2 * How to handle complex IL opcodes in an arch-independent way
4 Many IL opcodes are very simple: add, ldind etc.
5 Such opcodes can be implemented with a single cpu instruction
6 in most architectures (on some, a group of IL instructions
7 can be converted to a single cpu op).
8 There are many IL opcodes, though, that are more complex, but
9 can be expressed as a series of trees or a single tree of
10 simple operations. Such simple operations are architecture-independent.
11 It makes sense to decompose such complex IL instructions in their
12 simpler equivalent so that we gain in several ways:
13 *) porting effort is easier, because only the simple instructions
14 need to be implemented in arch-specific code
15 *) we could apply BURG rules to the trees and do pattern matching
16 on them to optimize the expressions according to the host cpu
18 The issue is: where do we do such conversion from coarse opcodes to
21 * Doing the conversion in method_to_ir ()
23 Some of these conversions can certainly be done in method_to_ir (),
24 but it's not always easy to decide which are better done there and
25 which in a different pass.
26 For example, let's take ldlen: in the mono implementation, ldlen
27 can be simply implemented with a load from a fixed position in the
30 len = [reg + maxlen_offset]
32 However, ldlen carries also semantics information: the result is the
33 length of the array, and since in the CLR arrays are of fixed size,
34 this information can be useful to later do bounds check removal.
35 If we convert this opcode in method_to_ir () we lost some useful
36 information for further optimizations.
38 In some other ways, decomposing an opcode in method_to_ir() may
39 allow for better optimizations later on (need to come up with an
42 * Doing the conversion in inssel.brg
44 Some conversion may be done inside the burg rules: this has the
45 disadvantage that the instruction selector is not run again on
46 the resulting expression tree and we could miss some optimization
47 (this is what effectively happens with the coarse opcodes in the old
48 jit). This may also interfere with an efficient local register allocator.
49 It may be possible to add an extension in monoburg that allows a rule
52 recheck: LDLEN (reg) {
53 create an expression tree representing LDLEN
57 When the monoburg label process gets back a recheck, it will run
58 the labeling again on the resulting expression tree.
59 If this is possible at all (and in an efficient way) is a
60 question for dietmar:-)
61 It should be noted, though, that this may not always work, since
62 some complex IL opcodes may require a series of expression trees
63 and handling such cases in monoburg could become quite hairy.
64 For example, think of opcode that need to do multiple actions on the
65 same object: this basically means a DUP...
66 On the other end, if a complex opcode needs a DUP, monoburg doesn't
67 actually need to create trees if it emits the instructions in
68 the correct sequence and maintains the right values in the registers
69 (usually the values that need a DUP are not changed...). How
70 this integrates with the current register allocator is not clear, since
71 that assigns registers based on the rule, but the instructions emitted
72 by the rules may be different (this already happens with the current JIT
73 where a MULT is replaced with lea etc...).
75 * Doing it in a separate pass.
77 Doing the conversion in a separate pass over the instructions
78 is another alternative. This can be done right after method_to_ir ()
79 or after the SSA pass (since the IR after the SSA pass should look
80 almost like the IR we get back from method_to_ir ()).
82 This has the following advantages:
83 *) monoburg will handle only the simple opcodes (makes porting easier)
84 *) the instruction selection will be run on all the additional trees
85 *) it's easier to support coarse opcodes that produce multiple expression
86 trees (and apply the monoburg selector on all of them)
87 *) the SSA optimizer will see the original opcodes and will be able to use
88 the semantic info associated with them
90 The disadvantage is that this is a separate pass on the code and
91 it takes time (how much has not been measured yet, though).
93 With this approach, we may also be able to have C implementations
94 of some of the opcodes: this pass would insert a function call to
95 the C implementation (for example in the cases when first porting
96 to a new arch and implemenating some stuff may be too hard in asm).
98 * Extended basic blocks
100 IL code needs a lot of checks, bounds checks, overflow checks,
101 type checks and so on. This potentially increases by a lot
102 the number of basic blocks in a control flow graph. However,
103 all such blocks end up with a throw opcode that gives control to the
104 exception handling mechanism.
105 After method_to_ir () a MonoBasicBlock can be considered a sort
106 of extended basic block where the additional exits don't point
107 to basic blocks in the same procedure (at least when the method
108 doesn't have exception tables).
109 We need to make sure the passes following method_to_ir () can cope
110 with such kinds of extended basic blocks (especially the passes
111 that we need to apply to all the methods: as a start, we could
112 skip SSA optimizations for methods with exception clauses...)