JIT: Poor codegen when calling an instance method on a temporary, unused struct

[Thealexbarney]

Description

Poor codegen in the following situation:

1. Initialize a new struct
2. Call an instance method on the struct that doesn't depend on any struct fields
3. Don't use the struct again

Configuration

Running a recent master build on Win 10 2004 x64

Other information

Given this code, the JIT produces this disassembly for Method1 and Method2:

public static int Method1() => new Rng().NextRandom();
public static int Method2() => CreateRng().NextRandom();

public static Rng CreateRng() => new Rng();

public struct Rng
{
    public int NextRandom() => 4;
}

If you do it the normal way (new Rng().NextRandom()) the JIT doesn't completely optimize it.
It insists on initializing the struct. If the struct is a larger size the JIT would zero the entire struct on the stack before returning 4.

; Assembly listing for method Program:Method1():int
; Emitting BLENDED_CODE for X64 CPU with AVX - Windows
; optimized code
; rsp based frame
; partially interruptible
; Final local variable assignments
;
;  V00 loc0         [V00    ] (  1,  1   )  struct ( 8) [rsp+0x00]   do-not-enreg[XS] must-init addr-exposed ld-addr-op
;# V01 OutArgs      [V01    ] (  1,  1   )  lclBlk ( 0) [rsp+0x00]   "OutgoingArgSpace"
;  V02 tmp1         [V02,T00] (  2,  4   )   byref  ->  rax         "dup spill"
;
; Lcl frame size = 8

G_M20050_IG01:
       50                   push     rax
       33C0                 xor      rax, rax
       48890424             mov      qword ptr [rsp], rax
						;; bbWeight=1    PerfScore 2.25
G_M20050_IG02:
       488D0424             lea      rax, bword ptr [rsp]
       33D2                 xor      edx, edx
       8810                 mov      byte  ptr [rax], dl
       B804000000           mov      eax, 4
						;; bbWeight=1    PerfScore 2.00
G_M20050_IG03:
       4883C408             add      rsp, 8
       C3                   ret      
						;; bbWeight=1    PerfScore 1.25

; Total bytes of code 25, prolog size 7, PerfScore 8.00, (MethodHash=ba7bb1ad) for method Program:Method1():int
; ============================================================

If you split new Rng() into its own function, the JIT does completely optimize it.

; Assembly listing for method Program:Method2():int
; Emitting BLENDED_CODE for X64 CPU with AVX - Windows
; optimized code
; rsp based frame
; partially interruptible
; Final local variable assignments
;
;* V00 loc0         [V00    ] (  0,  0   )  struct ( 8) zero-ref    do-not-enreg[SB] ld-addr-op
;# V01 OutArgs      [V01    ] (  1,  1   )  lclBlk ( 0) [rsp+0x00]   "OutgoingArgSpace"
;* V02 tmp1         [V02    ] (  0,  0   )  struct ( 8) zero-ref    do-not-enreg[SB] ld-addr-op "Inline ldloca(s) first use temp"
;
; Lcl frame size = 0

G_M12881_IG01:
						;; bbWeight=1    PerfScore 0.00
G_M12881_IG02:
       B804000000           mov      eax, 4
						;; bbWeight=1    PerfScore 0.25
G_M12881_IG03:
       C3                   ret      
						;; bbWeight=1    PerfScore 1.00

; Total bytes of code 6, prolog size 0, PerfScore 1.85, (MethodHash=7270cdae) for method Program:Method2():int
; ============================================================

category:cq
theme:structs
skill-level:intermediate
cost:medium

[EgorBo]

IR after Morph:

Method1

***** BB01
STMT00000 (IL 0x000...0x004)
               [000003] -A---+------              *  ASG       byref 
               [000002] D----+-N----              +--*  LCL_VAR   byref  V02 tmp1         
               [000001] -----+------              \--*  LCL_VAR_ADDR byref  V00 loc0         

***** BB01
STMT00001 (IL   ???...  ???)
               [000008] IA-X-+------              *  ASG       struct (init)
               [000007] ---X-+-N----              +--*  BLK       struct<1>
               [000005] -----+------              |  \--*  LCL_VAR   byref  V02 tmp1         
               [000006] -----+------              \--*  CNS_INT   int    0

***** BB01
STMT00003 (IL   ???...  ???)
               [000011] -----+------              *  RETURN    int   
               [000012] -----+------              \--*  CNS_INT   int    4

Method2

***** BB01
STMT00004 (IL 0x000...  ???)
               [000013] IA---+------              *  ASG       struct (init)
               [000010] D----+-N----              +--*  LCL_VAR   struct<Rng, 1> V02 tmp1         
               [000012] -----+------              \--*  CNS_INT   int    0

***** BB01
STMT00001 (IL   ???...  ???)
               [000004] IA---+------              *  ASG       struct (init)
               [000002] D----+-N----              +--*  LCL_VAR   struct<Rng, 1> V00 loc0         
               [000014] -----+------              \--*  CNS_INT   int    0

***** BB01
STMT00003 (IL   ???...  ???)
               [000009] -----+------              *  RETURN    int   
               [000016] -----+------              \--*  CNS_INT   int    4

[sandreenko]

This issue is interesting, there are 2 solutions, I think both have their own benefits and should be implemented:

1. Don't spill addr of local because they are trivial to clone,
   change
   

   runtime/src/coreclr/jit/importer.cpp

   Line 13719 in 355eff5

   if (!opts.compDbgCode && !op1->IsIntegralConst(0) && !op1->IsFPZero() && !op1->IsLocal())

   
   to

if (!opts.compDbgCode && !op1->IsIntegralConst(0) && !op1->IsFPZero() && !op1->IsLocal() && !op1->OperIsLocalAddr())

however, it won't work because importer sees it as "ADDR(LCL_VAR())` so we need a part of #11057.

2. continue the work to treat LCL_VAR_ADDR as consts started in CQ: Don't force const arguments to temps. #47786 and add O2K_LCL_VAR_ADDR, O2K_LCL_FLD_ADDR to optOp2Kind so VN created assertions like:
   V02 = &V01
   and uses it to see that some stores are dead.

Not sure if we have time for it in 6.0 but lets keep it here for now.

[JulieLeeMSFT]

Moving to .NET 8.