Teach InstCombine to optimize extract of a value from a vector add operation with...

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@172576 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
index dd7ea14e8a8989c99f35ac2ebe33323e050048e9..8bfcc805bc8c52037948ea3abd36bd0ef6accf7b 100644 (file)
--- a/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
+++ b/lib/Transforms/InstCombine/InstCombineVectorOps.cpp
@@ -13,7 +13,9 @@
 //===----------------------------------------------------------------------===//
 
 #include "InstCombine.h"
+#include "llvm/Support/PatternMatch.h"
 using namespace llvm;
+using namespace PatternMatch;
 
 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
 /// is to leave as a vector operation.  isConstant indicates whether we're
@@ -92,6 +94,13 @@ static Value *FindScalarElement(Value *V, unsigned EltNo) {
     return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth);
   }
 
+  // Extract a value from a vector add operation with a constant zero.
+  Value *Val = 0; Constant *Con = 0;
+  if (match(V, m_Add(m_Value(Val), m_Constant(Con)))) {
+    if (Con->getAggregateElement(EltNo)->isNullValue())
+      return FindScalarElement(Val, EltNo);
+  }
+
   // Otherwise, we don't know.
   return 0;
 }

diff --git a/test/Transforms/InstCombine/vec_extract_elt.ll b/test/Transforms/InstCombine/vec_extract_elt.ll
index 63e4ee2112d819e247d16cbdb1be8e878779792f..166066a201bf6b0ebd89dc98a6e09cea4b5c16b6 100644 (file)
--- a/test/Transforms/InstCombine/vec_extract_elt.ll
+++ b/test/Transforms/InstCombine/vec_extract_elt.ll
@@ -7,3 +7,13 @@ define i32 @test(float %f) {
         ret i32 %tmp19
 }
 
+define i64 @test2(i64 %in) {
+  %vec = insertelement <8 x i64> undef, i64 %in, i32 0
+  %splat = shufflevector <8 x i64> %vec, <8 x i64> undef, <8 x i32> zeroinitializer
+  %add = add <8 x i64> %splat, <i64 0, i64 1, i64 2, i64 3, i64 4, i64 5, i64 6, i64 7>
+  %scl1 = extractelement <8 x i64> %add, i32 0
+  %scl2 = extractelement <8 x i64> %add, i32 0
+  %r = add i64 %scl1, %scl2
+  ret i64 %r
+}
+