2 * Proposal for the local register allocator
4 The local register allocator deals with allocating registers
5 for temporaries inside a single basic block, while the global
6 register allocator is concerned with method-wide allocation of
8 The global register allocator uses callee-saved register for it's
9 purpouse so that there is no need to save and restore these registers
12 There are a number of issues the local allocator needs to deal with:
13 *) some instructions expect operands in specific registers (for example
14 the shl instruction on x86, or the call instruction with thiscall
15 convention, or the equivalent call instructions on other architectures,
16 such as the need to put output registers in %oX on sparc)
17 *) some instructions deliver results only in specific registers (for example
18 the div instruction on x86, or the call instructionson on almost all
20 *) it needs to know what registers may be clobbered by an instruction
21 (such as in a method call)
22 *) it should avoid excessive reloads or stores to improve performance
24 While which specific instructions have limitations is architecture-dependent,
25 the problem shold be solved in an arch-independent way to reduce code duplication.
26 The register allocator will be 'driven' by the arch-dependent code, but it's
27 implementation should be arch-independent.
29 To improve the current local register allocator, we need to
30 keep more state in it than the current setup that only keeps busy/free info.
32 Possible state information is:
34 free: the resgister is free to use and it doesn't contain useful info
35 freeable: the register contains data loaded from a local (there is
36 also info about _which_ local it contains) as a result from previous
37 instructions (like, there was a store from the register to the local)
38 moveable: it contains live data that is needed in a following instruction, but
39 the contents may be moved to a different register
40 busy: the register contains live data and it is placed there because
41 the following instructions need it exactly in that register
42 allocated: the register is used by the global allocator
44 The local register allocator will have the following interfaces:
47 Searches for a register in the free state. If it doesn't find it,
48 searches for a freeable register. Sets the status to moveable.
49 Looking for a 'free' register before a freeable one should allow for
50 removing a few redundant loads (though I'm still unsure if such
51 things should be delegated entirely to the peephole pass).
53 int get_register_force (int reg);
54 Returns 'reg' if it is free or freeable. If it is moveable, it moves it
55 to another free or freeable register.
56 Sets the status of 'reg' to busy.
58 void set_register_freeable (int reg);
59 Sets the status of 'reg' to freeable.
61 void set_register_free (int reg);
62 Sets the status of 'reg' to free.
64 void will_clobber (int reg);
65 Spills the register to the stack. Sets the status to freeable.
66 After the clobbering has occurred, set the status to free.
68 void register_unspill (int reg);
69 Un-spills register reg and sets the status to moveable.
71 FIXME: how is the 'local' information represented? Maybe a MonoInst* pointer.
73 Note: the register allocator will insert instructions in the basic block
74 during it's operation.
78 Given the tree (on x86 the right argument to shl needs to be in ecx):
80 store (local1, shl (local1, call (some_arg)))
82 At the start of the basic block, the registers are set to the free state.
83 The sequence of instructions may be:
84 instruction register status -> [%eax %ecx %edx]
86 eax = load local1 mov free free
87 /* call clobbers eax, ecx, edx */
88 spill eax free free free
90 /* now eax contains the right operand of the shl */
91 mov %eax -> %ecx free busy free
92 un-spill mov busy free
93 shl %cl, %eax mov free free
95 The resulting x86 code is:
101 mov $fff0(%ebp), %eax
104 Note that since shl could operate directly on memory, we could have:
111 The above example with loading the operand in a register is just to complicate
112 the example and show that the algorithm should be able to handle it.
114 Let's take another example with the this-call call convention (the first argument
116 In this case, will_clobber() will be called only on %eax and %edx, while %ecx
117 will be allocated with get_register_force ().
118 Note: when a register is allocated with get_register_force(), it should be set
119 to a different state as soon as possible.
121 store (local1, shl (local1, this-call (local1)))
123 instruction register status -> [%eax %ecx %edx]
125 eax = load local1 mov free free
126 /* force load in %ecx */
127 ecx = load local1 mov busy free
128 spill eax free busy free
130 /* now eax contains the right operand of the shl */
131 mov %eax -> %ecx free busy free
132 un-spill mov busy free
133 shl %cl, %eax mov free free
135 What happens when a register that we need to allocate with get_register_force ()
136 contains an operand for the next instruction?
138 instruction register status -> [%eax %ecx %edx]
139 eax = load local0 mov free free
140 ecx = load local1 mov mov free
141 get_register_force (ecx) here.
146 The first option is way better (and allows the peephole pass to
147 just load the value in %edx directly, instead of loading first to %ecx).
148 This doesn't work, though, if the instruction clobbers the %edx register
149 (like in a this-call). So, we first need to clobber the registers
150 (so the state of %ecx changes to freebale and there is no issue
151 with get_register_force ()).
152 What if an instruction both clobbers a register and requires it as
153 an operand? Lets' take the x86 idiv instruction as an example: it
154 requires the dividend in edx:eax and returns the result in eax,
155 with the modulus in edx.
157 store (local1, div (local1, local2))
159 instruction register status -> [%eax %ecx %edx]
160 eax = load local0 mov free free
161 will_clobber eax, edx free mov free
162 force mov %ecx, %eax busy free free
163 set %edx busy free busy
166 Note: edx is set to free after idiv, because the modulus is not needed
167 (if it was a rem, eax would have been freed).
168 If we load the divisor before will_clobber(), we'll have to spill
169 eax and reload it later. If we load it just after the idiv, there is no issue.
170 In any case, the algorithm should give the correct results and allow the operation.
172 Working recursively on the isntructions there shouldn't be huge issues
173 with this algorithm (though, of course, it's not optimal and it may
174 introduce excessive spills or register moves). The advantage over the current
175 local reg allocator is that:
176 1) the number of spills/moves would be smaller anyway
177 2) a separate peephole pass could be able to eliminate reg moves
178 3) we'll be able to remove the 'forced' spills we currently do with
179 the return value of method calls
183 How to best integrate such a reg allocator with the burg stuff.
185 Think about a call os sparc with two arguments: they got into %o0 and %o1
186 and each of them sets the register as busy. But what if the values to put there
187 are themselves the result of a call? %o0 is no problem, but for all the
188 next argument n the above algorithm would spill all the 0...n-1 registers...
192 More complex solutions to the local register allocator problem:
193 http://dimacs.rutgers.edu/TechnicalReports/abstracts/1997/97-33.html
195 Combining register allocation and instruction scheduling:
196 http://citeseer.nj.nec.com/motwani95combining.html
198 More on LRA euristics:
199 http://citeseer.nj.nec.com/liberatore97hardness.html
201 Linear-time optimal code scheduling for delayedload architectures
202 http://www.cs.wisc.edu/~fischer/cs701.f01/inst.sched.ps.gz
204 Precise Register Allocation for Irregular Architectures
205 http://citeseer.nj.nec.com/kong98precise.html
207 Allocate registers first to subtrees that need more of them.
208 http://www.upb.de/cs/ag-kastens/compii/folien/comment401-409.2.pdf