blob: 44e9fbf0e760aad91de27ec571909270ea411137 [file] [log] [blame]
Chris Lattner34cb1542009-04-01 21:11:04 +00001//===----------------------------------------------------------------------===//
2// Representing sign/zero extension of function results
3//===----------------------------------------------------------------------===//
4
5Mar 25, 2009 - Initial Revision
6
7Most ABIs specify that functions which return small integers do so in a
8specific integer GPR. This is an efficient way to go, but raises the question:
9if the returned value is smaller than the register, what do the high bits hold?
10
11There are three (interesting) possible answers: undefined, zero extended, or
12sign extended. The number of bits in question depends on the data-type that
13the front-end is referencing (typically i1/i8/i16/i32).
14
15Knowing the answer to this is important for two reasons: 1) we want to be able
16to implement the ABI correctly. If we need to sign extend the result according
17to the ABI, we really really do need to do this to preserve correctness. 2)
18this information is often useful for optimization purposes, and we want the
19mid-level optimizers to be able to process this (e.g. eliminate redundant
20extensions).
21
22For example, lets pretend that X86 requires the caller to properly extend the
23result of a return (I'm not sure this is the case, but the argument doesn't
24depend on this). Given this, we should compile this:
25
26int a();
27short b() { return a(); }
28
29into:
30
31_b:
32 subl $12, %esp
33 call L_a$stub
34 addl $12, %esp
35 cwtl
36 ret
37
38An optimization example is that we should be able to eliminate the explicit
39sign extension in this example:
40
41short y();
42int z() {
43 return ((int)y() << 16) >> 16;
44}
45
46_z:
47 subl $12, %esp
48 call _y
49 ;; movswl %ax, %eax -> not needed because eax is already sext'd
50 addl $12, %esp
51 ret
52
53//===----------------------------------------------------------------------===//
54// What we have right now.
55//===----------------------------------------------------------------------===//
56
57Currently, these sorts of things are modelled by compiling a function to return
58the small type and a signext/zeroext marker is used. For example, we compile
59Z into:
60
61define i32 @z() nounwind {
62entry:
63 %0 = tail call signext i16 (...)* @y() nounwind
64 %1 = sext i16 %0 to i32
65 ret i32 %1
66}
67
68and b into:
69
70define signext i16 @b() nounwind {
71entry:
72 %0 = tail call i32 (...)* @a() nounwind ; <i32> [#uses=1]
73 %retval12 = trunc i32 %0 to i16 ; <i16> [#uses=1]
74 ret i16 %retval12
75}
76
77This has some problems: 1) the actual precise semantics are really poorly
78defined (see PR3779). 2) some targets might want the caller to extend, some
79might want the callee to extend 3) the mid-level optimizer doesn't know the
80size of the GPR, so it doesn't know that %0 is sign extended up to 32-bits
81here, and even if it did, it could not eliminate the sext. 4) the code
82generator has historically assumed that the result is extended to i32, which is
83a problem on PIC16 (and is also probably wrong on alpha and other 64-bit
84targets).
85
86//===----------------------------------------------------------------------===//
87// The proposal
88//===----------------------------------------------------------------------===//
89
90I suggest that we have the front-end fully lower out the ABI issues here to
91LLVM IR. This makes it 100% explicit what is going on and means that there is
92no cause for confusion. For example, the cases above should compile into:
93
94define i32 @z() nounwind {
95entry:
96 %0 = tail call i32 (...)* @y() nounwind
97 %1 = trunc i32 %0 to i16
98 %2 = sext i16 %1 to i32
99 ret i32 %2
100}
101define i32 @b() nounwind {
102entry:
103 %0 = tail call i32 (...)* @a() nounwind
104 %retval12 = trunc i32 %0 to i16
105 %tmp = sext i16 %retval12 to i32
106 ret i32 %tmp
107}
108
109In this model, no functions will return an i1/i8/i16 (and on a x86-64 target
110that extends results to i64, no i32). This solves the ambiguity issue, allows us
111to fully describe all possible ABIs, and now allows the optimizers to reason
112about and eliminate these extensions.
113
114The one thing that is missing is the ability for the front-end and optimizer to
115specify/infer the guarantees provided by the ABI to allow other optimizations.
116For example, in the y/z case, since y is known to return a sign extended value,
117the trunc/sext in z should be eliminable.
118
119This can be done by introducing new sext/zext attributes which mean "I know
120that the result of the function is sign extended at least N bits. Given this,
121and given that it is stuck on the y function, the mid-level optimizer could
122easily eliminate the extensions etc with existing functionality.
123
124The major disadvantage of doing this sort of thing is that it makes the ABI
125lowering stuff even more explicit in the front-end, and that we would like to
126eventually move to having the code generator do more of this work. However,
127the sad truth of the matter is that this is a) unlikely to happen anytime in
128the near future, and b) this is no worse than we have now with the existing
129attributes.
130
131C compilers fundamentally have to reason about the target in many ways.
132This is ugly and horrible, but a fact of life.
133