From 4635abb06529379bf2641b26d07d29aed5ce59f4 Mon Sep 17 00:00:00 2001 From: Robert Lougher Date: Fri, 13 Mar 2015 18:33:27 +0000 Subject: [PATCH] [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 | 34 ++- .../Reassociate/fast-ReassociateVector.ll | 224 +++++++++++++++--- 2 files changed, 200 insertions(+), 58 deletions(-) diff --git a/lib/Transforms/Scalar/Reassociate.cpp b/lib/Transforms/Scalar/Reassociate.cpp index 98016b40c56..307cc73d991 100644 --- a/lib/Transforms/Scalar/Reassociate.cpp +++ b/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 ((!Ty->isIntegerTy() && !Ty->isFloatingPointTy()) || - (!BinaryOperator::isNot(I) && !BinaryOperator::isNeg(I) && - !BinaryOperator::isFNeg(I))) + if (!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) - : ConstantFP::get(Ty, -1.0); + Constant *NegOne = Ty->isIntOrIntVectorTy() ? + 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) - : ConstantFP::get(Ty, NumFound); + Constant *C = Ty->isIntOrIntVectorTy() ? + 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. - if (I->getType()->isVectorTy()) + // TODO: We should optimize vector Xor instructions, but they are + // 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 Tree; diff --git a/test/Transforms/Reassociate/fast-ReassociateVector.ll b/test/Transforms/Reassociate/fast-ReassociateVector.ll index eeae096bf94..9fbb5ccfe9a 100644 --- a/test/Transforms/Reassociate/fast-ReassociateVector.ll +++ b/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. -define <4 x float> @test1() { -; CHECK-LABEL: test1 -; CHECK-NEXT: %tmp1 = fsub fast <4 x float> zeroinitializer, zeroinitializer -; CHECK-NEXT: %tmp2 = fmul fast <4 x float> %tmp1, zeroinitializer +; Check that a*c+b*c is turned into (a+b)*c +define <4 x float> @test1(<4 x float> %a, <4 x float> %b, <4 x float> %c) { +; CHECK-LABEL: @test1 +; CHECK-NEXT: %tmp = fadd fast <4 x float> %b, %a +; 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 - %tmp2 = fmul fast <4 x float> zeroinitializer, %tmp1 - ret <4 x float> %tmp2 + %mul = fmul fast <4 x float> %a, %c + %mul1 = fmul fast <4 x float> %b, %c + %add = fadd fast <4 x float> %mul, %mul1 + ret <4 x float> %add } -; Commute integer vector operations. -define <2 x i32> @test2(<2 x i32> %x, <2 x i32> %y) { -; CHECK-LABEL: test2 -; CHECK-NEXT: %tmp1 = add <2 x i32> %x, %y -; CHECK-NEXT: %tmp2 = add <2 x i32> %x, %y -; CHECK-NEXT: %tmp3 = add <2 x i32> %tmp1, %tmp2 +; Check that a*a*b+a*a*c is turned into a*(a*(b+c)). +define <2 x float> @test2(<2 x float> %a, <2 x float> %b, <2 x float> %c) { +; CHECK-LABEL: @test2 +; CHECK-NEXT: fadd fast <2 x float> %c, %b +; CHECK-NEXT: fmul fast <2 x float> %a, %tmp2 +; CHECK-NEXT: fmul fast <2 x float> %tmp3, %a +; CHECK-NEXT: ret <2 x float> - %tmp1 = add <2 x i32> %x, %y - %tmp2 = add <2 x i32> %y, %x - %tmp3 = add <2 x i32> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %t0 = fmul fast <2 x float> %a, %b + %t1 = fmul fast <2 x float> %a, %t0 + %t2 = fmul fast <2 x float> %a, %c + %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-LABEL: test3 -; CHECK-NEXT: %tmp1 = mul <2 x i32> %x, %y -; CHECK-NEXT: %tmp2 = mul <2 x i32> %x, %y -; CHECK-NEXT: %tmp3 = mul <2 x i32> %tmp1, %tmp2 +; Check that a*b+a*c+d is turned into a*(b+c)+d. +define <2 x double> @test3(<2 x double> %a, <2 x double> %b, <2 x double> %c, <2 x double> %d) { +; CHECK-LABEL: @test3 +; CHECK-NEXT: fadd fast <2 x double> %c, %b +; 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 - %tmp2 = mul <2 x i32> %y, %x - %tmp3 = mul <2 x i32> %tmp1, %tmp2 - ret <2 x i32> %tmp3 + %t0 = fmul fast <2 x double> %a, %b + %t1 = fmul fast <2 x double> %a, %c + %t2 = fadd fast <2 x double> %t1, %d + %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) { -; CHECK-LABEL: test4 -; CHECK-NEXT: %tmp1 = and <2 x i32> %x, %y -; CHECK-NEXT: %tmp2 = and <2 x i32> %x, %y -; CHECK-NEXT: %tmp3 = and <2 x i32> %tmp1, %tmp2 +; No fast-math. +define <2 x float> @test4(<2 x float> %A) { +; CHECK-LABEL: @test4 +; CHECK-NEXT: %X = fadd <2 x float> %A, +; CHECK-NEXT: %Y = fadd <2 x float> %A, +; 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, +; CHECK-NEXT: ret <2 x float> + + %Y = fmul fast <2 x float> %X, + %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, +; 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, +; CHECK-NEXT: ret <2 x double> + + %X = fmul fast <2 x double> %W, + %Y = fmul fast <2 x double> %W, + %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, +; CHECK-NEXT: ret <2 x float> %tmp2 + + %tmp1 = fmul fast <2 x float> , %arg + %tmp2 = fmul fast <2 x float> %tmp1, + 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, +; CHECK-NEXT: ret <2 x double> + + %1 = fadd fast <2 x double> %a, + %2 = fadd fast <2 x double> %b, %1 + %3 = fsub fast <2 x double> , %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, +; CHECK-NEXT: fmul fast <2 x float> %e, %z +; CHECK-NEXT: ret <2 x float> + + %d = fmul fast <2 x float> %z, + %c = fsub fast <2 x float> , %d + %e = fmul fast <2 x float> %a, %c + %f = fsub fast <2 x float> , %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, +; 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, +; CHECK-NEXT: ret <2 x i64> %shl + + %mul = mul <2 x i64> %c, %b + %shl = shl <2 x i64> %mul, + 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, +; CHECK-NEXT: %add = fadd fast <4 x float> %mul, %a +; CHECK-NEXT: ret <4 x float> %add + + %mul = fmul fast <4 x float> , %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) { -; CHECK-LABEL: test5 -; CHECK-NEXT: %tmp1 = or <2 x i32> %x, %y -; CHECK-NEXT: %tmp2 = or <2 x i32> %x, %y -; CHECK-NEXT: %tmp3 = or <2 x i32> %tmp1, %tmp2 +define <2 x i32> @test16(<2 x i32> %x, <2 x i32> %y) { +; CHECK-LABEL: test16 +; CHECK-NEXT: %tmp3 = or <2 x i32> %y, %x +; CHECK-NEXT: ret <2 x i32> %tmp3 %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) { -; CHECK-LABEL: test6 +; FIXME: Optimize vector xor. Currently only commute operands. +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