JIT: perf regression from improved local struct copy prop

[AndyAyersMS]

See #74384 (comment)

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

Issue Details

---

See #74384 (comment)

Author:	AndyAyersMS
Assignees:	-
Labels:	area-CodeGen-coreclr, untriaged
Milestone:	-

[AndyAyersMS]

We now run asserion gen "early" for some block expansions, so we miss out on some assertions for some kinds of expansions.

In particular in the baseline jit we did:

fgMorphTree BB02, STMT00005 (before)
               [000026] IA---------                         *  ASG       struct (init)
               [000025] n------N---                         +--*  BLK       struct<8>
               [000023] -----------                         |  \--*  ADDR      byref 
               [000022] -------N---                         |     \--*  LCL_VAR   ref    V03 loc2         
               [000024] -----------                         \--*  CNS_INT   int    0
MorphInitBlock:
MorphBlock for dst tree, before:
               [000025] n----+-N---                         *  BLK       struct<8>
               [000023] -----+-----                         \--*  ADDR      byref 
               [000022] D----+-N---                            \--*  LCL_VAR   ref    V03 loc2         
MorphBlock after:
               [000025] n----+-N---                         *  BLK       struct<8>
               [000023] -----+-----                         \--*  ADDR      byref 
               [000022] D----+-N---                            \--*  LCL_VAR   ref    V03 loc2         
PrepareDst for [000025] have found a local var V03.
fgMorphOneAsgBlock (after):
               [000026] -A---------                         *  ASG       ref   
               [000022] D----+-N---                         +--*  LCL_VAR   ref    V03 loc2         
               [000024] -----+-----                         \--*  CNS_INT   ref    null
MorphInitBlock (after):
               [000026] -A---------                         *  ASG       ref   
               [000022] D----+-N---                         +--*  LCL_VAR   ref    V03 loc2         
               [000024] -----+-----                         \--*  CNS_INT   ref    null
GenTreeNode creates assertion:
               [000026] -A---------                         *  ASG       ref   
In BB02 New Local Constant Assertion: V03 == null, index = #01

fgMorphTree BB02, STMT00005 (after)
               [000026] -A---+-----                         *  ASG       ref   
               [000022] D----+-N---                         +--*  LCL_VAR   ref    V03 loc2         
               [000024] -----+-----                         \--*  CNS_INT   ref    null

fgMorphTree BB02, STMT00006 (before)
               [000029] --CXG------                         *  CALL help void   HELPER.CORINFO_HELP_ARRADDR_ST
               [000020] ----------- arg0                    +--*  LCL_VAR   ref    V01 loc0         
               [000028] ----------- arg1                    +--*  CAST      long <- int
               [000021] -----------                         |  \--*  LCL_VAR   int    V02 loc1         
               [000027] ----------- arg2                    \--*  LCL_VAR   ref    V03 loc2  

Initializing arg info for 29.CALL:
Args for call [000029] CALL after AddFinalArgsAndDetermineABIInfo:
CallArg[[000020].LCL_VAR ref (By value), 1 reg: rcx, byteAlignment=8]
CallArg[[000028].CAST long (By value), 1 reg: rdx, byteAlignment=8]
CallArg[[000027].LCL_VAR ref (By value), 1 reg: r8, byteAlignment=8]

Morphing args for 29.CALL:

Assertion prop in BB02:
Constant Assertion: V03 == null, index = #01
               [000027] -----------                         *  CNS_INT   ref    null

Sorting the arguments:
Deferred argument ('rdx'):
               [000028] -----+-----                         *  CAST      long <- int
               [000021] -----+-----                         \--*  LCL_VAR   int    V02 loc1         
Moved to late list
Deferred argument ('rcx'):
               [000020] -----+-----                         *  LCL_VAR   ref    V01 loc0         
Moved to late list
Deferred argument ('r8'):
               [000027] -----+-----                         *  CNS_INT   ref    null
Moved to late list

Register placement order:    rdx rcx r8 
Args for [000029].CALL after fgMorphArgs:
CallArg[[000020].LCL_VAR ref (By value), 1 reg: rcx, byteAlignment=8, isLate, processed]
CallArg[[000028].CAST long (By value), 1 reg: rdx, byteAlignment=8, isLate, processed]
CallArg[[000027].CNS_INT ref (By value), 1 reg: r8, byteAlignment=8, isLate, processed]
OutgoingArgsStackSize is 32

fgMorphIndexAddr (before remorph):
               [000049] ---X-------                         *  COMMA     byref 
               [000042] ---X-------                         +--*  BOUNDS_CHECK_Rng void  

In the current jit we don't get this assertion. This optimization happens later during assertion prop, but by then it is too late to hoist invariants. So, we end up with the following inner loop codegen where a load is not hoisted.

;; before
IN0007: 000013 align    [13 bytes for IG03]
G_M31925_IG03:        ; offs=000020H, size=0014H, bbWeight=4    PerfScore 16.00, gcrefRegs=00000001 {rax}, byrefRegs=00000000 {}, loop=IG03, BB02 [0001], byref, isz

IN0008: 000020 movsxd   r8, ecx
IN0009: 000023 cmp      r8, rdx
IN000a: 000026 jae      SHORT G_M31925_IG05
IN000b: 000028 xor      r9, r9
IN000c: 00002B mov      gword ptr [rax+8*r8+10H], r9
IN000d: 000030 dec      ecx
IN000e: 000032 jns      SHORT G_M31925_IG03

;; after
G_M46622_IG03:        ; offs=00000FH, size=001BH, bbWeight=4    PerfScore 25.00, gcrefRegs=00000001 {rax}, byrefRegs=00000000 {}, BB02 [0001], byref, isz

IN0005: 00000F movsxd   rcx, edx
IN0006: 000012 mov      r8d, dword ptr [rax+08H]
IN0007: 000016 cmp      rcx, r8
IN0008: 000019 jae      SHORT G_M46622_IG05
IN0009: 00001B movsxd   rcx, edx
IN000a: 00001E xor      r8, r8
IN000b: 000021 mov      gword ptr [rax+8*rcx+10H], r8
IN000c: 000026 dec      edx
IN000d: 000028 jns      SHORT G_M46622_IG03

Seems like we perhaps should re-run assertion gen if we do field-by field or "one asg block" expansion.

Also seems odd that we don't try and detect this optimization in preorder -- we could avoid going through call morphing just to undo the call later on in morph. Anyways...

[AndyAyersMS]