[mono] Unsafe.[Write/Read]Unaligned doesn't unroll 64B blocks

[matouskozak]

In dotnet/perf-autofiling-issues#33182, we discovered that MonoJIT generates two memcpy calls for

[StructLayout(LayoutKind.Sequential, Size = 64)]
public struct Block64 {}

Unsafe.WriteUnaligned(ref dest, Unsafe.ReadUnaligned<Block64>(ref src));

1  il_seq_point intr il: 0x0
2  il_seq_point il: 0x1
3  load_membase R37 <- [fp + 0x18]
4  add_imm R38 <- fp [32]
5  move R40 <- R38
6  move R41 <- R37
7  iconst R42 <- [64]
8  voidcall [void string:memcpy (byte*,byte*,int)] [r0 <- R40] [r1 <- R41] [r2 <- R42] clobbers: c
9  il_seq_point il: 0x8, nonempty-stack
10 add_imm R43 <- fp [32]
11 load_membase R44 <- [fp + 0x10]
12 move R46 <- R44
13 move R47 <- R43
14 iconst R48 <- [64]
15 voidcall [void string:memcpy (byte*,byte*,int)] [r0 <- R46] [r1 <- R47] [r2 <- R48] clobbers: c
16 il_seq_point il: 0xd, nonempty-stack
17 il_seq_point il: 0xd
18 il_seq_point il: 0xe

For this scenario, MonoJIT takes the "safe" path in mini_emit_memcpy_internal (passes size / align > MAX_INLINE_COPIES) instead of using mini_emit_memcpy that handles with copy unrolling.

In comparison Unsafe.As<byte, Block64>(ref dest) = Unsafe.As<byte, Block64>(ref src); leads to:

1  il_seq_point intr il: 0x0
2  il_seq_point il: 0x1
3  il_seq_point il: 0x7, nonempty-stack
4  il_seq_point il: 0xd, nonempty-stack
5  load_membase R39 <- [fp + 0x10]
6  nop
7  load_membase R40 <- [fp + 0x18]
8  nop
9  iconst R41 <- [64]
10 voidcall [void string:memcpy (byte*,byte*,int)] [r0 <- R39] [r1 <- R40] [r2 <- R41] clobbers: c
11 il_seq_point il: 0x17

This is causing serious regression on MonoJIT dotnet/perf-autofiling-issues#33182 and more. Fixing this would bring over 400+ microbenchmark improvements (dotnet/perf-autofiling-issues#41406 (comment))

[dotnet-policy-service]

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