Supporting Arrays and Allocatables in LFortran

Supporting Arrays and Allocatables in LFortran

• 1.
  Gagandeep Singh, SoftwareDeveloper (Quansight) [email protected] Supporting Arrays and Allocatables in LFortran International Fortran Conference, 2021
• 2.
  Overview • Background • ArrayDeclaration • Operations involving arrays • Allocatable arrays • Array as input and output to functions/subroutines • Automatic Deallocation
• 3.
  Background Internal representations ofcode used by LFortran • Abstract Syntax Tree (AST) - Contains all the syntax information in the input Fortran code. Each statement can be interpreted as a tree and then the whole code is just a forest of these trees. • Abstract Semantic Representation (ASR) - Contains all the semantic information such as symbol tables (containing functions, variables, references to module elements, etc.). All the heavy lifting (type checking, implicit casting) is done here. • Backend - Receives ASR as input and generates the code in desired language (LLVM, C++, etc.). My work involved dealing with LLVM backend. • ASR to ASR Passes - Takes ASR as input and transforms it into an equivalent ASR. For example, converting all the loops to while loops, select-case to if-else if-if ladders which helps in simplifying backend.
• 4.
  Array Declaration • Allthe dimensional and type information was already available in ASR representation of input code. • We used a structure to represent arrays in LLVM IR. It contains the following information, ArrayType* array - Pointer to 1D memory block containing the data. int64_t offset - This contains the o ff set value. As of now this is always set to 0. dim - This is simply the array of dimension_descriptor structure specifying the details of each dimension.
• 5.
  Array Declaration • Thedimension_descriptor structure has the following elements, lower_bound upper_bound size - size of the current dimension • For allocatable arrays, a 1 bit integer is also added to the array descriptor. It is used to check whether the pointer, ArrayType* array, is freed or not.
• 6.
  Operations involving Arrays •General approach - Convert any array operation to loops. For example, c = a + b is converted to a loop, c(i) = a(i) + b(i), for an iterator variable i. • Achieved by writing ASR to ASR passes. Input ASR pass contains original expressions with operations on arrays and the output ASR contains the loops implementing those operations.
• 7.
  Allocatable arrays • Thedescriptor for allocatable arrays is same as for “normal” arrays but contain an extra 1 bit integer to keep track whether the memory allocated is freed or not. • We use malloc in C to allocate memory on heap. It is called in LLVM IR. • Similarly, we use free in C to deallocate the memory allocated previously. In case of automatic deallocation (discussed later) we use the extra 1 bit integer to decide whether to call free.
• 8.
  Array as inputand output to functions/subroutines • Input to Functions/Subroutines - At LLVM level, pointer to the original array descriptor as passed as input to functions/subroutines. • Output from Subroutines - Pointer to array descriptor is passed which can be modi fi ed by the subroutine. • Output from Function - The function is fi rst converted to a subroutine with the array being returned as intent(out) argument. Then any call to this function is converted to a subroutine call. Achieved by writing ASR to ASR pass. Helps in avoiding copying date from temporary return variable to the desired destination variable.
• 9.
  Automatic Deallocation • Motivation Freethe memory on heap while leaving a scope (module, function, subroutine, program, etc.). Avoid double frees if already done explicitly by the user. Before calling a function/subroutine, automatically deallocate arrays with intent(out), allocatable attributes. • An ImplicitDeallocate node is appended to all the scopes in a ASR pass. It keeps track of only local variables. For example, input/output to a function/subroutine won’t be a ff ected.