Fold BOX+ISINST to null if possible #88989
Conversation
ghost
added
the
area-CodeGen-coreclr
|
Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch |
|
@MihuBot -tier0 |
src/coreclr/jit/importer.cpp
Outdated
| if ((opts != BoxPatterns::MakeInlineObservation) && | ||
| ((impStackTop().val->gtFlags & GTF_SIDE_EFFECT) == 0)) | ||
| { | ||
| if ((opts == BoxPatterns::IsByRefLike) || |
There was a problem hiding this comment.
Should we do this for byref-like types? Note that the exact patterns that are supported for them is documented here and I think it is expected that other patterns continue throwing (or, alternatively, that the pattern is documented).
A separate question -- doesn't this logic essentially subsume the logic that happens below?
There was a problem hiding this comment.
A separate question -- doesn't this logic essentially subsume the logic that happens below?
Sort of, in this case we only handle MustNot case while the logic below handles both Must and MustNot - I tried to share them but decided that it's not worth it, also, removed nullable<> support since it didn't add any diffs
Should we do this for byref-like types?
Let me remove it then and keep only the patterns listed there then
2 participants
Closes #88987
Tier0 Codegen diff for
Test<int>:; Assembly listing for method Program:Test[int](int) (Tier0) push rbp sub rsp, 48 lea rbp, [rsp+30H] xor eax, eax mov qword ptr [rbp-08H], rax - mov qword ptr [rbp-10H], rax mov dword ptr [rbp+10H], ecx - mov rcx, 0xD1FFAB1E ; System.Int32 - call CORINFO_HELP_NEWSFAST - mov gword ptr [rbp-10H], rax - mov rdx, gword ptr [rbp-10H] - mov ecx, dword ptr [rbp+10H] - mov dword ptr [rdx+08H], ecx - mov rdx, gword ptr [rbp-10H] - mov rcx, 0xD1FFAB1E ; System.Collections.IEnumerable - call [CORINFO_HELP_ISINSTANCEOFINTERFACE] - mov gword ptr [rbp-08H], rax + xor rcx, rcx + mov gword ptr [rbp-08H], rcx cmp gword ptr [rbp-08H], 0 je SHORT G_M5413_IG03 mov rcx, gword ptr [rbp-08H] call [Program:Consume(System.Object)] G_M5413_IG03: nop add rsp, 48 pop rbp ret -; Total bytes of code 100 +; Total bytes of code 49Consumecall is still here but it's underif (false)branch, we just can't properly fold that never-taken branch in Tier0 (needs forward sub in tier0) - not a big deal, at least, it no longer allocates.Some nice jit-diffs numbers locally.