ShuriZma
/
xenia
mirror of https://github.com/xenia-project/xenia.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity
xenia/TODO.md

87 lines
1.7 KiB
Markdown
Raw Normal View History

Adding TODO.
2013-01-22 08:54:49 +00:00
## Instructions
```
stbu
lwzu
addx
addix
addic
addzex
subfx
subfex
subficx
mulli
mullwx
divwux
andix
orx
rlwinmx
rldiclx
extsbx
slwx
srawix
Update TODO.md
2013-01-22 21:15:02 +00:00
# can be no-op, or @llvm.prefetch
Adding TODO.
2013-01-22 08:54:49 +00:00
dcbt
dcbtst
Update TODO.md
2013-01-22 21:15:02 +00:00
twi # @llvm.debugtrap ?
Adding TODO.
2013-01-22 08:54:49 +00:00
```
Update TODO.md
2013-01-22 21:15:02 +00:00
Overflow bits can be set via the intrinsics:
`llvm.sadd.with.overflow`/etc
It'd be nice to avoid doing this unless absolutely required. The SDB could
walk functions to see if they ever read or branch on the SO bit of things.
`@llvm.expect.i32`/`.i64` could be used with the BH bits in branches to
indicate expected values.
Adding TODO.
2013-01-22 08:54:49 +00:00
## Codegen
Update TODO.md
2013-01-22 18:50:20 +00:00
### Calling convention
Experiment with fastcc? May need typedef fn ptrs to call into the JITted code.
Adding better register tracking through locals and fixing branches. LLVM does an amazing job of optimizing this. There are many opportunities to make it better, too, by preventing spilling where not required or only spilling/filling things when needed.
2013-01-25 08:32:42 +00:00
### Function calling convention analysis
Update TODO.md
2013-01-22 21:15:02 +00:00
Adding better register tracking through locals and fixing branches. LLVM does an amazing job of optimizing this. There are many opportunities to make it better, too, by preventing spilling where not required or only spilling/filling things when needed.
2013-01-25 08:32:42 +00:00
Track functions to see if they follow the standard calling convention.
This could use the hints from the EH data in the XEX. Looking specifically for
stack prolog/epilog and branches to LR.
Adding TODO.
2013-01-22 08:54:49 +00:00
Adding better register tracking through locals and fixing branches. LLVM does an amazing job of optimizing this. There are many opportunities to make it better, too, by preventing spilling where not required or only spilling/filling things when needed.
2013-01-25 08:32:42 +00:00
Benefits:
- Optimized prolog/epilog generation.
- Local variables for stack storage (alloca/etc) instead of user memory.
- Better return detection and fast returns.
### Indirect branches (ctr/lr)
Adding TODO.
2013-01-22 08:54:49 +00:00
Update TODO.md
2013-01-22 18:50:20 +00:00
Return path:
- In SDB see if the function follows the 'return' semantic:
- mfspr LR / mtspr LR/CTR / bcctr -- at end?
- In codegen add a BB that is just return.
- If a block 'returns', branch to the return BB.
Tail calls:
- If in a call BB check next BB.
- If next is a return, optimize to a tail call.
Adding TODO.
2013-01-22 08:54:49 +00:00
Fast path:
- Every time LR would be stashed, add the value of LR (some NIA) to a lookup
table.
- When doing an indirect branch first lookup the address in the table.
- If not found, slow path, else jump.
Slow path:
- Call out and do an SDB lookup.
- If found, return, add to lookup table, and jump.
- If not found, need new function codegen!
```
Misc output cleanup and a test run script.
2013-01-24 05:59:26 +00:00
## Debugging