[Reassociate] Add initial support for vector instructions.

This patch adds initial support for vector instructions to the reassociation
pass. It enables most parts of the pass to work with vectors but to keep the
size of the patch small, optimization of Xor trees, canonicalization of
negative constants and converting shifts to muls, etc., have been left out.
This will be handled in later patches.

The patch is based on an initial patch by Chad Rosier.

Differential Revision: http://reviews.llvm.org/D7566


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@232190 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Transforms/Scalar/Reassociate.cpp

@@ -321,10 +321,8 @@ unsigned Reassociate::getRank(Value *V) {
 
   // If this is a not or neg instruction, do not count it for rank.  This
   // assures us that X and ~X will have the same rank.
-  Type *Ty = V->getType();
+  if  (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
-  if ((!Ty->isIntegerTy() && !Ty->isFloatingPointTy()) ||
+       !BinaryOperator::isFNeg(I))
-      (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) &&
-       !BinaryOperator::isFNeg(I)))
     ++Rank;
 
   DEBUG(dbgs() << "Calculated Rank[" << V->getName() << "] = " << Rank << "\n");
@@ -351,7 +349,7 @@ void Reassociate::canonicalizeOperands(Instruction *I) {
 
 static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
                                  Instruction *InsertBefore, Value *FlagsOp) {
-  if (S1->getType()->isIntegerTy())
+  if (S1->getType()->isIntOrIntVectorTy())
     return BinaryOperator::CreateAdd(S1, S2, Name, InsertBefore);
   else {
     BinaryOperator *Res =
@@ -363,7 +361,7 @@ static BinaryOperator *CreateAdd(Value *S1, Value *S2, const Twine &Name,
 
 static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
                                  Instruction *InsertBefore, Value *FlagsOp) {
-  if (S1->getType()->isIntegerTy())
+  if (S1->getType()->isIntOrIntVectorTy())
     return BinaryOperator::CreateMul(S1, S2, Name, InsertBefore);
   else {
     BinaryOperator *Res =
@@ -375,7 +373,7 @@ static BinaryOperator *CreateMul(Value *S1, Value *S2, const Twine &Name,
 
 static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
                                  Instruction *InsertBefore, Value *FlagsOp) {
-  if (S1->getType()->isIntegerTy())
+  if (S1->getType()->isIntOrIntVectorTy())
     return BinaryOperator::CreateNeg(S1, Name, InsertBefore);
   else {
     BinaryOperator *Res = BinaryOperator::CreateFNeg(S1, Name, InsertBefore);
@@ -388,8 +386,8 @@ static BinaryOperator *CreateNeg(Value *S1, const Twine &Name,
 ///
 static BinaryOperator *LowerNegateToMultiply(Instruction *Neg) {
   Type *Ty = Neg->getType();
-  Constant *NegOne = Ty->isIntegerTy() ? ConstantInt::getAllOnesValue(Ty)
+  Constant *NegOne = Ty->isIntOrIntVectorTy() ?
-                                       : ConstantFP::get(Ty, -1.0);
+    ConstantInt::getAllOnesValue(Ty) : ConstantFP::get(Ty, -1.0);
 
   BinaryOperator *Res = CreateMul(Neg->getOperand(1), NegOne, "", Neg, Neg);
   Neg->setOperand(1, Constant::getNullValue(Ty)); // Drop use of op.
@@ -872,7 +870,7 @@ void Reassociate::RewriteExprTree(BinaryOperator *I,
       Constant *Undef = UndefValue::get(I->getType());
       NewOp = BinaryOperator::Create(Instruction::BinaryOps(Opcode),
                                      Undef, Undef, "", I);
-      if (NewOp->getType()->isFloatingPointTy())
+      if (NewOp->getType()->isFPOrFPVectorTy())
         NewOp->setFastMathFlags(I->getFastMathFlags());
     } else {
       NewOp = NodesToRewrite.pop_back_val();
@@ -1520,8 +1518,8 @@ Value *Reassociate::OptimizeAdd(Instruction *I,
 
       // Insert a new multiply.
       Type *Ty = TheOp->getType();
-      Constant *C = Ty->isIntegerTy() ? ConstantInt::get(Ty, NumFound)
+      Constant *C = Ty->isIntOrIntVectorTy() ?
-                                      : ConstantFP::get(Ty, NumFound);
+        ConstantInt::get(Ty, NumFound) : ConstantFP::get(Ty, NumFound);
       Instruction *Mul = CreateMul(TheOp, C, "factor", I, I);
 
       // Now that we have inserted a multiply, optimize it. This allows us to
@@ -1661,7 +1659,7 @@ Value *Reassociate::OptimizeAdd(Instruction *I,
     // from an expression will drop a use of maxocc, and this can cause
     // RemoveFactorFromExpression on successive values to behave differently.
     Instruction *DummyInst =
-        I->getType()->isIntegerTy()
+        I->getType()->isIntOrIntVectorTy()
             ? BinaryOperator::CreateAdd(MaxOccVal, MaxOccVal)
             : BinaryOperator::CreateFAdd(MaxOccVal, MaxOccVal);
 
@@ -1792,7 +1790,7 @@ static Value *buildMultiplyTree(IRBuilder<> &Builder,
 
   Value *LHS = Ops.pop_back_val();
   do {
-    if (LHS->getType()->isIntegerTy())
+    if (LHS->getType()->isIntOrIntVectorTy())
       LHS = Builder.CreateMul(LHS, Ops.pop_back_val());
     else
       LHS = Builder.CreateFMul(LHS, Ops.pop_back_val());
@@ -2090,8 +2088,9 @@ void Reassociate::OptimizeInst(Instruction *I) {
   if (I->isCommutative())
     canonicalizeOperands(I);
 
-  // Don't optimize vector instructions.
+  // TODO: We should optimize vector Xor instructions, but they are
-  if (I->getType()->isVectorTy())
+  // currently unsupported.
+  if (I->getType()->isVectorTy() && I->getOpcode() == Instruction::Xor)
     return;
 
   // Don't optimize floating point instructions that don't have unsafe algebra.
@@ -2170,9 +2169,6 @@ void Reassociate::OptimizeInst(Instruction *I) {
 }
 
 void Reassociate::ReassociateExpression(BinaryOperator *I) {
-  assert(!I->getType()->isVectorTy() &&
-         "Reassociation of vector instructions is not supported.");
-
   // First, walk the expression tree, linearizing the tree, collecting the
   // operand information.
   SmallVector<RepeatedValue, 8> Tree;

test/Transforms/Reassociate/fast-ReassociateVector.ll

@@ -1,46 +1,192 @@
 ; RUN: opt < %s -reassociate -S | FileCheck %s
 
-; Canonicalize operands, but don't optimize floating point vector operations.
+; Check that a*c+b*c is turned into (a+b)*c
-define <4 x float> @test1() {
+define <4 x float> @test1(<4 x float> %a, <4 x float> %b, <4 x float> %c) {
-; CHECK-LABEL: test1
+; CHECK-LABEL: @test1
-; CHECK-NEXT: %tmp1 = fsub fast <4 x float> zeroinitializer, zeroinitializer
+; CHECK-NEXT: %tmp = fadd fast <4 x float> %b, %a
-; CHECK-NEXT: %tmp2 = fmul fast <4 x float> %tmp1, zeroinitializer
+; CHECK-NEXT: %tmp1 = fmul fast <4 x float> %tmp, %c
+; CHECK-NEXT: ret <4 x float> %tmp1
 
-  %tmp1 = fsub fast <4 x float> zeroinitializer, zeroinitializer
+  %mul = fmul fast <4 x float> %a, %c
-  %tmp2 = fmul fast <4 x float> zeroinitializer, %tmp1
+  %mul1 = fmul fast <4 x float> %b, %c
-  ret <4 x float> %tmp2
+  %add = fadd fast <4 x float> %mul, %mul1
+  ret <4 x float> %add
 }
 
-; Commute integer vector operations.
+; Check that a*a*b+a*a*c is turned into a*(a*(b+c)).
-define <2 x i32> @test2(<2 x i32> %x, <2 x i32> %y) {
+define <2 x float> @test2(<2 x float> %a, <2 x float> %b, <2 x float> %c) {
-; CHECK-LABEL: test2
+; CHECK-LABEL: @test2
-; CHECK-NEXT: %tmp1 = add <2 x i32> %x, %y
+; CHECK-NEXT: fadd fast <2 x float> %c, %b
-; CHECK-NEXT: %tmp2 = add <2 x i32> %x, %y
+; CHECK-NEXT: fmul fast <2 x float> %a, %tmp2
-; CHECK-NEXT: %tmp3 = add <2 x i32> %tmp1, %tmp2
+; CHECK-NEXT: fmul fast <2 x float> %tmp3, %a
+; CHECK-NEXT: ret <2 x float>
 
-  %tmp1 = add <2 x i32> %x, %y
+  %t0 = fmul fast <2 x float> %a, %b
-  %tmp2 = add <2 x i32> %y, %x
+  %t1 = fmul fast <2 x float> %a, %t0
-  %tmp3 = add <2 x i32> %tmp1, %tmp2
+  %t2 = fmul fast <2 x float> %a, %c
-  ret <2 x i32> %tmp3
+  %t3 = fmul fast <2 x float> %a, %t2
+  %t4 = fadd fast <2 x float> %t1, %t3
+  ret <2 x float> %t4
 }
 
-define <2 x i32> @test3(<2 x i32> %x, <2 x i32> %y) {
+; Check that a*b+a*c+d is turned into a*(b+c)+d.
-; CHECK-LABEL: test3
+define <2 x double> @test3(<2 x double> %a, <2 x double> %b, <2 x double> %c, <2 x double> %d) {
-; CHECK-NEXT: %tmp1 = mul <2 x i32> %x, %y
+; CHECK-LABEL: @test3
-; CHECK-NEXT: %tmp2 = mul <2 x i32> %x, %y
+; CHECK-NEXT: fadd fast <2 x double> %c, %b
-; CHECK-NEXT: %tmp3 = mul <2 x i32> %tmp1, %tmp2
+; CHECK-NEXT: fmul fast <2 x double> %tmp, %a
+; CHECK-NEXT: fadd fast <2 x double> %tmp1, %d
+; CHECK-NEXT: ret <2 x double>
 
-  %tmp1 = mul <2 x i32> %x, %y
+  %t0 = fmul fast <2 x double> %a, %b
-  %tmp2 = mul <2 x i32> %y, %x
+  %t1 = fmul fast <2 x double> %a, %c
-  %tmp3 = mul <2 x i32> %tmp1, %tmp2
+  %t2 = fadd fast <2 x double> %t1, %d
-  ret <2 x i32> %tmp3
+  %t3 = fadd fast <2 x double> %t0, %t2
+  ret <2 x double> %t3
 }
 
-define <2 x i32> @test4(<2 x i32> %x, <2 x i32> %y) {
+; No fast-math.
-; CHECK-LABEL: test4
+define <2 x float> @test4(<2 x float> %A) {
-; CHECK-NEXT: %tmp1 = and <2 x i32> %x, %y
+; CHECK-LABEL: @test4
-; CHECK-NEXT: %tmp2 = and <2 x i32> %x, %y
+; CHECK-NEXT: %X = fadd <2 x float> %A, <float 1.000000e+00, float 1.000000e+00>
-; CHECK-NEXT: %tmp3 = and <2 x i32> %tmp1, %tmp2
+; CHECK-NEXT: %Y = fadd <2 x float> %A, <float 1.000000e+00, float 1.000000e+00>
+; CHECK-NEXT: %R = fsub <2 x float> %X, %Y
+; CHECK-NEXT: ret <2 x float> %R
+
+  %X = fadd <2 x float> %A, < float 1.000000e+00, float 1.000000e+00 >
+  %Y = fadd <2 x float> %A, < float 1.000000e+00, float 1.000000e+00 >
+  %R = fsub <2 x float> %X, %Y
+  ret <2 x float> %R
+}
+
+; Check 47*X + 47*X -> 94*X.
+define <2 x float> @test5(<2 x float> %X) {
+; CHECK-LABEL: @test5
+; CHECK-NEXT: fmul fast <2 x float> %X, <float 9.400000e+01, float 9.400000e+01>
+; CHECK-NEXT: ret <2 x float>
+
+  %Y = fmul fast <2 x float> %X, <float 4.700000e+01, float 4.700000e+01>
+  %Z = fadd fast <2 x float> %Y, %Y
+  ret <2 x float> %Z
+}
+
+; Check X+X+X -> 3*X.
+define <2 x float> @test6(<2 x float> %X) {
+; CHECK-LABEL: @test6
+; CHECK-NEXT: fmul fast <2 x float> %X, <float 3.000000e+00, float 3.000000e+00>
+; CHECK-NEXT: ret <2 x float>
+
+  %Y = fadd fast <2 x float> %X ,%X
+  %Z = fadd fast <2 x float> %Y, %X
+  ret <2 x float> %Z
+}
+
+; Check 127*W+50*W -> 177*W.
+define <2 x double> @test7(<2 x double> %W) {
+; CHECK-LABEL: @test7
+; CHECK-NEXT: fmul fast <2 x double> %W, <double 1.770000e+02, double 1.770000e+02>
+; CHECK-NEXT: ret <2 x double>
+
+  %X = fmul fast <2 x double> %W, <double 127.0, double 127.0>
+  %Y = fmul fast <2 x double> %W, <double 50.0, double 50.0>
+  %Z = fadd fast <2 x double> %Y, %X
+  ret <2 x double> %Z
+}
+
+; Check X*12*12 -> X*144.
+define <2 x float> @test8(<2 x float> %arg) {
+; CHECK-LABEL: @test8
+; CHECK: fmul fast <2 x float> %arg, <float 1.440000e+02, float 1.440000e+02>
+; CHECK-NEXT: ret <2 x float> %tmp2
+
+  %tmp1 = fmul fast <2 x float> <float 1.200000e+01, float 1.200000e+01>, %arg
+  %tmp2 = fmul fast <2 x float> %tmp1, <float 1.200000e+01, float 1.200000e+01>
+  ret <2 x float> %tmp2
+}
+
+; Check (b+(a+1234))+-a -> b+1234.
+define <2 x double> @test9(<2 x double> %b, <2 x double> %a) {
+; CHECK-LABEL: @test9
+; CHECK: fadd fast <2 x double> %b, <double 1.234000e+03, double 1.234000e+03>
+; CHECK-NEXT: ret <2 x double>
+
+  %1 = fadd fast <2 x double> %a, <double 1.234000e+03, double 1.234000e+03>
+  %2 = fadd fast <2 x double> %b, %1
+  %3 = fsub fast <2 x double> <double 0.000000e+00, double 0.000000e+00>, %a
+  %4 = fadd fast <2 x double> %2, %3
+  ret <2 x double> %4
+}
+
+; Check -(-(z*40)*a) -> a*40*z.
+define <2 x float> @test10(<2 x float> %a, <2 x float> %b, <2 x float> %z) {
+; CHECK-LABEL: @test10
+; CHECK: fmul fast <2 x float> %a, <float 4.000000e+01, float 4.000000e+01>
+; CHECK-NEXT: fmul fast <2 x float> %e, %z
+; CHECK-NEXT: ret <2 x float>
+
+  %d = fmul fast <2 x float> %z, <float 4.000000e+01, float 4.000000e+01>
+  %c = fsub fast <2 x float> <float 0.000000e+00, float 0.000000e+00>, %d
+  %e = fmul fast <2 x float> %a, %c
+  %f = fsub fast <2 x float> <float 0.000000e+00, float 0.000000e+00>, %e
+  ret <2 x float> %f
+}
+
+; Check x*y+y*x -> x*y*2.
+define <2 x double> @test11(<2 x double> %x, <2 x double> %y) {
+; CHECK-LABEL: @test11
+; CHECK-NEXT: %factor = fmul fast <2 x double> %y, <double 2.000000e+00, double 2.000000e+00>
+; CHECK-NEXT: %tmp1 = fmul fast <2 x double> %factor, %x
+; CHECK-NEXT: ret <2 x double> %tmp1
+
+  %1 = fmul fast <2 x double> %x, %y
+  %2 = fmul fast <2 x double> %y, %x
+  %3 = fadd fast <2 x double> %1, %2
+  ret <2 x double> %3
+}
+
+; FIXME: shifts should be converted to mul to assist further reassociation.
+define <2 x i64> @test12(<2 x i64> %b, <2 x i64> %c) {
+; CHECK-LABEL: @test12
+; CHECK-NEXT:  %mul = mul <2 x i64> %c, %b
+; CHECK-NEXT:  %shl = shl <2 x i64> %mul, <i64 5, i64 5>
+; CHECK-NEXT:  ret <2 x i64> %shl
+
+  %mul = mul <2 x i64> %c, %b
+  %shl = shl <2 x i64> %mul, <i64 5, i64 5>
+  ret <2 x i64> %shl
+}
+
+; FIXME: expressions with a negative const should be canonicalized to assist
+; further reassociation.
+; We would expect (-5*b)+a -> a-(5*b) but only the constant operand is commuted.
+define <4 x float> @test13(<4 x float> %a, <4 x float> %b) {
+; CHECK-LABEL: @test13
+; CHECK-NEXT:  %mul = fmul fast <4 x float> %b, <float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00>
+; CHECK-NEXT:  %add = fadd fast <4 x float> %mul, %a
+; CHECK-NEXT:  ret <4 x float> %add
+
+  %mul = fmul fast <4 x float> <float -5.000000e+00, float -5.000000e+00, float -5.000000e+00, float -5.000000e+00>, %b
+  %add = fadd fast <4 x float> %mul, %a
+  ret <4 x float> %add
+}
+
+; Break up subtract to assist further reassociation.
+; Check a+b-c -> a+b+-c.
+define <2 x i64> @test14(<2 x i64> %a, <2 x i64> %b, <2 x i64> %c) {
+; CHECK-LABEL: @test14
+; CHECK-NEXT: %add = add <2 x i64> %b, %a
+; CHECK-NEXT: %c.neg = sub <2 x i64> zeroinitializer, %c
+; CHECK-NEXT: %sub = add <2 x i64> %add, %c.neg
+; CHECK-NEXT: ret <2 x i64> %sub
+
+  %add = add <2 x i64> %b, %a
+  %sub = sub <2 x i64> %add, %c
+  ret <2 x i64> %sub
+}
+
+define <2 x i32> @test15(<2 x i32> %x, <2 x i32> %y) {
+; CHECK-LABEL: test15
+; CHECK-NEXT: %tmp3 = and <2 x i32> %y, %x
+; CHECK-NEXT: ret <2 x i32> %tmp3
 
   %tmp1 = and <2 x i32> %x, %y
   %tmp2 = and <2 x i32> %y, %x
@@ -48,11 +194,10 @@ define <2 x i32> @test4(<2 x i32> %x, <2 x i32> %y) {
   ret <2 x i32> %tmp3
 }
 
-define <2 x i32> @test5(<2 x i32> %x, <2 x i32> %y) {
+define <2 x i32> @test16(<2 x i32> %x, <2 x i32> %y) {
-; CHECK-LABEL: test5
+; CHECK-LABEL: test16
-; CHECK-NEXT: %tmp1 = or <2 x i32> %x, %y
+; CHECK-NEXT: %tmp3 = or <2 x i32> %y, %x
-; CHECK-NEXT: %tmp2 = or <2 x i32> %x, %y
+; CHECK-NEXT: ret <2 x i32> %tmp3
-; CHECK-NEXT: %tmp3 = or <2 x i32> %tmp1, %tmp2
 
   %tmp1 = or <2 x i32> %x, %y
   %tmp2 = or <2 x i32> %y, %x
@@ -60,8 +205,9 @@ define <2 x i32> @test5(<2 x i32> %x, <2 x i32> %y) {
   ret <2 x i32> %tmp3
 }
 
-define <2 x i32> @test6(<2 x i32> %x, <2 x i32> %y) {
+; FIXME: Optimize vector xor.  Currently only commute operands.
-; CHECK-LABEL: test6
+define <2 x i32> @test17(<2 x i32> %x, <2 x i32> %y) {
+; CHECK-LABEL: test17
 ; CHECK-NEXT: %tmp1 = xor <2 x i32> %x, %y
 ; CHECK-NEXT: %tmp2 = xor <2 x i32> %x, %y
 ; CHECK-NEXT: %tmp3 = xor <2 x i32> %tmp1, %tmp2