Commit 48f77ccc authored by simonmar's avatar simonmar
Browse files

[project @ 2005-06-15 13:50:14 by simonmar]

Undo accidental commit
parent e6de0678
......@@ -80,58 +80,6 @@ The algorithm is roughly:
Possible plan for x86 floating pt register alloc:
- The standard reg alloc procedure allocates pretend floating point
registers to the GXXX instructions. We need to convert these GXXX
instructions to proper x86 FXXX instructions, using the FP stack for
We could do this in a separate pass, but it helps to have the
information about which real registers are live after the
instruction, so we do it at reg alloc time where that information
is already available.
- keep a mapping from %fakeN to FP stack slot in the monad.
- after assigning registers to the GXXX instruction, convert the
instruction to an FXXX instruction. eg.
- for GMOV just update the mapping, and ffree any dead regs.
- GLD: just fld and update mapping
GLDZ: just fldz and update mapping
GLD1: just fld1 and update mapping
- GST: just fst and update mapping, ffree dead regs.
- special case for GST reg, where reg is st(0), we can fstp.
- for GADD fp1, fp2, fp3:
- easy way: fld fp2
fld fp1
-- record that fp3 is now in %st(0), and all other
-- slots are pushed down one.
ffree fp1 -- if fp1 is dead now
ffree fp2 -- if fp2 is dead now
- optimisation #1
- if fp1 is in %st(0) and is dead afterward
fadd %st(0), fp2
-- record fp3 is in %st(0)
ffree fp2 -- if fp2 is dead now
- if fp2 is in %st(0) and is dead afterward
fadd %st(0), fp1
-- record fp3 is in %st(0)
- if fp1 is in %st(0), fp2 is dead afterward
fadd fp2, %st(0)
-- record fp3 is in fp2's locn
- if fp2 is in %st(0), fp1 is dead afterward
fadd fp1, %st(0)
-- record fp3 is in fp1's locn
- we should be able to avoid the nasty ffree problems of the current
scheme. The stack should be empty before doing a non-local
jump/call - we can assert that this is the case.
module RegisterAlloc (
) where
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment