Singlepass aarch64 #2750
Singlepass aarch64 #2750
Conversation
|
Not ready to merge. Mac support is not here yet. |
| .iter() | ||
| .cloned(), |
There was a problem hiding this comment.
Isn't it better to do into_iter() ?
There was a problem hiding this comment.
Sure, I can use into_iter() instead.
There was a problem hiding this comment.
But I'll still need a .cloned() so I think it will not change much in the end.
ptitSeb
force-pushed
the
singlepass_aarch64
branch
from
3f0f2ee to
2d218fb
Compare
|
Ready to merge now |
…ods (now 30 tests pass)
…09 tests passes now)
…tion yet (120 tests passes now)
ptitSeb
force-pushed
the
singlepass_aarch64
branch
from
2d218fb to
7f4fd02
Compare
There was a problem hiding this comment.
I can see a lot of unimplemented!() in machine_arm64.rs. Generally unimplemented! means the same thing as todo!: "this code is not implemented yet but needs to be implemented".
Use unreachable! instead to indicate "I am sure that this code is unreachable". They both panic, but it makes it clearer whether there is still work to be done.
| NEON::V31, | ||
| ]; | ||
| match n { | ||
| 0..=15 => Ok(REGS[n]), |
There was a problem hiding this comment.
Typo? Should this be 31? If not then add a comment explaining why.
| match n { | ||
| 0..=31 => Ok(REGS[n]), | ||
| _ => Err(()), | ||
| } |
There was a problem hiding this comment.
| match n { | |
| 0..=31 => Ok(REGS[n]), | |
| _ => Err(()), | |
| } | |
| REGS.get(n).ok_or(()) |
get returns an Option, ok_or converts it to a Result.
But perhaps this function should be changed to use plain indexing and simply panic if an out-of-range value is passed in since you're always unwraping the result anyways.
There was a problem hiding this comment.
it's not happy "expected enum arm64_decl::GPR, found &arm64_decl::GPR"
There was a problem hiding this comment.
You can used .cloned() to turn an Option<&T> to an Option<T>.
|
|
||
| type Assembler = VecAssembler<Aarch64Relocation>; | ||
|
|
||
| /// Force `dynasm!` to use the correct arch (x64) when cross-compiling. |
There was a problem hiding this comment.
| /// Force `dynasm!` to use the correct arch (x64) when cross-compiling. | |
| /// Force `dynasm!` to use the correct arch (aarch64) when cross-compiling. |
| if (disp & 0x7) != 0 || (disp >= 0x8000) { | ||
| unreachable!(); | ||
| } |
There was a problem hiding this comment.
Nit: this should use assert! instead.
|
|
||
| fn emit_mov_imm(&mut self, dst: Location, val: u64) { | ||
| match dst { | ||
| Location::GPR(dst) => { |
There was a problem hiding this comment.
Try to encode the immediate as a logical immediate so it can be instantiated with ORR dst, XZR, #imm.
There was a problem hiding this comment.
Yes. also can use MOVN if !val is small enough.
But all this are optimisation that can come later.
There was a problem hiding this comment.
Can we add a todo comment with that suggestion so we log it somewhere?
There was a problem hiding this comment.
The are a few relatively easy optimisation here and there, not just this one (like using LDP/STP more), maybe a notion page?
There was a problem hiding this comment.
Sounds good, let's create a notion doc!
| ), | ||
| } | ||
| } | ||
| fn emit_add2(&mut self, sz: Size, src: Location, dst: Location) { |
There was a problem hiding this comment.
Can this just be a wrapper over emit_add?
| offset: reloc_at as u32, | ||
| addend: 0, | ||
| }); | ||
| self.assembler.emit_movk(Location::GPR(GPR::X27), 0, 0); |
There was a problem hiding this comment.
The first one should be movz.
|
|
||
| pub struct MachineARM64 { | ||
| assembler: Assembler, | ||
| used_gprs: HashSet<GPR>, |
There was a problem hiding this comment.
Nit: This should probably be a bitset instead of a hash table.
There was a problem hiding this comment.
Because of the small size of the set? I have use HashSet to be consistant with the x64, that used HashSet too.
I have use |
|
I see |
|
bors r+ |
Description
Add Aarch64 support on Singlapass, for both Linux and MacOS.