mirror of
https://github.com/FEX-Emu/FEX.git
synced 2026-10-09 14:00:20 +02:00
Alyssa Rosenzweig
af21b8f3c7
It is not an external component, and it makes paths needlessly long. Ryan seemed amenable to this when we discussed on IRC earlier. Signed-off-by: Alyssa Rosenzweig <alyssa@rosenzweig.io>
37 lines
2.8 KiB
Markdown
37 lines
2.8 KiB
Markdown
# FEXCore IR Optimization passes
|
|
---
|
|
**This is very much a WIP since these optimization passes aren't in code yet**
|
|
## Pass Managers
|
|
* Need Function level optimization pass manager
|
|
* Need block level optimization pass manager
|
|
### Dead Store Elimination
|
|
We need to do dead store elimination because LLVM can't always handle elimination of our loadstores
|
|
This is very apparent when we are doing flag calculations and LLVM isn't able to remove them
|
|
This is mainly just an issue around the context loadstores.
|
|
We will want this more when the IRJIT comes online.
|
|
### Dead flag elimination
|
|
X86-64 is a fairly disguting ISA in that it calculates a bunch of flags on almost all instructions.
|
|
We need eliminate redundant flag calculations that end up being being overwritten without being used.
|
|
This happens *constantly* and in most cases the flag calculation takes significantly more work than the basic op by itself
|
|
Good chance that breaking out the flags to independent memory locations will make this easier. Or just adding ops for flag handling.
|
|
### Dead Code Elimination
|
|
There are a lot of cases that code will be generated that is immediately dead afterwards.
|
|
Flag calculation elimination will produce a lot of dead code that needs to get removed.
|
|
Additionally there are a decent amount of x86-64 instructions that store their results in to multiple registers and then the next instruction overwrites one of those instructions.
|
|
Multiply and Divide being a big one, since x86 calculates these at higher precision.
|
|
These can rely significantly tracking liveness between LoadContext and StoreContext ops
|
|
### ABI register elimination pass
|
|
This one is very fun and will reduce a decent amount of work that the JIT needs to do.
|
|
When we are targeting a specific x86-64 ABI and we know that we have translated a block of code that is the entire function.
|
|
We can eliminate stores to the context that by ABI standards is a temporary register.
|
|
We will be able to know exactly that these are dead and just remove the store (and run all the passes that optimize the rest away afterwards).
|
|
### Loadstore coalescing pass
|
|
Large amount of x86-64 instructions load or store registers in order from the context.
|
|
We can merge these in to loadstore pair ops to improve perf
|
|
### Function level heuristic pass
|
|
Once we know that a function is a true full recompile we can do some additional optimizations.
|
|
Remove any final flag stores. We know that a compiler won't pass flags past a function call boundary(It doesn't exist in the ABI)
|
|
Remove any loadstores to the context mid function, only do a final store at the end of the function and do loads at the start. Which means ops just map registers directly throughout the entire function.
|
|
### SIMD coalescing pass?
|
|
When operating on older MMX ops(64bit SIMD) and they may end up up generating some independent ops that can be coalesced in to a 128bit op
|