[libraries-jitstress] System.Tests.UInt128Tests_GenericMath.TryReadLittleEndianUInt64Test fails in CI

[jakobbotsch]

Example pipeline run: https://dev.azure.com/dnceng-public/public/_build/results?buildId=734477&view=results
Example console log: https://helixre107v0xdcypoyl9e7f.blob.core.windows.net/dotnet-runtime-refs-pull-104593-merge-a1079025a96a477299/System.Runtime.Tests/1/console.6e9bbd2a.log?helixlogtype=result

I verified that the problem reproduces with the latest main. It looks caused by #104467, though it must have interfered with another PR since libraries-jitstress didn't hit this failure on that PR.

[dotnet-policy-service]

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

[jakobbotsch]

The codegen for System.UInt128:op_LeftShift relies on the fact that the retbuffer does not alias alias any incoming implicit byref:

; Assembly listing for method System.UInt128:op_LeftShift(System.UInt128,int):System.UInt128 (FullOpts)
; Emitting BLENDED_CODE for X64 with AVX - Windows
; FullOpts code
; optimized code
; optimized using Unknown PGO
; rsp based frame
; partially interruptible
; with Unknown PGO: fgCalledCount is 160
; No PGO data
; 0 inlinees with PGO data; 2 single block inlinees; 0 inlinees without PGO data
; Final local variable assignments
;
;  V00 RetBuf       [V00,T02] (  3,  3   )   byref  ->   r8         single-def
;  V01 arg0         [V01,T00] (  3,  6   )   byref  ->  rdx         single-def
;  V02 arg1         [V02,T03] (  3,  3   )     int  ->  rcx         single-def
;  V03 loc0         [V03,T08] (  2,  0.52)    long  ->  rax        
;  V04 loc1         [V04,T07] (  1,  1   )     int  ->  [rsp+0x2C]  do-not-enreg[V] "GSCookie dummy"
;  V05 OutArgs      [V05    ] (  1,  1   )  struct (32) [rsp+0x00]  do-not-enreg[XS] addr-exposed "OutgoingArgSpace"
;* V06 tmp2         [V06    ] (  0,  0   )  struct (16) zero-ref    do-not-enreg[SF] ld-addr-op "NewObj constructor temp" <System.UInt128>
;* V07 tmp3         [V07    ] (  0,  0   )  struct (16) zero-ref    do-not-enreg[SF] ld-addr-op "NewObj constructor temp" <System.UInt128>
;* V08 tmp4         [V08    ] (  0,  0   )    long  ->  zero-ref    "Inlining Arg"
;* V09 tmp5         [V09    ] (  0,  0   )    long  ->  zero-ref    "Inlining Arg"
;* V10 tmp6         [V10    ] (  0,  0   )    long  ->  zero-ref    "V06.[000..008)"
;* V11 tmp7         [V11    ] (  0,  0   )    long  ->  zero-ref    "V06.[008..016)"
;* V12 tmp8         [V12    ] (  0,  0   )    long  ->  zero-ref    "V07.[000..008)"
;* V13 tmp9         [V13    ] (  0,  0   )    long  ->  zero-ref    "V07.[008..016)"
;  V14 GsCookie     [V14    ] (  1,  1   )    long  ->  [rsp+0x30]  do-not-enreg[X] addr-exposed "GSSecurityCookie"
;  V15 tmp11        [V15,T04] (  9,  3.76)   byref  ->   r8         single-def "shadowVar"
;  V16 tmp12        [V16,T05] (  6,  2.52)   byref  ->  rdx         single-def "shadowVar"
;  V17 tmp13        [V17,T01] (  8,  5.52)     int  ->  rcx         "shadowVar"
;  V18 cse0         [V18,T06] (  5,  1.78)     int  ->  r10         multi-def "CSE #01"
;
; Lcl frame size = 56

G_M38089_IG01:
       sub      rsp, 56
       mov      rax, 0xD1FFAB1E
       mov      qword ptr [rsp+0x30], rax
       mov      eax, r8d
       mov      r8, rcx
       mov      ecx, eax
						;; size=27 bbWeight=1 PerfScore 2.25
G_M38089_IG02:
       and      ecx, 127
       test     cl, 64
       je       SHORT G_M38089_IG06
						;; size=8 bbWeight=1 PerfScore 1.50
G_M38089_IG03:
       xor      eax, eax
       mov      qword ptr [r8], rax ; zeroing ret buffer
       mov      r10d, ecx
       and      r10d, 63
       mov      rax, qword ptr [rdx] ; accessing parameter that may have aliased the ret buffer
       shlx     rax, rax, r10
       mov      qword ptr [r8+0x08], rax
       mov      rax, r8
       mov      rcx, 0xD1FFAB1E
       cmp      qword ptr [rsp+0x30], rcx
       je       SHORT G_M38089_IG04
       call     CORINFO_HELP_FAIL_FAST
						;; size=49 bbWeight=0.50 PerfScore 4.88
G_M38089_IG04:
       nop      
						;; size=1 bbWeight=0.50 PerfScore 0.12
G_M38089_IG05:
       add      rsp, 56
       ret      
						;; size=5 bbWeight=0.50 PerfScore 0.62
G_M38089_IG06:
       test     ecx, ecx
       je       SHORT G_M38089_IG10
						;; size=4 bbWeight=0.50 PerfScore 0.62
G_M38089_IG07:
       mov      r10d, ecx
       and      r10d, 63
       mov      rax, qword ptr [rdx]
       shlx     rax, rax, r10
       mov      qword ptr [r8], rax
       mov      rax, qword ptr [rdx+0x08]
       shlx     rax, rax, r10
       neg      ecx
       add      ecx, 64
       mov      rdx, qword ptr [rdx]
       shrx     rcx, rdx, rcx
       or       rax, rcx
       mov      qword ptr [r8+0x08], rax
       mov      rax, r8
       mov      rcx, 0xD1FFAB1E
       cmp      qword ptr [rsp+0x30], rcx
       je       SHORT G_M38089_IG08
       call     CORINFO_HELP_FAIL_FAST
						;; size=72 bbWeight=0.26 PerfScore 3.96
G_M38089_IG08:
       nop      
						;; size=1 bbWeight=0.26 PerfScore 0.07
G_M38089_IG09:
       add      rsp, 56
       ret      
						;; size=5 bbWeight=0.26 PerfScore 0.33
G_M38089_IG10:
       vmovups  xmm0, xmmword ptr [rdx]
       vmovups  xmmword ptr [r8], xmm0
       mov      rax, r8
       mov      rcx, 0xD1FFAB1E
       cmp      qword ptr [rsp+0x30], rcx
       je       SHORT G_M38089_IG11
       call     CORINFO_HELP_FAIL_FAST
						;; size=34 bbWeight=0.24 PerfScore 2.52
G_M38089_IG11:
       nop      
						;; size=1 bbWeight=0.24 PerfScore 0.06
G_M38089_IG12:
       add      rsp, 56
       ret      
						;; size=5 bbWeight=0.24 PerfScore 0.30

; Total bytes of code 212, prolog size 19, PerfScore 17.23, instruction count 57, allocated bytes for code 212 (MethodHash=095f6b36) for method System.UInt128:op_LeftShift(System.UInt128,int):System.UInt128 (FullOpts)
; ============================================================

We either need to disable implicit byref copy elision at the caller, or make sure that we do not reorder this zeroing of the return buffer with parameters. (We already have to take care we do not do this, since the ret buffer can arbitrarily alias byrefs.)

[jakobbotsch]

I think it's more reasonable to go back to the old invariant (implicit byrefs do not alias anything), so I'll do the analysis to disable last-use copy elision if the argument overlaps the return buffer.