Jit should avoid allocations for T1 is T2 pattern in generics

[EgorBo]

Minimal repro from #58535

using System.Runtime.CompilerServices;

class Program
{
 [MethodImpl(MethodImplOptions.NoInlining)]
 static bool Test<T1, T2>(T1 source) where T1 : struct where T2 : struct 
 {
 return source is T2; // box, issinst, ldnull
 }

 static void Main()
 {
 Test<(int, string), (int, string)>((42, "42"));
 }
}

Test is compiled as:

G_M19373_IG01:
 push rdi
 push rsi
 push rbp
 push rbx
 sub rsp, 40
 mov qword ptr [rsp+20H], rcx
 mov rbx, rcx
 mov rsi, rdx
						;; bbWeight=1 PerfScore 5.75
G_M19373_IG02:
 mov rcx, qword ptr [rbx+56]
 mov rcx, qword ptr [rcx]
 call CORINFO_HELP_NEWSFAST
 mov rbp, rax
 lea rdi, bword ptr [rbp+8]
 call CORINFO_HELP_ASSIGN_BYREF
 movsq 
 mov rax, qword ptr [rbp]
 mov rdx, qword ptr [rbx+56]
 cmp rax, qword ptr [rdx+8]
 je SHORT G_M19373_IG04
						;; bbWeight=1 PerfScore 14.75
G_M19373_IG03:
 xor rbp, rbp
						;; bbWeight=0.12 PerfScore 0.03
G_M19373_IG04:
 test rbp, rbp
 setne al
 movzx rax, al
						;; bbWeight=1 PerfScore 1.50
G_M19373_IG05:
 add rsp, 40
 pop rbx
 pop rbp
 pop rsi
 pop rdi
 ret 
						;; bbWeight=1 PerfScore 3.25

while we could avoid the boxing here and compare generic contexts?

category:performance
theme:generics
skill-level:intermediate
cost:small
impact:small

Tagging subscribers to this area: @JulieLeeMSFT
See info in area-owners.md if you want to be subscribed.

Issue Details

---

Minimal repro from #58535

using System.Runtime.CompilerServices;

class Program
{
 [MethodImpl(MethodImplOptions.NoInlining)]
 static bool Test<T1, T2>(T1 source) => source is T2;

 static void Main()
 {
 Test<(int, string), (int, string)>((42, "42"));
 }
}

Test is compiled as:

G_M19373_IG01:
 push rdi
 push rsi
 push rbp
 push rbx
 sub rsp, 40
 mov qword ptr [rsp+20H], rcx
 mov rbx, rcx
 mov rsi, rdx
						;; bbWeight=1 PerfScore 5.75
G_M19373_IG02:
 mov rcx, qword ptr [rbx+56]
 mov rcx, qword ptr [rcx]
 call CORINFO_HELP_NEWSFAST
 mov rbp, rax
 lea rdi, bword ptr [rbp+8]
 call CORINFO_HELP_ASSIGN_BYREF
 movsq 
 mov rax, qword ptr [rbp]
 mov rdx, qword ptr [rbx+56]
 cmp rax, qword ptr [rdx+8]
 je SHORT G_M19373_IG04
						;; bbWeight=1 PerfScore 14.75
G_M19373_IG03:
 xor rbp, rbp
						;; bbWeight=0.12 PerfScore 0.03
G_M19373_IG04:
 test rbp, rbp
 setne al
 movzx rax, al
						;; bbWeight=1 PerfScore 1.50
G_M19373_IG05:
 add rsp, 40
 pop rbx
 pop rbp
 pop rsi
 pop rdi
 ret 
						;; bbWeight=1 PerfScore 3.25

while we could avoid the boxing here and compare generic contexts?

Author:	EgorBo
Assignees:	-
Labels:	tenet-performance, area-CodeGen-coreclr, untriaged
Milestone:	-

[AndyAyersMS]

Seems like another box pattern match case?

[acaly]

Shouldn't this be better written as default(T1) is T2? JIT can optimize that properly.

I think unless we need the dynamic checking, we should prefer checking with default(T) instead of something with type T (in this case the parameter).

[AndyAyersMS]

Fixed by #103361