Don't expand pinvokes in Tier0 #77545
Conversation
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch Issue DetailsGetId(false);
static int GetId(bool condition)
{
if (condition)
return GetCurrentThreadId();
return 0;
}
[DllImport("kernel32.dll")]
static extern int GetCurrentThreadId();Current Tier0 codegen for ; Assembly listing for method Program:GetId(bool):int
; Tier-0 compilation
; MinOpts code
G_M000_IG01: ;; offset=0000H
55 push rbp
4157 push r15
4156 push r14
4155 push r13
4154 push r12
57 push rdi
56 push rsi
53 push rbx
4883EC78 sub rsp, 120
488DAC24B0000000 lea rbp, [rsp+B0H]
894D10 mov dword ptr [rbp+10H], ecx
G_M000_IG02: ;; offset=001BH
488D8D78FFFFFF lea rcx, [rbp-88H]
498BD2 mov rdx, r10
E81628A55F call CORINFO_HELP_INIT_PINVOKE_FRAME
488945B8 mov qword ptr [rbp-48H], rax
488BC4 mov rax, rsp
48894598 mov qword ptr [rbp-68H], rax
488BC5 mov rax, rbp
488945A8 mov qword ptr [rbp-58H], rax
8B4510 mov eax, dword ptr [rbp+10H]
0FB6C0 movzx rax, al
85C0 test eax, eax
745E je SHORT G_M000_IG06
48B8F0ECDC5BF87F0000 mov rax, 0x7FF85BDCECF0
48894588 mov qword ptr [rbp-78H], rax
488D0521000000 lea rax, G_M000_IG04
488945A0 mov qword ptr [rbp-60H], rax
488B45B8 mov rax, qword ptr [rbp-48H]
488D9578FFFFFF lea rdx, bword ptr [rbp-88H]
48895010 mov qword ptr [rax+10H], rdx
488B45B8 mov rax, qword ptr [rbp-48H]
C6400C00 mov byte ptr [rax+0CH], 0
G_M000_IG03: ;; offset=0076H
FF15E4E50B00 call [Program:GetCurrentThreadId():int]
G_M000_IG04: ;; offset=007CH
488B55B8 mov rdx, qword ptr [rbp-48H]
C6420C01 mov byte ptr [rdx+0CH], 1
833D4903E35F00 cmp dword ptr [(reloc 0x7ff8bbb40ab4)], 0
7406 je SHORT G_M000_IG05
FF15052CE25F call [CORINFO_HELP_STOP_FOR_GC]
G_M000_IG05: ;; offset=0093H
488B55B8 mov rdx, qword ptr [rbp-48H]
488B4D80 mov rcx, bword ptr [rbp-80H]
48894A10 mov qword ptr [rdx+10H], rcx
8945C4 mov dword ptr [rbp-3CH], eax
EB05 jmp SHORT G_M000_IG07
G_M000_IG06: ;; offset=00A4H
33C0 xor eax, eax
8945C4 mov dword ptr [rbp-3CH], eax
G_M000_IG07: ;; offset=00A9H
8B45C4 mov eax, dword ptr [rbp-3CH]
G_M000_IG08: ;; offset=00ACH
4883C478 add rsp, 120
5B pop rbx
5E pop rsi
5F pop rdi
415C pop r12
415D pop r13
415E pop r14
415F pop r15
5D pop rbp
C3 ret
; Total bytes of code 189New Tier0 codegen for ; Assembly listing for method Program:GetId(bool):int
; Tier-0 compilation
; MinOpts code
G_M31913_IG01: ;; offset=0000H
55 push rbp
4883EC20 sub rsp, 32
488D6C2420 lea rbp, [rsp+20H]
894D10 mov dword ptr [rbp+10H], ecx
G_M31913_IG02: ;; offset=000DH
8B4510 mov eax, dword ptr [rbp+10H]
0FB6C0 movzx rax, al
85C0 test eax, eax
740C je SHORT G_M31913_IG04
E8D47BE7FF call Program:GetCurrentThreadId():int
90 nop
G_M31913_IG03: ;; offset=001DH
4883C420 add rsp, 32
5D pop rbp
C3 ret
G_M31913_IG04: ;; offset=0023H
33C0 xor eax, eax
G_M31913_IG05: ;; offset=0025H
4883C420 add rsp, 32
5D pop rbp
C3 ret
; Total bytes of code 43-146 bytes of codegen for an used codepath. If it turns out to be used an Pros:
Cons:
Overall I think it's a reasonable tradeoff PTAL @dotnet/jit-contrib
|
|
@EgorBo This could have an impact on startup paths for users with many single use and blittable signatures. I assume when we hit tier 1, the inlining will occur, which is good. However, for a Release build blittable P/Invokes have been inlined for a long time, changing that is something to be leery of. Can you check if there is a behavioral change when the export of the library isn't found? |
|
Do you some numbers for the cost of pros and cons in some real apps? |
|
If we're going to measure things, I'd like to see 3 measurements
I'm concerned about this change, and am in particular concerned about the startup time impact on things like winforms, as
|
Looks like we don't throw DllNotFoundException out of JIT when we inline a pinvoke do we? I've just tested it on this sample and it runs without exceptions (with and without the PR): public class Program
{
public static void Main(string[] args)
{
if (args.Length == 42)
Test();
}
[DllImport("non-existing-lib.dll")]
static extern int Test();
}
Ok, I'm closing this for now and if I have data I'll reopen. It's something I found in #77465 and decided to push separately to raise a discussion. Will workaround in #77465 to keep the current behavior. |
ghost
locked as resolved and limited conversation to collaborators
5 participants
Current Tier0 codegen for
GetID:New Tier0 codegen for
GetID:-146 bytes of codegen for an used codepath. If it turns out to be used an
IL_STUB_PInvokewill be compiled on demand.Pros:
IL_STUB_Pinvokecan already be compiled by someone else so we won't have to jit it again if it's usedCons:
Overall I think it's a reasonable tradeoff
PTAL @dotnet/jit-contrib