1 //===- SimplifyIntrinsics.cpp -- replace intrinsics with simpler form -----===//
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
7 //===----------------------------------------------------------------------===//
9 //===----------------------------------------------------------------------===//
11 /// This pass looks for suitable calls to runtime library for intrinsics that
12 /// can be simplified/specialized and replaces with a specialized function.
14 /// For example, SUM(arr) can be specialized as a simple function with one loop,
15 /// compared to the three arguments (plus file & line info) that the runtime
16 /// call has - when the argument is a 1D-array (multiple loops may be needed
17 // for higher dimension arrays, of course)
19 /// The general idea is that besides making the call simpler, it can also be
20 /// inlined by other passes that run after this pass, which further improves
21 /// performance, particularly when the work done in the function is trivial
22 /// and small in size.
23 //===----------------------------------------------------------------------===//
25 #include "flang/Optimizer/Builder/BoxValue.h"
26 #include "flang/Optimizer/Builder/FIRBuilder.h"
27 #include "flang/Optimizer/Builder/Todo.h"
28 #include "flang/Optimizer/Dialect/FIROps.h"
29 #include "flang/Optimizer/Dialect/FIRType.h"
30 #include "flang/Optimizer/Support/FIRContext.h"
31 #include "flang/Optimizer/Transforms/Passes.h"
32 #include "flang/Runtime/entry-names.h"
33 #include "mlir/Dialect/LLVMIR/LLVMDialect.h"
34 #include "mlir/IR/Matchers.h"
35 #include "mlir/IR/TypeUtilities.h"
36 #include "mlir/Pass/Pass.h"
37 #include "mlir/Transforms/DialectConversion.h"
38 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
39 #include "mlir/Transforms/RegionUtils.h"
40 #include "llvm/ADT/Optional.h"
41 #include "llvm/Support/Debug.h"
42 #include "llvm/Support/raw_ostream.h"
45 #define GEN_PASS_DEF_SIMPLIFYINTRINSICS
46 #include "flang/Optimizer/Transforms/Passes.h.inc"
49 #define RTNAME_STRINGIFY2(x) #x
50 #define RTNAME_STRINGIFY(x) RTNAME_STRINGIFY2(x)
51 #define RTNAME_STRING(x) RTNAME_STRINGIFY(RTNAME(x))
53 #define DEBUG_TYPE "flang-simplify-intrinsics"
57 class SimplifyIntrinsicsPass
58 : public fir::impl::SimplifyIntrinsicsBase<SimplifyIntrinsicsPass> {
59 using FunctionTypeGeneratorTy =
60 llvm::function_ref<mlir::FunctionType(fir::FirOpBuilder &)>;
61 using FunctionBodyGeneratorTy =
62 llvm::function_ref<void(fir::FirOpBuilder &, mlir::func::FuncOp &)>;
63 using GenReductionBodyTy = llvm::function_ref<void(
64 fir::FirOpBuilder &builder, mlir::func::FuncOp &funcOp, unsigned rank)>;
67 /// Generate a new function implementing a simplified version
68 /// of a Fortran runtime function defined by \p basename name.
69 /// \p typeGenerator is a callback that generates the new function's type.
70 /// \p bodyGenerator is a callback that generates the new function's body.
71 /// The new function is created in the \p builder's Module.
72 mlir::func::FuncOp getOrCreateFunction(fir::FirOpBuilder &builder,
73 const mlir::StringRef &basename,
74 FunctionTypeGeneratorTy typeGenerator,
75 FunctionBodyGeneratorTy bodyGenerator);
76 void runOnOperation() override;
77 void getDependentDialects(mlir::DialectRegistry ®istry) const override;
80 /// Helper function to replace a reduction type of call with its
81 /// simplified form. The actual function is generated using a callback
83 /// \p call is the call to be replaced
84 /// \p kindMap is used to create FIROpBuilder
85 /// \p genBodyFunc is the callback that builds the replacement function
86 void simplifyReduction(fir::CallOp call, const fir::KindMapping &kindMap,
87 GenReductionBodyTy genBodyFunc);
92 /// Generate function type for the simplified version of RTNAME(Sum) and
93 /// similar functions with a fir.box<none> type returning \p elementType.
94 static mlir::FunctionType genNoneBoxType(fir::FirOpBuilder &builder,
95 const mlir::Type &elementType) {
96 mlir::Type boxType = fir::BoxType::get(builder.getNoneType());
97 return mlir::FunctionType::get(builder.getContext(), {boxType},
101 using BodyOpGeneratorTy = llvm::function_ref<mlir::Value(
102 fir::FirOpBuilder &, mlir::Location, const mlir::Type &, mlir::Value,
104 using InitValGeneratorTy = llvm::function_ref<mlir::Value(
105 fir::FirOpBuilder &, mlir::Location, const mlir::Type &)>;
107 /// Generate the reduction loop into \p funcOp.
109 /// \p initVal is a function, called to get the initial value for
110 /// the reduction value
111 /// \p genBody is called to fill in the actual reduciton operation
112 /// for example add for SUM, MAX for MAXVAL, etc.
113 /// \p rank is the rank of the input argument.
114 static void genReductionLoop(fir::FirOpBuilder &builder,
115 mlir::func::FuncOp &funcOp,
116 InitValGeneratorTy initVal,
117 BodyOpGeneratorTy genBody, unsigned rank) {
118 auto loc = mlir::UnknownLoc::get(builder.getContext());
119 mlir::Type elementType = funcOp.getResultTypes()[0];
120 builder.setInsertionPointToEnd(funcOp.addEntryBlock());
122 mlir::IndexType idxTy = builder.getIndexType();
124 mlir::Block::BlockArgListType args = funcOp.front().getArguments();
125 mlir::Value arg = args[0];
127 mlir::Value zeroIdx = builder.createIntegerConstant(loc, idxTy, 0);
129 fir::SequenceType::Shape flatShape(rank,
130 fir::SequenceType::getUnknownExtent());
131 mlir::Type arrTy = fir::SequenceType::get(flatShape, elementType);
132 mlir::Type boxArrTy = fir::BoxType::get(arrTy);
133 mlir::Value array = builder.create<fir::ConvertOp>(loc, boxArrTy, arg);
134 mlir::Value init = initVal(builder, loc, elementType);
136 llvm::SmallVector<mlir::Value, 15> bounds;
138 assert(rank > 0 && "rank cannot be zero");
139 mlir::Value one = builder.createIntegerConstant(loc, idxTy, 1);
141 // Compute all the upper bounds before the loop nest.
142 // It is not strictly necessary for performance, since the loop nest
143 // does not have any store operations and any LICM optimization
144 // should be able to optimize the redundancy.
145 for (unsigned i = 0; i < rank; ++i) {
146 mlir::Value dimIdx = builder.createIntegerConstant(loc, idxTy, i);
148 builder.create<fir::BoxDimsOp>(loc, idxTy, idxTy, idxTy, array, dimIdx);
149 mlir::Value len = dims.getResult(1);
150 // We use C indexing here, so len-1 as loopcount
151 mlir::Value loopCount = builder.create<mlir::arith::SubIOp>(loc, len, one);
152 bounds.push_back(loopCount);
155 // Create a loop nest consisting of DoLoopOp operations.
156 // Collect the loops' induction variables into indices array,
157 // which will be used in the innermost loop to load the input
159 // The loops are generated such that the innermost loop processes
161 llvm::SmallVector<mlir::Value, 15> indices;
162 for (unsigned i = rank; 0 < i; --i) {
163 mlir::Value step = one;
164 mlir::Value loopCount = bounds[i - 1];
165 auto loop = builder.create<fir::DoLoopOp>(loc, zeroIdx, loopCount, step,
167 /*finalCountValue=*/false, init);
168 init = loop.getRegionIterArgs()[0];
169 indices.push_back(loop.getInductionVar());
170 // Set insertion point to the loop body so that the next loop
171 // is inserted inside the current one.
172 builder.setInsertionPointToStart(loop.getBody());
175 // Reverse the indices such that they are ordered as:
176 // <dim-0-idx, dim-1-idx, ...>
177 std::reverse(indices.begin(), indices.end());
179 // We are in the innermost loop: generate the reduction body.
180 mlir::Type eleRefTy = builder.getRefType(elementType);
182 builder.create<fir::CoordinateOp>(loc, eleRefTy, array, indices);
183 mlir::Value elem = builder.create<fir::LoadOp>(loc, addr);
185 mlir::Value reductionVal = genBody(builder, loc, elementType, elem, init);
187 // Unwind the loop nest and insert ResultOp on each level
188 // to return the updated value of the reduction to the enclosing
190 for (unsigned i = 0; i < rank; ++i) {
191 auto result = builder.create<fir::ResultOp>(loc, reductionVal);
192 // Proceed to the outer loop.
193 auto loop = mlir::cast<fir::DoLoopOp>(result->getParentOp());
194 reductionVal = loop.getResult(0);
195 // Set insertion point after the loop operation that we have
197 builder.setInsertionPointAfter(loop.getOperation());
200 // End of loop nest. The insertion point is after the outermost loop.
201 // Return the reduction value from the function.
202 builder.create<mlir::func::ReturnOp>(loc, reductionVal);
205 /// Generate function body of the simplified version of RTNAME(Sum)
206 /// with signature provided by \p funcOp. The caller is responsible
207 /// for saving/restoring the original insertion point of \p builder.
208 /// \p funcOp is expected to be empty on entry to this function.
209 /// \p rank specifies the rank of the input argument.
210 static void genRuntimeSumBody(fir::FirOpBuilder &builder,
211 mlir::func::FuncOp &funcOp, unsigned rank) {
212 // function RTNAME(Sum)<T>x<rank>_simplified(arr)
213 // T, dimension(:) :: arr
216 // do iter = 0, extent(arr)
217 // sum = sum + arr[iter]
219 // RTNAME(Sum)<T>x<rank>_simplified = sum
220 // end function RTNAME(Sum)<T>x<rank>_simplified
221 auto zero = [](fir::FirOpBuilder builder, mlir::Location loc,
222 mlir::Type elementType) {
223 if (auto ty = elementType.dyn_cast<mlir::FloatType>()) {
224 const llvm::fltSemantics &sem = ty.getFloatSemantics();
225 return builder.createRealConstant(loc, elementType,
226 llvm::APFloat::getZero(sem));
228 return builder.createIntegerConstant(loc, elementType, 0);
231 auto genBodyOp = [](fir::FirOpBuilder builder, mlir::Location loc,
232 mlir::Type elementType, mlir::Value elem1,
233 mlir::Value elem2) -> mlir::Value {
234 if (elementType.isa<mlir::FloatType>())
235 return builder.create<mlir::arith::AddFOp>(loc, elem1, elem2);
236 if (elementType.isa<mlir::IntegerType>())
237 return builder.create<mlir::arith::AddIOp>(loc, elem1, elem2);
239 llvm_unreachable("unsupported type");
243 genReductionLoop(builder, funcOp, zero, genBodyOp, rank);
246 static void genRuntimeMaxvalBody(fir::FirOpBuilder &builder,
247 mlir::func::FuncOp &funcOp, unsigned rank) {
248 auto init = [](fir::FirOpBuilder builder, mlir::Location loc,
249 mlir::Type elementType) {
250 if (auto ty = elementType.dyn_cast<mlir::FloatType>()) {
251 const llvm::fltSemantics &sem = ty.getFloatSemantics();
252 return builder.createRealConstant(
253 loc, elementType, llvm::APFloat::getLargest(sem, /*Negative=*/true));
255 unsigned bits = elementType.getIntOrFloatBitWidth();
256 int64_t minInt = llvm::APInt::getSignedMinValue(bits).getSExtValue();
257 return builder.createIntegerConstant(loc, elementType, minInt);
260 auto genBodyOp = [](fir::FirOpBuilder builder, mlir::Location loc,
261 mlir::Type elementType, mlir::Value elem1,
262 mlir::Value elem2) -> mlir::Value {
263 if (elementType.isa<mlir::FloatType>())
264 return builder.create<mlir::arith::MaxFOp>(loc, elem1, elem2);
265 if (elementType.isa<mlir::IntegerType>())
266 return builder.create<mlir::arith::MaxSIOp>(loc, elem1, elem2);
268 llvm_unreachable("unsupported type");
271 genReductionLoop(builder, funcOp, init, genBodyOp, rank);
274 /// Generate function type for the simplified version of RTNAME(DotProduct)
275 /// operating on the given \p elementType.
276 static mlir::FunctionType genRuntimeDotType(fir::FirOpBuilder &builder,
277 const mlir::Type &elementType) {
278 mlir::Type boxType = fir::BoxType::get(builder.getNoneType());
279 return mlir::FunctionType::get(builder.getContext(), {boxType, boxType},
283 /// Generate function body of the simplified version of RTNAME(DotProduct)
284 /// with signature provided by \p funcOp. The caller is responsible
285 /// for saving/restoring the original insertion point of \p builder.
286 /// \p funcOp is expected to be empty on entry to this function.
287 /// \p arg1ElementTy and \p arg2ElementTy specify elements types
288 /// of the underlying array objects - they are used to generate proper
289 /// element accesses.
290 static void genRuntimeDotBody(fir::FirOpBuilder &builder,
291 mlir::func::FuncOp &funcOp,
292 mlir::Type arg1ElementTy,
293 mlir::Type arg2ElementTy) {
294 // function RTNAME(DotProduct)<T>_simplified(arr1, arr2)
295 // T, dimension(:) :: arr1, arr2
298 // do iter = 0, extent(arr1)
299 // product = product + arr1[iter] * arr2[iter]
301 // RTNAME(ADotProduct)<T>_simplified = product
302 // end function RTNAME(DotProduct)<T>_simplified
303 auto loc = mlir::UnknownLoc::get(builder.getContext());
304 mlir::Type resultElementType = funcOp.getResultTypes()[0];
305 builder.setInsertionPointToEnd(funcOp.addEntryBlock());
307 mlir::IndexType idxTy = builder.getIndexType();
310 resultElementType.isa<mlir::FloatType>()
311 ? builder.createRealConstant(loc, resultElementType, 0.0)
312 : builder.createIntegerConstant(loc, resultElementType, 0);
314 mlir::Block::BlockArgListType args = funcOp.front().getArguments();
315 mlir::Value arg1 = args[0];
316 mlir::Value arg2 = args[1];
318 mlir::Value zeroIdx = builder.createIntegerConstant(loc, idxTy, 0);
320 fir::SequenceType::Shape flatShape = {fir::SequenceType::getUnknownExtent()};
321 mlir::Type arrTy1 = fir::SequenceType::get(flatShape, arg1ElementTy);
322 mlir::Type boxArrTy1 = fir::BoxType::get(arrTy1);
323 mlir::Value array1 = builder.create<fir::ConvertOp>(loc, boxArrTy1, arg1);
324 mlir::Type arrTy2 = fir::SequenceType::get(flatShape, arg2ElementTy);
325 mlir::Type boxArrTy2 = fir::BoxType::get(arrTy2);
326 mlir::Value array2 = builder.create<fir::ConvertOp>(loc, boxArrTy2, arg2);
327 // This version takes the loop trip count from the first argument.
328 // If the first argument's box has unknown (at compilation time)
329 // extent, then it may be better to take the extent from the second
330 // argument - so that after inlining the loop may be better optimized, e.g.
331 // fully unrolled. This requires generating two versions of the simplified
332 // function and some analysis at the call site to choose which version
333 // is more profitable to call.
334 // Note that we can assume that both arguments have the same extent.
336 builder.create<fir::BoxDimsOp>(loc, idxTy, idxTy, idxTy, array1, zeroIdx);
337 mlir::Value len = dims.getResult(1);
338 mlir::Value one = builder.createIntegerConstant(loc, idxTy, 1);
339 mlir::Value step = one;
341 // We use C indexing here, so len-1 as loopcount
342 mlir::Value loopCount = builder.create<mlir::arith::SubIOp>(loc, len, one);
343 auto loop = builder.create<fir::DoLoopOp>(loc, zeroIdx, loopCount, step,
345 /*finalCountValue=*/false, zero);
346 mlir::Value sumVal = loop.getRegionIterArgs()[0];
349 mlir::OpBuilder::InsertPoint loopEndPt = builder.saveInsertionPoint();
350 builder.setInsertionPointToStart(loop.getBody());
352 mlir::Type eleRef1Ty = builder.getRefType(arg1ElementTy);
353 mlir::Value index = loop.getInductionVar();
355 builder.create<fir::CoordinateOp>(loc, eleRef1Ty, array1, index);
356 mlir::Value elem1 = builder.create<fir::LoadOp>(loc, addr1);
357 // Convert to the result type.
358 elem1 = builder.create<fir::ConvertOp>(loc, resultElementType, elem1);
360 mlir::Type eleRef2Ty = builder.getRefType(arg2ElementTy);
362 builder.create<fir::CoordinateOp>(loc, eleRef2Ty, array2, index);
363 mlir::Value elem2 = builder.create<fir::LoadOp>(loc, addr2);
364 // Convert to the result type.
365 elem2 = builder.create<fir::ConvertOp>(loc, resultElementType, elem2);
367 if (resultElementType.isa<mlir::FloatType>())
368 sumVal = builder.create<mlir::arith::AddFOp>(
369 loc, builder.create<mlir::arith::MulFOp>(loc, elem1, elem2), sumVal);
370 else if (resultElementType.isa<mlir::IntegerType>())
371 sumVal = builder.create<mlir::arith::AddIOp>(
372 loc, builder.create<mlir::arith::MulIOp>(loc, elem1, elem2), sumVal);
374 llvm_unreachable("unsupported type");
376 builder.create<fir::ResultOp>(loc, sumVal);
378 builder.restoreInsertionPoint(loopEndPt);
380 mlir::Value resultVal = loop.getResult(0);
381 builder.create<mlir::func::ReturnOp>(loc, resultVal);
384 mlir::func::FuncOp SimplifyIntrinsicsPass::getOrCreateFunction(
385 fir::FirOpBuilder &builder, const mlir::StringRef &baseName,
386 FunctionTypeGeneratorTy typeGenerator,
387 FunctionBodyGeneratorTy bodyGenerator) {
388 // WARNING: if the function generated here changes its signature
389 // or behavior (the body code), we should probably embed some
390 // versioning information into its name, otherwise libraries
391 // statically linked with older versions of Flang may stop
392 // working with object files created with newer Flang.
393 // We can also avoid this by using internal linkage, but
394 // this may increase the size of final executable/shared library.
395 std::string replacementName = mlir::Twine{baseName, "_simplified"}.str();
396 mlir::ModuleOp module = builder.getModule();
397 // If we already have a function, just return it.
398 mlir::func::FuncOp newFunc =
399 fir::FirOpBuilder::getNamedFunction(module, replacementName);
400 mlir::FunctionType fType = typeGenerator(builder);
402 assert(newFunc.getFunctionType() == fType &&
403 "type mismatch for simplified function");
407 // Need to build the function!
408 auto loc = mlir::UnknownLoc::get(builder.getContext());
410 fir::FirOpBuilder::createFunction(loc, module, replacementName, fType);
411 auto inlineLinkage = mlir::LLVM::linkage::Linkage::LinkonceODR;
413 mlir::LLVM::LinkageAttr::get(builder.getContext(), inlineLinkage);
414 newFunc->setAttr("llvm.linkage", linkage);
416 // Save the position of the original call.
417 mlir::OpBuilder::InsertPoint insertPt = builder.saveInsertionPoint();
419 bodyGenerator(builder, newFunc);
421 // Now back to where we were adding code earlier...
422 builder.restoreInsertionPoint(insertPt);
427 fir::ConvertOp expectConvertOp(mlir::Value val) {
428 if (fir::ConvertOp op =
429 mlir::dyn_cast_or_null<fir::ConvertOp>(val.getDefiningOp()))
431 LLVM_DEBUG(llvm::dbgs() << "Didn't find expected fir::ConvertOp\n");
435 static bool isOperandAbsent(mlir::Value val) {
436 if (auto op = expectConvertOp(val)) {
437 assert(op->getOperands().size() != 0);
438 return mlir::isa_and_nonnull<fir::AbsentOp>(
439 op->getOperand(0).getDefiningOp());
444 static bool isZero(mlir::Value val) {
445 if (auto op = expectConvertOp(val)) {
446 assert(op->getOperands().size() != 0);
447 if (mlir::Operation *defOp = op->getOperand(0).getDefiningOp())
448 return mlir::matchPattern(defOp, mlir::m_Zero());
453 static mlir::Value findBoxDef(mlir::Value val) {
454 if (auto op = expectConvertOp(val)) {
455 assert(op->getOperands().size() != 0);
456 if (auto box = mlir::dyn_cast_or_null<fir::EmboxOp>(
457 op->getOperand(0).getDefiningOp()))
458 return box.getResult();
459 if (auto box = mlir::dyn_cast_or_null<fir::ReboxOp>(
460 op->getOperand(0).getDefiningOp()))
461 return box.getResult();
466 static unsigned getDimCount(mlir::Value val) {
467 // In order to find the dimensions count, we look for EmboxOp/ReboxOp
468 // and take the count from its *result* type. Note that in case
469 // of sliced emboxing the operand and the result of EmboxOp/ReboxOp
470 // have different types.
471 // Actually, we can take the box type from the operand of
472 // the first ConvertOp that has non-opaque box type that we meet
473 // going through the ConvertOp chain.
474 if (mlir::Value emboxVal = findBoxDef(val))
475 if (auto boxTy = emboxVal.getType().dyn_cast<fir::BoxType>())
476 if (auto seqTy = boxTy.getEleTy().dyn_cast<fir::SequenceType>())
477 return seqTy.getDimension();
481 /// Given the call operation's box argument \p val, discover
482 /// the element type of the underlying array object.
483 /// \returns the element type or llvm::None if the type cannot
484 /// be reliably found.
485 /// We expect that the argument is a result of fir.convert
486 /// with the destination type of !fir.box<none>.
487 static llvm::Optional<mlir::Type> getArgElementType(mlir::Value val) {
488 mlir::Operation *defOp;
490 defOp = val.getDefiningOp();
491 // Analyze only sequences of convert operations.
492 if (!mlir::isa<fir::ConvertOp>(defOp))
494 val = defOp->getOperand(0);
495 // The convert operation is expected to convert from one
496 // box type to another box type.
497 auto boxType = val.getType().cast<fir::BoxType>();
498 auto elementType = fir::unwrapSeqOrBoxedSeqType(boxType);
499 if (!elementType.isa<mlir::NoneType>())
504 void SimplifyIntrinsicsPass::simplifyReduction(fir::CallOp call,
505 const fir::KindMapping &kindMap,
506 GenReductionBodyTy genBodyFunc) {
507 mlir::SymbolRefAttr callee = call.getCalleeAttr();
508 mlir::Operation::operand_range args = call.getArgs();
509 // args[1] and args[2] are source filename and line number, ignored.
510 const mlir::Value &dim = args[3];
511 const mlir::Value &mask = args[4];
512 // dim is zero when it is absent, which is an implementation
513 // detail in the runtime library.
514 bool dimAndMaskAbsent = isZero(dim) && isOperandAbsent(mask);
515 unsigned rank = getDimCount(args[0]);
516 if (dimAndMaskAbsent && rank > 0) {
517 mlir::Location loc = call.getLoc();
518 fir::FirOpBuilder builder(call, kindMap);
520 // Support only floating point and integer results now.
521 mlir::Type resultType = call.getResult(0).getType();
522 if (!resultType.isa<mlir::FloatType>() &&
523 !resultType.isa<mlir::IntegerType>())
526 auto argType = getArgElementType(args[0]);
529 assert(*argType == resultType &&
530 "Argument/result types mismatch in reduction");
532 auto typeGenerator = [&resultType](fir::FirOpBuilder &builder) {
533 return genNoneBoxType(builder, resultType);
535 auto bodyGenerator = [&rank, &genBodyFunc](fir::FirOpBuilder &builder,
536 mlir::func::FuncOp &funcOp) {
537 genBodyFunc(builder, funcOp, rank);
539 // Mangle the function name with the rank value as "x<rank>".
540 std::string funcName =
541 (mlir::Twine{callee.getLeafReference().getValue(), "x"} +
544 mlir::func::FuncOp newFunc =
545 getOrCreateFunction(builder, funcName, typeGenerator, bodyGenerator);
547 builder.create<fir::CallOp>(loc, newFunc, mlir::ValueRange{args[0]});
548 call->replaceAllUsesWith(newCall.getResults());
549 call->dropAllReferences();
554 void SimplifyIntrinsicsPass::runOnOperation() {
555 LLVM_DEBUG(llvm::dbgs() << "=== Begin " DEBUG_TYPE " ===\n");
556 mlir::ModuleOp module = getOperation();
557 fir::KindMapping kindMap = fir::getKindMapping(module);
558 module.walk([&](mlir::Operation *op) {
559 if (auto call = mlir::dyn_cast<fir::CallOp>(op)) {
560 if (mlir::SymbolRefAttr callee = call.getCalleeAttr()) {
561 mlir::StringRef funcName = callee.getLeafReference().getValue();
562 // Replace call to runtime function for SUM when it has single
563 // argument (no dim or mask argument) for 1D arrays with either
564 // Integer4 or Real8 types. Other forms are ignored.
565 // The new function is added to the module.
567 // Prototype for runtime call (from sum.cpp):
568 // RTNAME(Sum<T>)(const Descriptor &x, const char *source, int line,
569 // int dim, const Descriptor *mask)
571 if (funcName.startswith(RTNAME_STRING(Sum))) {
572 simplifyReduction(call, kindMap, genRuntimeSumBody);
575 if (funcName.startswith(RTNAME_STRING(DotProduct))) {
576 LLVM_DEBUG(llvm::dbgs() << "Handling " << funcName << "\n");
577 LLVM_DEBUG(llvm::dbgs() << "Call operation:\n"; op->dump();
578 llvm::dbgs() << "\n");
579 mlir::Operation::operand_range args = call.getArgs();
580 const mlir::Value &v1 = args[0];
581 const mlir::Value &v2 = args[1];
582 mlir::Location loc = call.getLoc();
583 fir::FirOpBuilder builder(op, kindMap);
585 mlir::Type type = call.getResult(0).getType();
586 if (!type.isa<mlir::FloatType>() && !type.isa<mlir::IntegerType>())
589 // Try to find the element types of the boxed arguments.
590 auto arg1Type = getArgElementType(v1);
591 auto arg2Type = getArgElementType(v2);
593 if (!arg1Type || !arg2Type)
596 // Support only floating point and integer arguments
597 // now (e.g. logical is skipped here).
598 if (!arg1Type->isa<mlir::FloatType>() &&
599 !arg1Type->isa<mlir::IntegerType>())
601 if (!arg2Type->isa<mlir::FloatType>() &&
602 !arg2Type->isa<mlir::IntegerType>())
605 auto typeGenerator = [&type](fir::FirOpBuilder &builder) {
606 return genRuntimeDotType(builder, type);
608 auto bodyGenerator = [&arg1Type,
609 &arg2Type](fir::FirOpBuilder &builder,
610 mlir::func::FuncOp &funcOp) {
611 genRuntimeDotBody(builder, funcOp, *arg1Type, *arg2Type);
614 // Suffix the function name with the element types
616 std::string typedFuncName(funcName);
617 llvm::raw_string_ostream nameOS(typedFuncName);
619 arg1Type->print(nameOS);
621 arg2Type->print(nameOS);
623 mlir::func::FuncOp newFunc = getOrCreateFunction(
624 builder, typedFuncName, typeGenerator, bodyGenerator);
625 auto newCall = builder.create<fir::CallOp>(loc, newFunc,
626 mlir::ValueRange{v1, v2});
627 call->replaceAllUsesWith(newCall.getResults());
628 call->dropAllReferences();
631 LLVM_DEBUG(llvm::dbgs() << "Replaced with:\n"; newCall.dump();
632 llvm::dbgs() << "\n");
635 if (funcName.startswith(RTNAME_STRING(Maxval))) {
636 simplifyReduction(call, kindMap, genRuntimeMaxvalBody);
642 LLVM_DEBUG(llvm::dbgs() << "=== End " DEBUG_TYPE " ===\n");
645 void SimplifyIntrinsicsPass::getDependentDialects(
646 mlir::DialectRegistry ®istry) const {
647 // LLVM::LinkageAttr creation requires that LLVM dialect is loaded.
648 registry.insert<mlir::LLVM::LLVMDialect>();
650 std::unique_ptr<mlir::Pass> fir::createSimplifyIntrinsicsPass() {
651 return std::make_unique<SimplifyIntrinsicsPass>();