Unnecessary call to CORINFO_HELP_RNGCHKFAIL with subtraction and bitwise and #96046
Description
Reproduction:
static ref readonly byte X(int a) => ref _0xFFThen0x00_512[64 - (a & 63)];
static ReadOnlySpan<byte> _0xFFThen0x00_512
{
[MethodImpl(MethodImplOptions.AggressiveInlining)]
get => MemoryMarshal.AsBytes<ulong>
([
0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF,
0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF, 0xFFFFFFFFFFFFFFFF,
0x0000000000000000, 0x0000000000000000, 0x0000000000000000, 0x0000000000000000,
0x0000000000000000, 0x0000000000000000, 0x0000000000000000, 0x0000000000000000,
]);
}Godbolt link: https://godbolt.org/z/8q9fq6Ksn
Codegen:
C:X(int):byref (FullOpts):
G_M48249_IG01: ;; offset=0x0000
push rax
G_M48249_IG02: ;; offset=0x0001
and edi, 63
mov eax, edi
neg eax
add eax, 64
cmp eax, 128
jae SHORT G_M48249_IG04
mov rcx, 0x7FB1C7369BD0 ; static handle
add rax, rcx
G_M48249_IG03: ;; offset=0x001F
add rsp, 8
ret
G_M48249_IG04: ;; offset=0x0024
call CORINFO_HELP_RNGCHKFAIL
int3 The range check fail call is unnecessary since a & 63 is within 0->63, and then 64 - (a & 63) is then clearly within 1->64, which is clearly within the 0->127.
The issue seems to be caused by the 64 - as opposed to any other part of it.
/cc @EgorBo whomst I was discussing this on Discord with.