Repository navigation
[JIT][arm64] CSE for VSD calls #60025
Description
void Test(IDisposable d) => d.Dispose();Currently emits:
; Method Program:Test(System.IDisposable):this
G_M50981_IG01: ;; offset=0000H
A9BF7BFD stp fp, lr, [sp,#-16]!
910003FD mov fp, sp
;; bbWeight=1 PerfScore 1.50
G_M50981_IG02: ;; offset=0008H
AA0103E0 mov x0, x1
D280080B movz x11, #64
F2B9BB8B movk x11, #0xcddc LSL #16
F2CFFFAB movk x11, #0x7ffd LSL #32
D2800801 movz x1, #64
F2B9BB81 movk x1, #0xcddc LSL #16
F2CFFFA1 movk x1, #0x7ffd LSL #32
F9400021 ldr x1, [x1]
;; bbWeight=1 PerfScore 6.50
G_M50981_IG03: ;; offset=0028H
A8C17BFD ldp fp, lr, [sp],#16
D61F0020 br x1
;; bbWeight=1 PerfScore 2.00
; Total bytes of code: 48Expected codegen:
; Method Program:Test(System.IDisposable):this
G_M50981_IG01: ;; offset=0000H
A9BF7BFD stp fp, lr, [sp,#-16]!
910003FD mov fp, sp
;; bbWeight=1 PerfScore 1.50
G_M50981_IG02: ;; offset=0008H
AA0103E0 mov x0, x1
D2800801 movz x1, #64
F2B9BB81 movk x1, #0xcddc LSL #16
F2CFFFA1 movk x1, #0x7ffd LSL #32
F9400021 ldr x1, [x1]
;; bbWeight=1 PerfScore 6.50
G_M50981_IG03: ;; offset=0028H
A8C17BFD ldp fp, lr, [sp],#16
D61F0020 br x1
;; bbWeight=1 PerfScore 2.00/cc @dotnet/jit-contrib
Metadata
Metadata
Assignees
Labels
Type
Projects
- Status
Activity
AndyAyersMS commentedon Oct 5, 2021
You mean instead of those composed immediates? I don't think there's much we can do on that front.
We should however do a better job of CSEing these, or (if this happens in say lower) be smarter about not gratuitously duplicating handle nodes.
EgorBo commentedon Oct 5, 2021
good point, but looks like the second constant appears in lowering after CSE (#47808 related?)
AndyAyersMS commentedon Oct 5, 2021
Similar -- #47808 was for virtual calls. This is an interface call (VSD).
We could try to expose this earlier too, and give CSE a chance to work, but it might also be fairly easy to fix in lowering itself as we have to load the thing in a register anyways.
EgorBo commentedon Oct 5, 2021
@AndyAyersMS regarding the relocs - I was thinking of a dedicated reg always pointing to the start of a separated heap for Type objects (and string literals) so we could always use relative addressing for those - but it's just thinking out loud
[-][JIT][arm64] relocs for isinst and VSD[/-][+][JIT][arm64] CSE for VSD calls[/+]AndyAyersMS commentedon Oct 5, 2021
RelatedShared constant CSEs should get you something similar, right? Just that the "base" address would be selected based on the actual constants that appear in the IR.Right now we allow a limited form of this, perhaps we should expand it?
jakobbotsch commentedon Oct 5, 2021
We already do the optimization but it's turned off for tailcalls. I have #59602 that fixes it.
jakobbotsch commentedon Oct 8, 2021
This was actually fixed with #56669.