[LV] Consider users that are memory accesses in uniforms expansion step

When expanding the set of uniform instructions beyond the seed instructions
(e.g., consecutive pointers), we mark a new instruction uniform if all its
loop-varying users are uniform. We should also allow users that are consecutive
or interleaved memory accesses. This fixes cases where we have an instruction
that is used as the pointer operand of a consecutive access but also used by a
non-memory instruction that later becomes uniform as part of the expansion.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@297179 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/Vectorize/LoopVectorize.cpp b/lib/Transforms/Vectorize/LoopVectorize.cpp
index 763ce083837753f0fbaadd402f0b8f23052fb729..836a38d98138720db216e2b59aae6b280b7c602b 100644 (file)
--- a/lib/Transforms/Vectorize/LoopVectorize.cpp
+++ b/lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -5675,7 +5675,9 @@ void LoopVectorizationCostModel::collectLoopUniforms(unsigned VF) {
         continue;
       auto *OI = cast<Instruction>(OV);
       if (all_of(OI->users(), [&](User *U) -> bool {
-            return isOutOfScope(U) || Worklist.count(cast<Instruction>(U));
+            auto *J = cast<Instruction>(U);
+            return !TheLoop->contains(J) || Worklist.count(J) ||
+                   (OI == getPointerOperand(J) && isUniformDecision(J, VF));
           })) {
         Worklist.insert(OI);
         DEBUG(dbgs() << "LV: Found uniform instruction: " << *OI << "\n");

diff --git a/test/Transforms/LoopVectorize/consecutive-ptr-uniforms.ll b/test/Transforms/LoopVectorize/consecutive-ptr-uniforms.ll
index 88b2aa36b08c40722e86644351752906b6240ab0..125829090c3f2dd6ee4d5d6c5482d32fd2bc359d 100644 (file)
--- a/test/Transforms/LoopVectorize/consecutive-ptr-uniforms.ll
+++ b/test/Transforms/LoopVectorize/consecutive-ptr-uniforms.ll
@@ -438,3 +438,53 @@ for.end:
   %tmp5 = phi i32 [ %tmp2, %for.body ]
   ret i32 %tmp5
 }
+
+; INTER-LABEL: bitcast_pointer_operand
+;
+; Check that a pointer operand having a user other than a memory access is
+; recognized as uniform after vectorization. In this test case, %tmp1 is a
+; bitcast that is used by a load and a getelementptr instruction (%tmp2). Once
+; %tmp2 is marked uniform, %tmp1 should be marked uniform as well.
+;
+; INTER:       LV: Found uniform instruction: %cond = icmp slt i64 %i.next, %n
+; INTER-NEXT:  LV: Found uniform instruction: %tmp2 = getelementptr inbounds i8, i8* %tmp1, i64 3
+; INTER-NEXT:  LV: Found uniform instruction: %tmp6 = getelementptr inbounds i8, i8* %B, i64 %i
+; INTER-NEXT:  LV: Found uniform instruction: %tmp1 = bitcast i64* %tmp0 to i8*
+; INTER-NEXT:  LV: Found uniform instruction: %tmp0 = getelementptr inbounds i64, i64* %A, i64 %i
+; INTER-NEXT:  LV: Found uniform instruction: %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+; INTER-NEXT:  LV: Found uniform instruction: %i.next = add nuw nsw i64 %i, 1
+; INTER:       vector.body:
+; INTER-NEXT:    [[INDEX:%.*]] = phi i64 [ 0, %vector.ph ], [ [[INDEX_NEXT:%.*]], %vector.body ]
+; INTER-NEXT:    [[TMP4:%.*]] = getelementptr inbounds i64, i64* %A, i64 [[INDEX]]
+; INTER-NEXT:    [[TMP5:%.*]] = bitcast i64* [[TMP4]] to <32 x i8>*
+; INTER-NEXT:    [[WIDE_VEC:%.*]] = load <32 x i8>, <32 x i8>* [[TMP5]], align 1
+; INTER-NEXT:    [[STRIDED_VEC:%.*]] = shufflevector <32 x i8> [[WIDE_VEC]], <32 x i8> undef, <4 x i32> <i32 0, i32 8, i32 16, i32 24>
+; INTER-NEXT:    [[STRIDED_VEC5:%.*]] = shufflevector <32 x i8> [[WIDE_VEC]], <32 x i8> undef, <4 x i32> <i32 3, i32 11, i32 19, i32 27>
+; INTER-NEXT:    [[TMP6:%.*]] = xor <4 x i8> [[STRIDED_VEC5]], [[STRIDED_VEC]]
+; INTER-NEXT:    [[TMP7:%.*]] = getelementptr inbounds i8, i8* %B, i64 [[INDEX]]
+; INTER-NEXT:    [[TMP8:%.*]] = bitcast i8* [[TMP7]] to <4 x i8>*
+; INTER-NEXT:    store <4 x i8> [[TMP6]], <4 x i8>* [[TMP8]], align 1
+; INTER-NEXT:    [[INDEX_NEXT]] = add i64 [[INDEX]], 4
+; INTER:         br i1 {{.*}}, label %middle.block, label %vector.body
+;
+define void @bitcast_pointer_operand(i64* %A, i8* %B, i64 %n) {
+entry:
+  br label %for.body
+
+for.body:
+  %i = phi i64 [ 0, %entry ], [ %i.next, %for.body ]
+  %tmp0 = getelementptr inbounds i64, i64* %A, i64 %i
+  %tmp1 = bitcast i64* %tmp0 to i8*
+  %tmp2 = getelementptr inbounds i8, i8* %tmp1, i64 3
+  %tmp3 = load i8, i8* %tmp2, align 1
+  %tmp4 = load i8, i8* %tmp1, align 1
+  %tmp5 = xor i8 %tmp3, %tmp4
+  %tmp6 = getelementptr inbounds i8, i8* %B, i64 %i
+  store i8 %tmp5, i8* %tmp6
+  %i.next = add nuw nsw i64 %i, 1
+  %cond = icmp slt i64 %i.next, %n
+  br i1 %cond, label %for.body, label %for.end
+
+for.end:
+  ret void
+}