Sorry, something went wrong.
|
I had to revert the x86-64 changes as they are more involved and require more investigation. I'm going to create a new issue for changing x86-64 Linux target: #148256 |
Sorry, something went wrong.
|
Do you have any numbers for this? |
Sorry, something went wrong.
|
Do you have any numbers for this?
Before all the calls were done via GOT, now only 55% are done via the GOT and the rest using a direct bl |
Sorry, something went wrong.
|
Sorry, didn’t have a chance to look at this yet. What happens if the jump is outside the 26 bit range? The trampolines get used? Were the trampolines being used at all before? |
Sorry, something went wrong.
|
Sorry, didn’t have a chance to look at this yet. What happens if the jump is outside the 26 bit range? The trampolines get used? Were the trampolines being used at all before? Earlier no trampolines were used at all. Now if a jump is out of range, the trampoline is generated. |
Sorry, something went wrong.
Removing -fno-plt makes Linux stencils use more direct call relocations and fewer GOT-indirect call sequences. On AArch64 it means CALL26 instead of adrp/ldr/blr GOT calls
This improves code size and some complexity of the JIT stencils. It reduces the amount of GOT-specific patching the JIT has to do.