xref: /llvm-project/flang/docs/fstack-arrays.md (revision 2d416219af5c0091f7887e4d4463e63f5a37d811)
1# Stack arrays pass
2## Problem Description
3In gfortran, `-fstack-arrays` will cause all local arrays, including those of
4unknown size, to be allocated from stack memory. Gfortran enables this flag by
5default at `-Ofast`.
6
7Flang already allocates all local arrays on the stack (the
8`memory-allocation-opt` pass can move large allocations to the heap, but by
9default it will not). But there are some cases where temporary arrays are
10created on the heap by Flang. Moving these temporary arrays is the goal here.
11
12## Proposed Solution
13One approach would be to write a pass which attempts to move heap allocations to
14the stack (like the `memory-allocation-opt` pass in reverse). This approach has
15the advantage of a self-contained implementation and ensuring that newly added
16cases are covered.
17
18Another approach would be to modify each place in which arrays are allocated on
19the heap to instead generate a stack allocation. The advantage of the second
20approach is that it would be difficult to match up all `fir.allocmem` operations
21with their associated `fir.freemem`. In general, the lifetimes of heap allocated
22memory are not constrained by the current function's stack frame and so cannot
23be always converted to stack allocations. It is much easier to swap heap
24allocations for stack allocations when they are first generated because the
25lifetime information is conveniently available.
26
27For example, to rewrite the heap allocation in the `array-value-copy` pass with
28a stack allocation using the first approach would require analysis to ensure
29that the heap allocation is always freed before the function returns. This is
30much more complex than never generating a heap allocation (and free) in the
31first place (the second approach).
32
33The plan is to take the more complex first approach so that newly added changes
34to lowering code do not need to be made to support the stack arrays option. The
35general problem of determining heap allocation lifetimes can be simplified in
36this case because only those allocations which are always freed before exit from
37the same function are possible to be moved to heap allocations in that
38function's stack frame. Furthermore, the aim of this pass would be to only move
39those allocations generated by Flang, and so some of the more difficult cases
40can be avoided. Where the transformation is not possible, the existing heap
41allocation will be kept.
42
43## Implementation details overview
44In order to limit the complexity of the proposed pass, it is useful to
45understand the situations in which Flang will generate heap allocations.
46
47### Known Heap Array Allocations
48Flang allocates most arrays on the stack by default, but there are a few cases
49where temporary arrays are allocated on the heap:
50- `flang/lib/Optimizer/Transforms/ArrayValueCopy.cpp`
51- `flang/lib/Optimizer/Transforms/MemoryAllocation.cpp`
52- `flang/lib/Lower/ConvertExpr.cpp`
53- `flang/lib/Lower/IntrinsicCall.cpp`
54- `flang/lib/Lower/ConvertVariable.cpp`
55
56Lowering code is being updated and in the future, temporaries for expressions
57will be created in the HLFIR bufferization pass in
58`flang/lib/Optimizer/HLFIR/Trnasforms/BufferizeHLFIR.cpp`.
59
60#### `ArrayValueCopy.cpp`
61Memory is allocated for a temporary array in `allocateArrayTemp()`. This
62temporary array is used to ensure that assignments of one array to itself
63produce the required value. E.g.
64
65```
66integer, dimension(5), intent(inout) :: x
67x(3,4) = x(1,2)
68```
69
70#### `MemoryAllocation.cpp`
71The default options for the Memory Allocation transformation ensure that no
72array allocations, no matter how large, are moved from the stack to the heap.
73
74#### `ConvertExpr.cpp`
75`ConvertExpr.cpp` allocates many array temporaries on the heap:
76  - Passing array arguments by value or when they need re-shaping
77  - Lowering elemental array expressions
78  - Lowering mask expressions
79  - Array constructors
80
81The last two of these cases are **not** covered by the current stack arrays pass
82design.
83
84The FIR code generated for mask expressions (the WHERE construct) sets a
85boolean variable to indicate whether a heap allocation was necessary. The
86allocation is only freed if the variable indicates that the allocation was
87performed to begin with. The proposed dataflow analysis is not intelligent
88enough to statically determine that the boolean variable will always be true
89when the allocation is performed. Beyond this, the control flow in the generated
90FIR code passes the allocated memory through `fir.result`, resulting in a
91different SSA value to be allocated and freed, causing the analysis not to
92realise that the allocated memory is freed. The most convenient solution here
93would be to generate less complicated FIR code, as the existing codegen has
94known bugs: https://github.com/llvm/llvm-project/issues/56921,
95https://github.com/llvm/llvm-project/issues/59803.
96
97Code generated for array constructors uses `realloc()` to grow the allocated
98buffer because the size of the resulting array cannot always be determined
99ahead of running the constructor. This makes this temporary unsuitable
100for allocation on the stack.
101
102#### `IntrinsicCall.cpp`
103The existing design is for the runtime to do the allocation and the lowering
104code to insert `fir.freemem` to remove the allocation. It is not clear whether
105this can be replaced by adapting lowering code to do stack allocations and to
106pass these to the runtime. This would be a significant change and so is out of
107scope of `-fstack-arrays`.
108
109#### `ConvertVariable.cpp`
110See `Fortran::lower::MapSymbolAttributes`:
111
112Dummy arguments that are not declared in the current entry point require a
113skeleton definition. Most such "unused" dummies will not survive into final
114generated code, but some will. It is illegal to reference one at run time if it
115does. There are three ways these dummies can be mapped to a value:
116- a `fir::UndefOp` value: preferable but can't be used for dummies with dynamic
117  bounds or used to define another dummy.
118- A stack slot. This has intermediate-weight cost but may not be usable for
119  objects with dynamic bounds.
120- A heap box/descriptor is the heaviest weight option, only used for dynamic
121  objects.
122
123These heap allocations will be out of scope for `-fstack-arrays` because the
124existing implementation already uses stack allocations (or better) where
125possible, and most such dummy arguments will be removed from generated code.
126
127#### `BufferizeHLFIR.cpp`
128TODO
129
130### Detecting Allocations to Move
131Allocations which could be moved to the stack will be detected by performing a
132forward dense data flow analysis using `mlir::dataflow::DenseForwardDataFlowAnalysis`.
133This analysis will search for SSA values created by a `fir.allocmem` which are
134always freed using `fir.freemem` within the same function.
135
136Tracking allocations by SSA values will miss some cases where address
137calculations are duplicated in different blocks: resulting in different SSA
138values as arguments for the allocation and the free. It is believed that simple
139tracking of SSA values is sufficient for detecting the allocations for array
140temporaries added by Flang, because these temporaries should be simple arrays:
141not nested inside of derived types or other arrays.
142
143Arrays allocated by an `allocate` statement in source code should not be moved
144to the stack. These will be identified by adding an attribute to these
145`fir.allocmem` operations when they are lowered. Intrinsics such as `allocated`
146and `move_alloc` do not need to be supported because the array temporaries moved
147to the stack will never be visible in user code.
148
149Only allocations for arrays will be considered for moving to the stack.
150
151When conducting the dense data flow analysis, the pass will assume that existing
152allocations are not freed inside called functions. This is true for the existing
153heap array temporary allocations generated by Flang. If an allocation were to be
154freed inside of a called function, the allocation would presumably not also be
155freed later in the caller function (as this would be a double-free). Therefore
156the stack arrays pass would not find where the allocation was freed and so not
157transform the allocation. It is necessary to make this assumption so that the
158stack arrays pass can transform array allocations created for pass-by-value
159function arguments.
160
161### Moving Allocations
162The `fir.allocmem` will be replaced by a `fir.alloca` with the same arguments.
163The `fir.freemem` will be removed entirely.
164
165Where possible, the `fir.alloca` should be placed in the function entry block
166(so we can be sure it only happens once). This may not be possible if the array
167has non-constant extents (causing the `fir.alloca` to have SSA values as
168operands). In this case, the `fir.alloca` will be placed immediately after the
169last operand becomes available.
170
171If this location is inside a loop (either an `mlir::LoopLikeOpInterface` or a
172cyclic CFG), the transformation should attempt to use the `llvm.stacksave`/
173`llvm.stackrestore` intrinsics to ensure that the stack does not grow on every
174loop iteration. Use of these intrinsics is considered valid when the allocation
175and deallocation occur within the same block and there are no other stack
176allocations between them. If this is not the case, the existing heap allocation
177will be preserved.
178
179### FIR attribute
180A FIR attribute will be added to distinguish `fir.allocmem` arising from
181`allocate` statements in source from `fir.allocmem` operations  added by Flang.
182The attribute will be called `"fir.must_be_heap"` and will have a boolean value:
183`true` meaning that the stack arrays pass must not move the allocation, `false`
184meaning that stack arrays may move the allocation. Not specifying the attribute
185will be equivalent to setting it to `false`.
186
187## Testing Plan
188FileCheck tests will be written to check each of the above identified sources of
189heap allocated array temporaries are detected and converted by the new pass.
190
191Another test will check that `allocate` statements in source code will not be
192moved to the stack.
193