; CHECK-NOT: Statements
-;
-; REMARK:        remark: <unknown>:0:0: SCoP begins here.
-; REMARK-NEXT:   remark: <unknown>:0:0: Possibly aliasing pointer, use restrict keyword.
-; REMARK-NEXT:   remark: <unknown>:0:0: Possibly aliasing pointer, use restrict keyword.
-; REMARK-NEXT:   remark: <unknown>:0:0: No-aliasing assumption: {  : 1 = 0 }
-; REMARK-NEXT:   remark: <unknown>:0:0: SCoP ends here but was dismissed.
-;
-; ModuleID = 'bugpoint-reduced-simplified.bc'
+; @test3
+; FIXME: We should reject this.  The runtime check is in some sense correct,
+; but it's impossible to satisfy.
+; CHECK: if (1 && (&MemRef_A[2305843009213693952] <= &MemRef_B[0] || &MemRef_B[n] <= &MemRef_A[-2305843009213693952]))
+
 target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
 
 %struct.info = type { i32, %struct.ctr*, i32, %struct.ord*, %struct.ctr*, i32, i8*, i32, i32, double }
 %struct.ctr = type { i32, i8, i8, i32 }
 %struct.ord = type { i32, i8 }
 
 ; Function Attrs: argmemonly nounwind
 declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1) #0
 
-; Function Attrs: nounwind uwtable
-define void @bestVirtualIndex(%struct.info** %ppIdxInfo) {
+define void @test(%struct.info** %ppIdxInfo) {
 entry:
   %0 = load %struct.info*, %struct.info** %ppIdxInfo, align 8
   br label %if.end125
 
 if.end125:                                        ; preds = %entry
   %1 = load %struct.ctr*, %struct.ctr** undef, align 8
   br label %for.end143
 
 for.end143:                                       ; preds = %if.end125
   %2 = bitcast %struct.ctr* %1 to i8*
   tail call void @llvm.memset.p0i8.i64(i8* %2, i8 0, i64 32, i32 4, i1 false)
   %needToFreeIdxStr = getelementptr inbounds %struct.info, %struct.info* %0, i64 0, i32 7
   %3 = load i32, i32* %needToFreeIdxStr, align 8
   br i1 false, label %if.end149, label %if.then148
 
 if.then148:                                       ; preds = %for.end143
   br label %if.end149
 
 if.end149:                                        ; preds = %if.then148, %for.end143
   unreachable
 }
 
+define void @test2(%struct.info** %ppIdxInfo, i64 %n) {
+entry:
+  %0 = load %struct.info*, %struct.info** %ppIdxInfo, align 8
+  br label %if.end125
+
+if.end125:                                        ; preds = %entry
+  %1 = load %struct.ctr*, %struct.ctr** undef, align 8
+  br label %for.end143
+
+for.end143:                                       ; preds = %if.end125
+  %2 = bitcast %struct.ctr* %1 to i8*
+  tail call void @llvm.memset.p0i8.i64(i8* %2, i8 0, i64 %n, i32 4, i1 false)
+  %needToFreeIdxStr = getelementptr inbounds %struct.info, %struct.info* %0, i64 0, i32 7
+  %3 = load i32, i32* %needToFreeIdxStr, align 8
+  br i1 false, label %if.end149, label %if.then148
+
+if.then148:                                       ; preds = %for.end143
+  br label %if.end149
+
+if.end149:                                        ; preds = %if.then148, %for.end143
+  unreachable
+}
+
+define void @test3(i32* nocapture readonly %A, i32* nocapture %B, i32 %n) local_unnamed_addr #0 {
+entry:
+  br label %entry.split
+
+entry.split:                                      ; preds = %entry
+  %conv = zext i32 %n to i64
+  %cmp8 = icmp eq i32 %n, 0
+  br i1 %cmp8, label %for.cond.cleanup, label %for.body.preheader
+
+for.body.preheader:                               ; preds = %entry.split
+  br label %for.body
+
+for.cond.cleanup.loopexit:                        ; preds = %for.body
+  br label %for.cond.cleanup
+
+for.cond.cleanup:                                 ; preds = %for.cond.cleanup.loopexit, %entry.split
+  ret void
+
+for.body:                                         ; preds = %for.body.preheader, %for.body
+  %i.09 = phi i64 [ %inc, %for.body ], [ 0, %for.body.preheader ]
+  %mul = mul nsw i64 %i.09, %i.09
+  %arrayidx = getelementptr inbounds i32, i32* %A, i64 %mul
+  %0 = load i32, i32* %arrayidx, align 4
+  %arrayidx2 = getelementptr inbounds i32, i32* %B, i64 %i.09
+  %1 = load i32, i32* %arrayidx2, align 4
+  %add = add nsw i32 %1, %0
+  store i32 %add, i32* %arrayidx2, align 4
+  %inc = add nuw nsw i64 %i.09, 1
+  %exitcond = icmp eq i64 %inc, %conv
+  br i1 %exitcond, label %for.cond.cleanup.loopexit, label %for.body
+}
+
+
 attributes #0 = { argmemonly nounwind }