JIT emits unnecessary movsxd instructions when calling into Span indexer

[GrabYourPitchforks]

When passing a non-constant value into the Span<T> and ReadOnlySpan<T> indexer, the JIT will emit an unnecessary movsxd instruction on x64. The repro is fairly simple:

for (int i = 0; i < ints.Length; i++)
{
    retVal += ints[i];
}

Current codegen:

00007ffd`2a2e7291 85c9            test    ecx,ecx
00007ffd`2a2e7293 7e0f            jle     <AFTER_LOOP>
00007ffd`2a2e7295 4d63d1          movsxd  r10,r9d
00007ffd`2a2e7298 42030492        add     eax,dword ptr [rdx+r10*4]
00007ffd`2a2e729c 41ffc1          inc     r9d
00007ffd`2a2e729f 443bc9          cmp     r9d,ecx
00007ffd`2a2e72a2 7cf1            jl      00007ffd`2a2e7295

I prototyped the below change in my local branch by modifying the logic in importer.cpp to use zero-extension instead of signed-extension for the span indexer and ran a benchmark. The modified code took approximately one-third less time to run. This optimization may be worth investigating if we believe that developers are iterating over spans in hot loops. (Admittedly, any more complex logic within the loop would almost certainly overwhelm these benchmark results.)

            // Element access
            GenTree*             indexIntPtr = gtNewCastNode(TYP_U_IMPL, indexClone, true /* fromUnsigned */, TYP_U_IMPL);   // <-- modified line
            GenTree*             sizeofNode  = gtNewIconNode(elemSize);
            GenTree*             mulNode     = gtNewOperNode(GT_MUL, TYP_U_IMPL, indexIntPtr, sizeofNode);   // <-- modified line

Method	Toolchain	SpanLength	Mean	Error	StdDev	Ratio	RatioSD
SumInts	baseline	48	2,921.32 us	53.786 us	44.914 us	1.00	0.00
SumInts	modified	48	1,964.96 us	38.825 us	43.154 us	0.67	0.02
SumInts	baseline	512	35,429.46 us	574.800 us	537.669 us	1.00	0.00
SumInts	modified	512	23,219.06 us	457.335 us	698.398 us	0.67	0.02
SumInts	baseline	2048	139,664.62 us	1,799.241 us	1,683.011 us	1.00	0.00
SumInts	modified	2048	93,175.18 us	1,838.916 us	3,586.665 us	0.66	0.04

/cc @dotnet/jit-contrib

category:cq
theme:basic-cq
skill-level:expert
cost:medium
impact:medium

[gfoidl]

The same movsxd is generated when indexing arrays.

The zero-extension could be omitted, if the JIT recognizes this pattern, and operate with the native int size (i.e. ulong on x64, uint), hence emitting code as follows (for x64):

G_M56184_IG03:
       03048A               add      eax, dword ptr [rdx+4*rdi]
       FFC7                 inc      rdi
       3BFE                 cmp      rdi, rsi
       7CF4                 jl       SHORT G_M56184_IG03

[BruceForstall]

@AndyAyersMS

[AndyAyersMS]

Range prop may be able to determine i cannot be negative (definitely for arrays, maybe for spans) and could possibly rewrite the in-loop uses that feed sign-extensions. Might get the simple cases.

Worth taking a quick look.

[mikedn]

@AndyAyersMS I tried that in the past but for some reason I gave up. I'll check what I did back then.

[GrabYourPitchforks]

@AndyAyersMS I didn't investigate anything quite as complex as flowing context that i must be non-negative or backing i with a native int to begin with. In the quick changes I experimented with, there was still a zero-extension (the movsxd r10, r9d instruction became mov r10d, r9d), and this was sufficient to get the performance gain mentioned. I suspect this was a zero-latency mov within the processor?

[mikedn]

and this was sufficient to get the performance gain mentioned. I suspect this was a zero-latency mov within the processor?

Yes. I tried that in the past with array indices. It was working but in some cases involving pointer arithmetic there was a regression due to CSE no longer picking up the casts, those from array indices where zero extending and those from pointer arithmetic were sign extending.

[mikedn]

@AndyAyersMS FWIW I have this in mikedn/coreclr@e0a6e91 It does seem to work (diff corelib -> -368 bytes) but I don't remember if the change was "done" or if there was some kind of problem that made me abandon it. It could be that it was just my general dislike of RangeCheck, I generally don't trust that code. I'll try to figure out what's up with it.

[mikedn]

Ran a diff on the entire FX (after patching a strange issue with missing assertions):

Total bytes of diff: -1414 (-0.01% of base)
    diff is an improvement.
Top file improvements by size (bytes):
        -365 : System.Private.CoreLib.dasm (-0.01% of base)
        -186 : System.Data.Common.dasm (-0.02% of base)
        -160 : System.Private.Xml.dasm (-0.01% of base)
         -87 : Microsoft.CodeAnalysis.CSharp.dasm (0.00% of base)
         -82 : Microsoft.CSharp.dasm (-0.03% of base)
75 total files with size differences (75 improved, 0 regressed), 54 unchanged.
Top method improvements by size (bytes):
         -28 (-0.23% of base) : System.Private.CoreLib.dasm - Dictionary`2:OnDeserialization(ref):this (28 methods)
         -16 (-0.20% of base) : System.Data.Common.dasm - XmlTreeGen:SchemaTree(ref,ref,ref,ref,bool):this
         -15 (-0.47% of base) : System.Private.Xml.dasm - XmlSchemaValidator:EndElementIdentityConstraints(ref,ref,ref):this
         -12 (-7.14% of base) : Microsoft.CSharp.dasm - MethodTypeInferrer:DeduceDependencies():bool:this
         -12 (-0.79% of base) : System.Private.CoreLib.dasm - EnumeratorToIteratorAdapter`1:GetMany(ref):int:this (13 methods)
Top method improvements by size (percentage):
         -12 (-7.14% of base) : Microsoft.CSharp.dasm - MethodTypeInferrer:DeduceDependencies():bool:this
          -6 (-3.02% of base) : Microsoft.CSharp.dasm - MethodSymbol:InferenceMustFail():bool:this
          -6 (-2.67% of base) : Microsoft.CodeAnalysis.CSharp.dasm - PointerTypeSymbol:Equals(ref,bool,bool):bool:this (2 methods)
          -5 (-2.42% of base) : Microsoft.CSharp.dasm - UserStringBuilder:ErrAppendTypeParameters(ref,ref):this
          -2 (-2.06% of base) : Microsoft.CSharp.dasm - MethodTypeInferrer:ExactTypeArgumentInference(ref,ref):this
870 total methods with size differences (870 improved, 0 regressed), 123816 unchanged.

Not a lot, as this saves at most one byte per cast (sometimes mov has one byte less than movsxd, sometimes they have the same size). I'd guesstimate that there are ~5000 casts that are transformed.

PIN data doesn't look too good, a 0.5% regression. Didn't check memory usage but I don't think it will be good either. I was afraid of that, RangeCheck is just inefficient.

[AndyAyersMS]

We're in the 3.0 endgame for the jit, and there doesn't seem to be any cheap / safe way to fix this, so marking as future.

[GrabYourPitchforks]

Related to dotnet/coreclr#21553.

[GrabYourPitchforks]

Related to https://github.com/dotnet/coreclr/issues/23666.