RyuJIT: Don't convert GT/LE to EQ/NE for "ARR_LEN cmp 0"

[EgorBo]

RyuJIT converts (uint)x > 0 to x != 0 (same for <=) but I suggest we keep it as it is for things like (uint)array.Length > 0 since != and == operators are not bound checks elimination friendly, e.g.:

publicstaticvoidFoo1(int[]array){if((uint)array.Length>0)array[0]=42;}publicstaticvoidFoo2(int[]array){if(0<(uint)array.Length)// same condition, different orderarray[0]=42;}

Before my changes

; Method Egor:Foo1(System.Int32[]) 4883EC28 subrsp,40 8B4108 moveax, dword ptr [rcx+8] 85C0 testeax,eax 740C je SHORT G_M52406_IG04 83F800 cmpeax,0 760C jbe SHORT G_M52406_IG05 C741102A000000 mov dword ptr [rcx+16],42G_M52406_IG04: 4883C428 addrsp,40 C3 retG_M52406_IG05: E81FC5485F call CORINFO_HELP_RNGCHKFAIL CC int3; Total bytes of code: 34; Method Egor:Foo2(System.Int32[]) 8B4108 moveax, dword ptr [rcx+8] 85C0 testeax,eax7407je SHORT G_M52789_IG04 C741102A000000 mov dword ptr [rcx+16],42G_M52789_IG04: C3 ret; Total bytes of code: 15

^ As you can see, bound checks are removed only for "unnatural" (IMO) condition in Foo3
That's why we have this "unnatural" conditions in String.IsNullOrEmpty and other places.

After my changes

; Method Egor:Foo1(System.Int32[]) 8B4108 moveax, dword ptr [rcx+8] 85C0 testeax,eax7407je SHORT G_M52406_IG04 C741102A000000 mov dword ptr [rcx+16],42G_M52406_IG04: C3 ret; Total bytes of code: 15; Method Egor:Foo2(System.Int32[]) 8B4108 moveax, dword ptr [rcx+8] 85C0 testeax,eax7407je SHORT G_M52789_IG04 C741102A000000 mov dword ptr [rcx+16],42G_M52789_IG04: C3 ret; Total bytes of code: 15

[benaadams]

Span<int> has same issue

publicstaticvoidFoo1(Span<int>array){if((uint)array.Length>0)array[0]=42;}publicstaticvoidFoo2(Span<int>array){if(0<(uint)array.Length)// same condition, different orderarray[0]=42;}

C.Foo1(System.Span`1<Int32>) L0000: pushebp L0001: movebp,esp L0003: moveax,[ebp+0xc] L0006: testeax,eax L0008: jz L0018 L000a: cmpeax,0x0 L000d: jbe L001c L000f: moveax,[ebp+0x8] L0012: mov dword [eax],0x2a L0018: popebp L0019: ret0x8 L001c: call0x62022680 L0021: int3C.Foo2(System.Span`1<Int32>) L0000: moveax,[esp+0x8] L0004: testeax,eax L0006: jz L0012 L0008: moveax,[esp+0x4] L000c: mov dword [eax],0x2a L0012: ret0x8

[sandreenko]

Do you know how hard it is to make it friendly?
As I see we use

runtime/src/coreclr/src/jit/gentree.h

Line 1563 in 8527a99

staticboolOperIsBoundsCheck(genTreeOps op)

to find what we optimize as a bound check and we don't have EQ/NE there, should we add them?

[EgorBo]

As far as I understand it's already handled in #40180

[nathan-moore]

As far as I understand it's already handled in #40180

I checked, it indeed is :). Should also handle the span case.