JIT: optimize some cases of unused structs

[AndyAyersMS]

In some cases CSC will use ldobj; pop to null check a pointer to
struct. This change avoids copying the struct value for such constructs.

Codegen may still redundantly null check, if there are multiple such checks
in a method.

Fixes #18710

[AndyAyersMS]

@briansull PTAL
cc @dotnet/jit-contrib

No diffs in fx for crossgen or pmi. Possibly because CSC will sometimes use ldfld instead which is less costly and easier to dead code.

Diffs on example code from #18710 (added here as tests):

Summary:
(Lower is better)
Total bytes of diff: -216 (-25.50% of base)
    diff is an improvement.
Top file improvements by size (bytes):
        -194 : structpop2.dasm (-38.42% of base)
         -22 : structpop.dasm (-6.43% of base)
2 total files with size differences (2 improved, 0 regressed), 0 unchanged.
Top method improvements by size (bytes):
        -194 : structpop2.dasm - P:TestByPtr(long):int
         -22 : structpop.dasm - P:TestByPtr(long):int
2 total methods with size differences (2 improved, 0 regressed), 11 unchanged.

For instance:

 ; Assembly listing for method P:TestByPtr(long):int
 ; Emitting BLENDED_CODE for X64 CPU with AVX
 ; optimized code
 ; rsp based frame
 ; partially interruptible
 ; Final local variable assignments
 ;
-;  V00 arg0         [V00,T00] (  7,  7   )    long  ->  rcx
+;  V00 arg0         [V00,T00] (  5,  5   )    long  ->  rcx
 ;  V01 loc0         [V01,T02] (  2,  2   )     int  ->  rax
 ;  V02 OutArgs      [V02    ] (  1,  1   )  lclBlk (32) [rsp+0x00]
 ;  V03 tmp1         [V03,T01] (  2,  4   )     int  ->  rdx
-;  V04 tmp2         [V04    ] (  3,  6   )  struct (16) [rsp+0x28]   do-not-enreg[XS] addr-exposed
-;  V05 tmp3         [V05    ] (  1,  1   )     int  ->  [rsp+0x28]   do-not-enreg[X] addr-exposed V04.F1(offs=0x00) P-DEP
-;  V06 tmp4         [V06    ] (  1,  1   )     int  ->  [rsp+0x2C]   do-not-enreg[X] addr-exposed V04.F2(offs=0x04) P-DEP
-;  V07 tmp5         [V07    ] (  1,  1   )     int  ->  [rsp+0x30]   do-not-enreg[X] addr-exposed V04.F3(offs=0x08) P-DEP
 ;
-; Lcl frame size = 56
+; Lcl frame size = 40
 G_M47241_IG01:
-       sub      rsp, 56
+       sub      rsp, 40
 G_M47241_IG02:
        mov      edx, dword ptr [rcx]
        mov      eax, dword ptr [rcx+4]
-       mov      r8d, dword ptr [rcx]
-       mov      dword ptr [rsp+28H], r8d
-       mov      r8d, dword ptr [rcx+4]
-       mov      dword ptr [rsp+2CH], r8d
-       mov      ecx, dword ptr [rcx+8]
-       mov      dword ptr [rsp+30H], ecx
+       cmp      dword ptr [rcx], ecx
        mov      ecx, edx
        mov      edx, eax
        call     P:Do(int,int):int
        nop
 G_M47241_IG03:
-       add      rsp, 56
+       add      rsp, 40
        ret
-; Total bytes of code 48, prolog size 4 for method P:TestByPtr(long):int
+; Total bytes of code 26, prolog size 4 for method P:TestByPtr(long):int

[mikedn]

Hmm, any volatile concerns?

[AndyAyersMS]

Hmm, maybe so. Thanks for bringing that up.

A quick glance suggests the jit won't even record volatile prefixes for ldobj. Need to dig in some more.

[mikedn]

Ah, this case is for structs. It's not clear to me if volatile applies to structs. ECMA-335 doesn't seem to say that volatile is not allowed on structs but I don't think I've ever seen code using that.

But then it looks like the importer already drops popped volatile reads - #6172

[AndyAyersMS]

Looks like CI is still not running normally.

[briansull]

Do you know what is up with "Windows_NT x64 Checked CoreFX Tests"

I am also seeing one test in that leg report a failure as well.

[CarolEidt]

Do you know what is up with "Windows_NT x64 Checked CoreFX Tests"

I opened #18840 for this yesterday, which has now been fixed with #18844.

[AndyAyersMS]

@briansull did you have any feedback?

[briansull]

Looks fine to me