1 //===- ExpandOps.cpp - Pass to legalize Arithmetic ops for LLVM lowering --===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 #include "PassDetail.h"
10 #include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
11 #include "mlir/Dialect/Arithmetic/Transforms/Passes.h"
12 #include "mlir/Dialect/StandardOps/IR/Ops.h"
13 #include "mlir/IR/TypeUtilities.h"
14 #include "mlir/Transforms/DialectConversion.h"
15 
16 using namespace mlir;
17 
18 /// Create an integer or index constant.
19 static Value createConst(Location loc, Type type, int value,
20                          PatternRewriter &rewriter) {
21   return rewriter.create<arith::ConstantOp>(
22       loc, rewriter.getIntegerAttr(type, value));
23 }
24 
25 namespace {
26 
27 /// Expands CeilDivUIOp (n, m) into
28 ///  n == 0 ? 0 : ((n-1) / m) + 1
29 struct CeilDivUIOpConverter : public OpRewritePattern<arith::CeilDivUIOp> {
30   using OpRewritePattern::OpRewritePattern;
31   LogicalResult matchAndRewrite(arith::CeilDivUIOp op,
32                                 PatternRewriter &rewriter) const final {
33     Location loc = op.getLoc();
34     Value a = op.getLhs();
35     Value b = op.getRhs();
36     Value zero = createConst(loc, a.getType(), 0, rewriter);
37     Value compare =
38         rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::eq, a, zero);
39     Value one = createConst(loc, a.getType(), 1, rewriter);
40     Value minusOne = rewriter.create<arith::SubIOp>(loc, a, one);
41     Value quotient = rewriter.create<arith::DivUIOp>(loc, minusOne, b);
42     Value plusOne = rewriter.create<arith::AddIOp>(loc, quotient, one);
43     rewriter.replaceOpWithNewOp<SelectOp>(op, compare, zero, plusOne);
44     return success();
45   }
46 };
47 
48 /// Expands CeilDivSIOp (n, m) into
49 ///   1) x = (m > 0) ? -1 : 1
50 ///   2) (n*m>0) ? ((n+x) / m) + 1 : - (-n / m)
51 struct CeilDivSIOpConverter : public OpRewritePattern<arith::CeilDivSIOp> {
52   using OpRewritePattern::OpRewritePattern;
53   LogicalResult matchAndRewrite(arith::CeilDivSIOp op,
54                                 PatternRewriter &rewriter) const final {
55     Location loc = op.getLoc();
56     Type type = op.getType();
57     Value a = op.getLhs();
58     Value b = op.getRhs();
59     Value plusOne = createConst(loc, type, 1, rewriter);
60     Value zero = createConst(loc, type, 0, rewriter);
61     Value minusOne = createConst(loc, type, -1, rewriter);
62     // Compute x = (b>0) ? -1 : 1.
63     Value compare =
64         rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::sgt, b, zero);
65     Value x = rewriter.create<SelectOp>(loc, compare, minusOne, plusOne);
66     // Compute positive res: 1 + ((x+a)/b).
67     Value xPlusA = rewriter.create<arith::AddIOp>(loc, x, a);
68     Value xPlusADivB = rewriter.create<arith::DivSIOp>(loc, xPlusA, b);
69     Value posRes = rewriter.create<arith::AddIOp>(loc, plusOne, xPlusADivB);
70     // Compute negative res: - ((-a)/b).
71     Value minusA = rewriter.create<arith::SubIOp>(loc, zero, a);
72     Value minusADivB = rewriter.create<arith::DivSIOp>(loc, minusA, b);
73     Value negRes = rewriter.create<arith::SubIOp>(loc, zero, minusADivB);
74     // Result is (a*b>0) ? pos result : neg result.
75     // Note, we want to avoid using a*b because of possible overflow.
76     // The case that matters are a>0, a==0, a<0, b>0 and b<0. We do
77     // not particuliarly care if a*b<0 is true or false when b is zero
78     // as this will result in an illegal divide. So `a*b<0` can be reformulated
79     // as `(a<0 && b<0) || (a>0 && b>0)' or `(a<0 && b<0) || (a>0 && b>=0)'.
80     // We pick the first expression here.
81     Value aNeg =
82         rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::slt, a, zero);
83     Value aPos =
84         rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::sgt, a, zero);
85     Value bNeg =
86         rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::slt, b, zero);
87     Value bPos =
88         rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::sgt, b, zero);
89     Value firstTerm = rewriter.create<arith::AndIOp>(loc, aNeg, bNeg);
90     Value secondTerm = rewriter.create<arith::AndIOp>(loc, aPos, bPos);
91     Value compareRes =
92         rewriter.create<arith::OrIOp>(loc, firstTerm, secondTerm);
93     // Perform substitution and return success.
94     rewriter.replaceOpWithNewOp<SelectOp>(op, compareRes, posRes, negRes);
95     return success();
96   }
97 };
98 
99 /// Expands FloorDivSIOp (n, m) into
100 ///   1)  x = (m<0) ? 1 : -1
101 ///   2)  return (n*m<0) ? - ((-n+x) / m) -1 : n / m
102 struct FloorDivSIOpConverter : public OpRewritePattern<arith::FloorDivSIOp> {
103   using OpRewritePattern::OpRewritePattern;
104   LogicalResult matchAndRewrite(arith::FloorDivSIOp op,
105                                 PatternRewriter &rewriter) const final {
106     Location loc = op.getLoc();
107     Type type = op.getType();
108     Value a = op.getLhs();
109     Value b = op.getRhs();
110     Value plusOne = createConst(loc, type, 1, rewriter);
111     Value zero = createConst(loc, type, 0, rewriter);
112     Value minusOne = createConst(loc, type, -1, rewriter);
113     // Compute x = (b<0) ? 1 : -1.
114     Value compare =
115         rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::slt, b, zero);
116     Value x = rewriter.create<SelectOp>(loc, compare, plusOne, minusOne);
117     // Compute negative res: -1 - ((x-a)/b).
118     Value xMinusA = rewriter.create<arith::SubIOp>(loc, x, a);
119     Value xMinusADivB = rewriter.create<arith::DivSIOp>(loc, xMinusA, b);
120     Value negRes = rewriter.create<arith::SubIOp>(loc, minusOne, xMinusADivB);
121     // Compute positive res: a/b.
122     Value posRes = rewriter.create<arith::DivSIOp>(loc, a, b);
123     // Result is (a*b<0) ? negative result : positive result.
124     // Note, we want to avoid using a*b because of possible overflow.
125     // The case that matters are a>0, a==0, a<0, b>0 and b<0. We do
126     // not particuliarly care if a*b<0 is true or false when b is zero
127     // as this will result in an illegal divide. So `a*b<0` can be reformulated
128     // as `(a>0 && b<0) || (a>0 && b<0)' or `(a>0 && b<0) || (a>0 && b<=0)'.
129     // We pick the first expression here.
130     Value aNeg =
131         rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::slt, a, zero);
132     Value aPos =
133         rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::sgt, a, zero);
134     Value bNeg =
135         rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::slt, b, zero);
136     Value bPos =
137         rewriter.create<arith::CmpIOp>(loc, arith::CmpIPredicate::sgt, b, zero);
138     Value firstTerm = rewriter.create<arith::AndIOp>(loc, aNeg, bPos);
139     Value secondTerm = rewriter.create<arith::AndIOp>(loc, aPos, bNeg);
140     Value compareRes =
141         rewriter.create<arith::OrIOp>(loc, firstTerm, secondTerm);
142     // Perform substitution and return success.
143     rewriter.replaceOpWithNewOp<SelectOp>(op, compareRes, negRes, posRes);
144     return success();
145   }
146 };
147 
148 template <typename OpTy, arith::CmpFPredicate pred>
149 struct MaxMinFOpConverter : public OpRewritePattern<OpTy> {
150 public:
151   using OpRewritePattern<OpTy>::OpRewritePattern;
152 
153   LogicalResult matchAndRewrite(OpTy op,
154                                 PatternRewriter &rewriter) const final {
155     Value lhs = op.getLhs();
156     Value rhs = op.getRhs();
157 
158     Location loc = op.getLoc();
159     Value cmp = rewriter.create<arith::CmpFOp>(loc, pred, lhs, rhs);
160     Value select = rewriter.create<SelectOp>(loc, cmp, lhs, rhs);
161 
162     auto floatType = getElementTypeOrSelf(lhs.getType()).cast<FloatType>();
163     Value isNaN = rewriter.create<arith::CmpFOp>(loc, arith::CmpFPredicate::UNO,
164                                                  lhs, rhs);
165 
166     Value nan = rewriter.create<arith::ConstantFloatOp>(
167         loc, APFloat::getQNaN(floatType.getFloatSemantics()), floatType);
168     if (VectorType vectorType = lhs.getType().dyn_cast<VectorType>())
169       nan = rewriter.create<SplatOp>(loc, vectorType, nan);
170 
171     rewriter.replaceOpWithNewOp<SelectOp>(op, isNaN, nan, select);
172     return success();
173   }
174 };
175 
176 template <typename OpTy, arith::CmpIPredicate pred>
177 struct MaxMinIOpConverter : public OpRewritePattern<OpTy> {
178 public:
179   using OpRewritePattern<OpTy>::OpRewritePattern;
180   LogicalResult matchAndRewrite(OpTy op,
181                                 PatternRewriter &rewriter) const final {
182     Value lhs = op.getLhs();
183     Value rhs = op.getRhs();
184 
185     Location loc = op.getLoc();
186     Value cmp = rewriter.create<arith::CmpIOp>(loc, pred, lhs, rhs);
187     rewriter.replaceOpWithNewOp<SelectOp>(op, cmp, lhs, rhs);
188     return success();
189   }
190 };
191 
192 struct ArithmeticExpandOpsPass
193     : public ArithmeticExpandOpsBase<ArithmeticExpandOpsPass> {
194   void runOnFunction() override {
195     RewritePatternSet patterns(&getContext());
196     ConversionTarget target(getContext());
197 
198     arith::populateArithmeticExpandOpsPatterns(patterns);
199 
200     target.addLegalDialect<arith::ArithmeticDialect, StandardOpsDialect>();
201     // clang-format off
202     target.addIllegalOp<
203       arith::CeilDivSIOp,
204       arith::CeilDivUIOp,
205       arith::FloorDivSIOp,
206       arith::MaxFOp,
207       arith::MaxSIOp,
208       arith::MaxUIOp,
209       arith::MinFOp,
210       arith::MinSIOp,
211       arith::MinUIOp
212     >();
213     // clang-format on
214     if (failed(
215             applyPartialConversion(getFunction(), target, std::move(patterns))))
216       signalPassFailure();
217   }
218 };
219 
220 } // namespace
221 
222 void mlir::arith::populateArithmeticExpandOpsPatterns(
223     RewritePatternSet &patterns) {
224   // clang-format off
225   patterns.add<
226     CeilDivSIOpConverter,
227     CeilDivUIOpConverter,
228     FloorDivSIOpConverter,
229     MaxMinFOpConverter<MaxFOp, arith::CmpFPredicate::OGT>,
230     MaxMinFOpConverter<MinFOp, arith::CmpFPredicate::OLT>,
231     MaxMinIOpConverter<MaxSIOp, arith::CmpIPredicate::sgt>,
232     MaxMinIOpConverter<MaxUIOp, arith::CmpIPredicate::ugt>,
233     MaxMinIOpConverter<MinSIOp, arith::CmpIPredicate::slt>,
234     MaxMinIOpConverter<MinUIOp, arith::CmpIPredicate::ult>
235    >(patterns.getContext());
236   // clang-format on
237 }
238 
239 std::unique_ptr<Pass> mlir::arith::createArithmeticExpandOpsPass() {
240   return std::make_unique<ArithmeticExpandOpsPass>();
241 }
242