[llvm] r282592 - [InstSimplify] allow or-of-icmps folds with vector splat constants

Sanjay Patel via llvm-commits llvm-commits at lists.llvm.org
Wed Sep 28 07:27:21 PDT 2016 

Author: spatel
Date: Wed Sep 28 09:27:21 2016
New Revision: 282592

URL: http://llvm.org/viewvc/llvm-project?rev=282592&view=rev
Log:
[InstSimplify] allow or-of-icmps folds with vector splat constants

Modified:
    llvm/trunk/lib/Analysis/InstructionSimplify.cpp
    llvm/trunk/test/Transforms/InstSimplify/AndOrXor.ll

Modified: llvm/trunk/lib/Analysis/InstructionSimplify.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Analysis/InstructionSimplify.cpp?rev=282592&r1=282591&r2=282592&view=diff
==============================================================================
--- llvm/trunk/lib/Analysis/InstructionSimplify.cpp (original)
+++ llvm/trunk/lib/Analysis/InstructionSimplify.cpp Wed Sep 28 09:27:21 2016
@@ -1687,27 +1687,26 @@ static Value *SimplifyOrOfICmps(ICmpInst
   if (Value *X = simplifyUnsignedRangeCheck(Op0, Op1, /*IsAnd=*/false))
     return X;
 
-  // FIXME: Use m_APInt to allow vector splat matches.
+  // (icmp (add V, C0), C1) | (icmp V, C0)
   ICmpInst::Predicate Pred0, Pred1;
-  ConstantInt *CI1, *CI2;
+  const APInt *C0, *C1;
   Value *V;
-  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_ConstantInt(CI1)),
-                         m_ConstantInt(CI2))))
+  if (!match(Op0, m_ICmp(Pred0, m_Add(m_Value(V), m_APInt(C0)), m_APInt(C1))))
     return nullptr;
 
-  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Specific(CI1))))
+  if (!match(Op1, m_ICmp(Pred1, m_Specific(V), m_Value())))
     return nullptr;
 
-  Type *ITy = Op0->getType();
-
   auto *AddInst = cast<BinaryOperator>(Op0->getOperand(0));
+  if (AddInst->getOperand(1) != Op1->getOperand(1))
+    return nullptr;
+
+  Type *ITy = Op0->getType();
   bool isNSW = AddInst->hasNoSignedWrap();
   bool isNUW = AddInst->hasNoUnsignedWrap();
 
-  const APInt &CI1V = CI1->getValue();
-  const APInt &CI2V = CI2->getValue();
-  const APInt Delta = CI2V - CI1V;
-  if (CI1V.isStrictlyPositive()) {
+  const APInt Delta = *C1 - *C0;
+  if (C0->isStrictlyPositive()) {
     if (Delta == 2) {
       if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_SLE)
         return getTrue(ITy);
@@ -1721,7 +1720,7 @@ static Value *SimplifyOrOfICmps(ICmpInst
         return getTrue(ITy);
     }
   }
-  if (CI1V.getBoolValue() && isNUW) {
+  if (C0->getBoolValue() && isNUW) {
     if (Delta == 2)
       if (Pred0 == ICmpInst::ICMP_UGE && Pred1 == ICmpInst::ICMP_ULE)
         return getTrue(ITy);

Modified: llvm/trunk/test/Transforms/InstSimplify/AndOrXor.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstSimplify/AndOrXor.ll?rev=282592&r1=282591&r2=282592&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/InstSimplify/AndOrXor.ll (original)
+++ llvm/trunk/test/Transforms/InstSimplify/AndOrXor.ll Wed Sep 28 09:27:21 2016
@@ -170,8 +170,6 @@ define <2 x i1> @and_of_icmps5_vec(<2 x
   ret <2 x i1> %cmp
 }
 
-; FIXME: Vector splats should fold the same way as scalars in the next 6 pairs of tests.
-
 define i1 @or_of_icmps0(i32 %b) {
 ; CHECK-LABEL: @or_of_icmps0(
 ; CHECK-NEXT:    ret i1 true
@@ -185,11 +183,7 @@ define i1 @or_of_icmps0(i32 %b) {
 
 define <2 x i1> @or_of_icmps0_vec(<2 x i32> %b) {
 ; CHECK-LABEL: @or_of_icmps0_vec(
-; CHECK-NEXT:    [[TMP1:%.*]] = add <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp uge <2 x i32> [[TMP1]], <i32 4, i32 4>
-; CHECK-NEXT:    [[CMP3:%.*]] = icmp sle <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[CMP:%.*]] = or <2 x i1> [[TMP2]], [[CMP:%.*]]3
-; CHECK-NEXT:    ret <2 x i1> [[CMP]]
+; CHECK-NEXT:    ret <2 x i1> <i1 true, i1 true>
 ;
   %1 = add <2 x i32> %b, <i32 2, i32 2>
   %2 = icmp uge <2 x i32> %1, <i32 4, i32 4>
@@ -211,11 +205,7 @@ define i1 @or_of_icmps1(i32 %b) {
 
 define <2 x i1> @or_of_icmps1_vec(<2 x i32> %b) {
 ; CHECK-LABEL: @or_of_icmps1_vec(
-; CHECK-NEXT:    [[TMP1:%.*]] = add nsw <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp sge <2 x i32> [[TMP1]], <i32 4, i32 4>
-; CHECK-NEXT:    [[CMP3:%.*]] = icmp sle <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[CMP:%.*]] = or <2 x i1> [[TMP2]], [[CMP:%.*]]3
-; CHECK-NEXT:    ret <2 x i1> [[CMP]]
+; CHECK-NEXT:    ret <2 x i1> <i1 true, i1 true>
 ;
   %1 = add nsw <2 x i32> %b, <i32 2, i32 2>
   %2 = icmp sge <2 x i32> %1, <i32 4, i32 4>
@@ -237,11 +227,7 @@ define i1 @or_of_icmps2(i32 %b) {
 
 define <2 x i1> @or_of_icmps2_vec(<2 x i32> %b) {
 ; CHECK-LABEL: @or_of_icmps2_vec(
-; CHECK-NEXT:    [[TMP1:%.*]] = add <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp ugt <2 x i32> [[TMP1]], <i32 3, i32 3>
-; CHECK-NEXT:    [[CMP3:%.*]] = icmp sle <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[CMP:%.*]] = or <2 x i1> [[TMP2]], [[CMP:%.*]]3
-; CHECK-NEXT:    ret <2 x i1> [[CMP]]
+; CHECK-NEXT:    ret <2 x i1> <i1 true, i1 true>
 ;
   %1 = add <2 x i32> %b, <i32 2, i32 2>
   %2 = icmp ugt <2 x i32> %1, <i32 3, i32 3>
@@ -263,11 +249,7 @@ define i1 @or_of_icmps3(i32 %b) {
 
 define <2 x i1> @or_of_icmps3_vec(<2 x i32> %b) {
 ; CHECK-LABEL: @or_of_icmps3_vec(
-; CHECK-NEXT:    [[TMP1:%.*]] = add nsw <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp sgt <2 x i32> [[TMP1]], <i32 3, i32 3>
-; CHECK-NEXT:    [[CMP3:%.*]] = icmp sle <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[CMP:%.*]] = or <2 x i1> [[TMP2]], [[CMP:%.*]]3
-; CHECK-NEXT:    ret <2 x i1> [[CMP]]
+; CHECK-NEXT:    ret <2 x i1> <i1 true, i1 true>
 ;
   %1 = add nsw <2 x i32> %b, <i32 2, i32 2>
   %2 = icmp sgt <2 x i32> %1, <i32 3, i32 3>
@@ -289,11 +271,7 @@ define i1 @or_of_icmps4(i32 %b) {
 
 define <2 x i1> @or_of_icmps4_vec(<2 x i32> %b) {
 ; CHECK-LABEL: @or_of_icmps4_vec(
-; CHECK-NEXT:    [[TMP1:%.*]] = add nuw <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp uge <2 x i32> [[TMP1]], <i32 4, i32 4>
-; CHECK-NEXT:    [[CMP3:%.*]] = icmp ule <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[CMP:%.*]] = or <2 x i1> [[TMP2]], [[CMP:%.*]]3
-; CHECK-NEXT:    ret <2 x i1> [[CMP]]
+; CHECK-NEXT:    ret <2 x i1> <i1 true, i1 true>
 ;
   %1 = add nuw <2 x i32> %b, <i32 2, i32 2>
   %2 = icmp uge <2 x i32> %1, <i32 4, i32 4>
@@ -315,11 +293,7 @@ define i1 @or_of_icmps5(i32 %b) {
 
 define <2 x i1> @or_of_icmps5_vec(<2 x i32> %b) {
 ; CHECK-LABEL: @or_of_icmps5_vec(
-; CHECK-NEXT:    [[TMP1:%.*]] = add nuw <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[TMP2:%.*]] = icmp ugt <2 x i32> [[TMP1]], <i32 3, i32 3>
-; CHECK-NEXT:    [[CMP3:%.*]] = icmp ule <2 x i32> %b, <i32 2, i32 2>
-; CHECK-NEXT:    [[CMP:%.*]] = or <2 x i1> [[TMP2]], [[CMP:%.*]]3
-; CHECK-NEXT:    ret <2 x i1> [[CMP]]
+; CHECK-NEXT:    ret <2 x i1> <i1 true, i1 true>
 ;
   %1 = add nuw <2 x i32> %b, <i32 2, i32 2>
   %2 = icmp ugt <2 x i32> %1, <i32 3, i32 3>

More information about the llvm-commits mailing list