[PATCH] D37619: [LV] Fix PR34523 - avoid generating redundant selects

[Ayal Zaks via Phabricator via llvm-commits]

Ayal created this revision.

When converting a PHI into a series of 'select' instructions to combine the
incoming values together according their edge masks, initialize the first
value to the incoming value In0 of the first predecessor, instead of
generating a redundant identity 'select(Cond[0], In0, In0)'. The latter
fails when the Cond[0] mask of the first predecessor is null, representing
a full mask, which can happen only when there's a single incoming value.

Added the testcase from the PR.

This fix follows https://reviews.llvm.org/D35725, which introduced using null to represent full masks.


https://reviews.llvm.org/D37619

Files:
  lib/Transforms/Vectorize/LoopVectorize.cpp
  test/Transforms/LoopVectorize/if-conversion.ll


Index: test/Transforms/LoopVectorize/if-conversion.ll
===================================================================
--- test/Transforms/LoopVectorize/if-conversion.ll
+++ test/Transforms/LoopVectorize/if-conversion.ll
@@ -168,3 +168,30 @@
 for.end:
   ret i32 %or
 }
+
+; Handle PHI with single incoming value having a full mask.
+; PR34523
+
+; CHECK-LABEL: PR34523
+; CHECK: vector.body
+
+define void @PR34523() {
+bb1:
+  br label %bb2
+
+bb2:                                             ; preds = %bb4, %bb1
+  %i = phi i16 [ undef, %bb1 ], [ %_tmp2, %bb4 ]
+  br label %bb3
+
+bb3:                                             ; preds = %bb2
+  %_tmp1 = phi [1 x [1 x i32]]* [ undef, %bb2 ]
+  br label %bb4
+
+bb4:                                             ; preds = %bb3
+  %_tmp2 = add i16 %i, 1
+  %_tmp3 = icmp slt i16 %_tmp2, 2
+  br i1 %_tmp3, label %bb2, label %bb5
+
+bb5:                                             ; preds = %bb4
+  unreachable
+}
Index: lib/Transforms/Vectorize/LoopVectorize.cpp
===================================================================
--- lib/Transforms/Vectorize/LoopVectorize.cpp
+++ lib/Transforms/Vectorize/LoopVectorize.cpp
@@ -4526,10 +4526,10 @@
 
       for (unsigned Part = 0; Part < UF; ++Part) {
         Value *In0 = getOrCreateVectorValue(P->getIncomingValue(In), Part);
-        // We might have single edge PHIs (blocks) - use an identity
-        // 'select' for the first PHI operand.
+        assert((Cond[Part] || NumIncoming == 1) &&
+               "Multiple predecessors with one predecessor having a full mask");
         if (In == 0)
-          Entry[Part] = Builder.CreateSelect(Cond[Part], In0, In0);
+          Entry[Part] = In0; // Initialize with the first incoming value.
         else
           // Select between the current value and the previous incoming edge
           // based on the incoming mask.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D37619.114335.patch
Type: text/x-patch
Size: 1891 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170908/6abb1517/attachment.bin>