-
Notifications
You must be signed in to change notification settings - Fork 0
/
TODO
59 lines (48 loc) · 2.63 KB
/
TODO
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
- NEW SSA/arg handling:
- run the first phase of SSA_GVN as part of IR construction:
- Get rid of "variables" as instructions in block 0
- read/write from vars X nodes table while filling nodes
- conditionally get rid of "args" as instruction in node 0:
- depending on if we actually have calls or not, we might want to get args out of the way to other regs
- always get rid of "phi" nodes from join nodes blocks
- don't put "putphi" instructions in block as part of phi constructions
- only schedule these as part of "resolve_moveblock" later
- calls:
- vregs+temps now have for all clobbered regs, not just "overlaps a call" bit but we need to use it
- for syscalls: much fewer clobbers
- handle constrained instructions like "shr reg, ECX" properly (note: input-constraint is not a clobber, but it works as a zero-order approx)
- FIXME: general paralell move handling. paralell moves currently needs to be handled:
- block of putphi instructions
- block of callarg instructions
- block of arg instructions (we shouldn't need to do swaps here, but why not support to)
- regalloc:
- implement dummy "spill everything!" allocator just to test spill code in isolation
-implement liveness interval spliting:
- reference: "Linear Scan Register Allocation on SSA Form"; Wimmer, Franz; 2010
- [DONE] allocation when no spilling is needed
- decide how do do spills:
- in advance as part of instruction selection (keeping pressure)
- in the midle
- consider "call near" test case as of 638022887f2f060ecf6eb0aacf211f93a05ed89a:
\\func twokube
\\ %x = arg
\\ %y = arg
\\ %xx = call kuben %x
\\ %yy = call kuben %y
\\ %summa = add %xx %yy
\\ ret %summa
\\end
- this uses two saved regs (for $xx and $y overlapping), but in theory only one is needed
(between the calls, $y could be moved from SAVED to .rdi and then %xx from .rax to SAVED)
- make a test case where we actually hit the register limit
- for this case: allow interval to be split into two distinct registers, whithout any memory spilling
- follow up project: attempt to generalize reverse-mode linear scan to general (at least reducible) CFG:s.
(only worth it if the struct/algo becomes simpler than
WIP: integrate codebase completely with EIRI.
Try zig package manager and see make Eiri depend on FLIR!
pub fn vmovdq_vg(self: *Self, quad: bool, dst: u4, src: IPReg) !void {
try self.new_inst(@returnAddress());
try self.vex3(wide, dst.ext(), src1.ext(), false, .h0F38, src2.id(), false, pp);
}
std.debug.print("\nnee: {}", .{std.os.system.getpid()});
std.time.sleep(1000 * 1e9);