Inlining FAILED: has ldstr VM restriction

[benaadams]

If one assembly has a static property as follows:

public static class HeaderNames
{
    public static string Date => "Date";
}

Using it in another assembly

if (ReferenceEquals(HeaderNames.Date, key))

Fails to inline the property at Jit time with

[FAILED: has ldstr VM restriction] HeaderNames:get_Date():ref

Is this by design? /cc @AndyAyersMS @jkotas

[benaadams]

Workaround is to lift the string load to a readonly static

public static class HeaderNames
{
    private readonly static string s_date = "Date";

    public static string Date => s_date;
}

[AndyAyersMS]

I don't know the specifics here, so will defer to Jan.

[benaadams]

Note, the behaviour I'm trying to capture is reference equality which direct string consts cross-assembly don't have (as they embed in the call site assembly); so using properties to achieve this.

[jkotas]

This check is there to ensure that the inlining won't change the string instance returned for constant string.

It looks like a fragile NGen optimization. I think this restriction can be relaxed without giving up much since we are not optimizing for fragile NGen anymore.

[benaadams]

My aim is to maintain reference equality so not propagate the const cross assembly at AoT (as string consts are not deduped cross assembly at assembly load). This is to achieve a fast path for string IgnoreCase comparisons for known strings (e.g. Header names dotnet/aspnetcore#9341)

With that goal in mind would it make sense to still block for crossgen/R2R; but allow at Jit/Tier1?

i.e. would a cross-assembly ldstr maintain reference equality at Jit time?

[jkotas]

C# sets CompilationRelaxations.NoStringInterning by default that gives runtime freedom in whether to de-dup cross-assembly.

I think it should be possible to adjust the system to de-dup more often or with less restrictions; but I am not sure whether we would want to start guaranteeing that everything is de-duped perfectly (ie ignore NoStringInterning).

[benaadams]

Maybe hiding a readonly static string behind a static property getter is the best approach then?

As when it inlines it will just inline the pointer; whereas if it was a string const it would also inline the load?

[benaadams]

e.g. with the readonly static + property I get

Inlines into 0600056A HttpRequestHeaders:TryGetValueFast(ref,byref):bool:this
  [1 IL=0143 TR=003198 060000E3] [below ALWAYS_INLINE size] HeaderNames:get_TE():ref
  [2 IL=0190 TR=003205 060000E3] [below ALWAYS_INLINE size] HeaderNames:get_TE():ref
  [0 IL=0197 TR=003210 060006FD] [FAILED: noinline per IL/cached result] String:Equals(ref,int):bool:this
  [3 IL=0246 TR=003070 060000ED] [below ALWAYS_INLINE size] HeaderNames:get_Via():ref
  [4 IL=0290 TR=003077 060000C5] [below ALWAYS_INLINE size] HeaderNames:get_DNT():ref
  [5 IL=0337 TR=003084 060000ED] [below ALWAYS_INLINE size] HeaderNames:get_Via():ref
  [0 IL=0344 TR=003089 060006FD] [FAILED: noinline per IL/cached result] String:Equals(ref,int):bool:this
  [6 IL=0387 TR=003096 060000C5] [below ALWAYS_INLINE size] HeaderNames:get_DNT():ref
  [0 IL=0394 TR=003101 060006FD] [FAILED: noinline per IL/cached result] String:Equals(ref,int):bool:this
  [7 IL=0443 TR=002875 060000CA] [below ALWAYS_INLINE size] HeaderNames:get_Host():ref
  [8 IL=0487 TR=002882 060000C4] [below ALWAYS_INLINE size] HeaderNames:get_Date():ref
  [9 IL=0527 TR=002889 060000C9] [below ALWAYS_INLINE size] HeaderNames:get_From():ref
  [10 IL=0571 TR=002896 060000CA] [below ALWAYS_INLINE size] HeaderNames:get_Host():ref
  [0 IL=0578 TR=002901 060006FD] [FAILED: noinline per IL/cached result] String:Equals(ref,int):bool:this
  [11 IL=0621 TR=002908 060000C4] [below ALWAYS_INLINE size] HeaderNames:get_Date():ref
  [0 IL=0628 TR=002913 060006FD] [FAILED: noinline per IL/cached result] String:Equals(ref,int):bool:this
  [12 IL=0667 TR=002920 060000C9] [below ALWAYS_INLINE size] HeaderNames:get_From():ref
  [0 IL=0674 TR=002925 060006FD] [FAILED: noinline per IL/cached result] String:Equals(ref,int):bool:this
  [13 IL=0720 TR=002747 060000B3] [below ALWAYS_INLINE size] HeaderNames:get_Allow():ref
  [14 IL=0764 TR=002754 060000DA] [below ALWAYS_INLINE size] HeaderNames:get_Range():ref
  [15 IL=0811 TR=002761 060000B3] [below ALWAYS_INLINE size] HeaderNames:get_Allow():ref
  [0 IL=0818 TR=002766 060006FD] [FAILED: noinline per IL/cached result] String:Equals(ref,int):bool:this
  [16 IL=0861 TR=002773 060000DA] [below ALWAYS_INLINE size] HeaderNames:get_Range():ref
  [0 IL=0868 TR=002778 060006FD] [FAILED: noinline per IL/cached result] String:Equals(ref,int):bool:this
  [17 IL=0917 TR=002424 060000A5] [below ALWAYS_INLINE size] HeaderNames:get_Accept():ref
  [18 IL=0961 TR=002431 060000D7] [below ALWAYS_INLINE size] HeaderNames:get_Pragma():ref
  [19 IL=1002 TR=002438 060000C3] [below ALWAYS_INLINE size] HeaderNames:get_Cookie():ref
  [20 IL=1046 TR=002445 060000C8] [below ALWAYS_INLINE size] HeaderNames:get_Expect():ref
  [21 IL=1090 TR=002452 060000D5] [below ALWAYS_INLINE size] HeaderNames:get_Origin():ref
  [22 IL=1137 TR=002459 060000A5] [below ALWAYS_INLINE size] HeaderNames:get_Accept():ref
  [0 IL=1144 TR=002464 060006FD] [FAILED: noinline per IL/cached result] String:Equals(ref,int):bool:this
  etc..

And the inlined string becomes a constant mov e.g.

if (ReferenceEquals(HeaderNames.Host, key))

becomes

G_M13073_IG06:
       mov      rdx, 0x15B8E12EBA8        ; HeaderNames:get_Host():ref
       cmp      gword ptr [rdx], rdi      ; ReferenceEquals
       jne      SHORT G_M13073_IG07