[llvm] r295719 - [InstCombine] Do not exercise nested max/min pattern on abs

Anna Thomas via llvm-commits llvm-commits at lists.llvm.org
Tue Feb 21 06:40:29 PST 2017


Author: annat
Date: Tue Feb 21 08:40:28 2017
New Revision: 295719

URL: http://llvm.org/viewvc/llvm-project?rev=295719&view=rev
Log:
[InstCombine] Do not exercise nested max/min pattern on abs

Summary:
This is a fix for assertion failure in
`getInverseMinMaxSelectPattern` when ABS is passed in as a select pattern.

We should not be invoking the simplification rule for
ABS(MIN(~ x,y))) or ABS(MAX(~x,y)) combinations.

Added a test case which would cause an assertion failure without the patch.

Reviewers: sanjoy, majnemer

Subscribers: llvm-commits

Differential Revision: https://reviews.llvm.org/D30051

Modified:
    llvm/trunk/lib/Transforms/InstCombine/InstCombineSelect.cpp
    llvm/trunk/test/Transforms/InstCombine/max-of-nots.ll

Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineSelect.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineSelect.cpp?rev=295719&r1=295718&r2=295719&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/InstCombine/InstCombineSelect.cpp (original)
+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineSelect.cpp Tue Feb 21 08:40:28 2017
@@ -785,7 +785,9 @@ Instruction *InstCombiner::foldSPFofSPF(
   // This transform is performance neutral if we can elide at least one xor from
   // the set of three operands, since we'll be tacking on an xor at the very
   // end.
-  if (IsFreeOrProfitableToInvert(A, NotA, ElidesXor) &&
+  if (SelectPatternResult::isMinOrMax(SPF1) &&
+      SelectPatternResult::isMinOrMax(SPF2) &&
+      IsFreeOrProfitableToInvert(A, NotA, ElidesXor) &&
       IsFreeOrProfitableToInvert(B, NotB, ElidesXor) &&
       IsFreeOrProfitableToInvert(C, NotC, ElidesXor) && ElidesXor) {
     if (!NotA)

Modified: llvm/trunk/test/Transforms/InstCombine/max-of-nots.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/max-of-nots.ll?rev=295719&r1=295718&r2=295719&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/InstCombine/max-of-nots.ll (original)
+++ llvm/trunk/test/Transforms/InstCombine/max-of-nots.ll Tue Feb 21 08:40:28 2017
@@ -90,6 +90,28 @@ define i32 @max_of_nots(i32 %x, i32 %y)
   ret i32 %smax96
 }
 
+ ; negative test case (i.e. can not simplify) : ABS(MIN(NOT x,y))
+define i32 @abs_of_min_of_not(i32 %x, i32 %y) {
+; CHECK-LABEL: @abs_of_min_of_not(
+; CHECK-NEXT: xor
+; CHECK-NEXT: add
+; CHECK-NEXT: icmp sge
+; CHECK-NEXT: select
+; CHECK-NEXT: icmp sgt
+; CHECK-NEXT: sub
+; CHECK-NEXT: select
+; CHECK-NEXT: ret
+
+  %xord = xor i32 %x, -1
+  %yadd = add i32 %y, 2
+  %cond.i = icmp sge i32 %yadd, %xord
+  %min = select i1 %cond.i, i32 %xord, i32 %yadd
+  %cmp2 = icmp sgt i32 %min, -1
+  %sub = sub i32 0, %min
+  %abs = select i1 %cmp2, i32 %min, i32 %sub
+  ret i32  %abs
+}
+
 define <2 x i32> @max_of_nots_vec(<2 x i32> %x, <2 x i32> %y) {
 ; CHECK-LABEL: @max_of_nots_vec(
 ; CHECK-NEXT:    [[TMP1:%.*]] = icmp sgt <2 x i32> %y, zeroinitializer




More information about the llvm-commits mailing list