Implement SVE2 ShiftLeftAndInsert

[snickolls-arm]

@a74nh @kunalspathak

Contributes to #115479

[kunalspathak]

you shouldn't need this because that should be already calculated as part of line 411 above under HW_Category_ShiftLeftByImmediate

[kunalspathak]

can you share sample disassembly for 3 scenarios?

• When shift is constant and within bounds
• When shift is constant and out of bounds
• When shift is variable

[snickolls-arm]

can you share sample disassembly for 3 scenarios?

• When shift is constant and within bounds
• When shift is constant and out of bounds
• When shift is variable

1 generates the following instruction:

// Console.WriteLine(Sve2.ShiftLeftAndInsert(u, v, 21));
...
IN0009: 00002C      sli     z16.s, z8.s, #21
...

2 and 3 both generate a runtime exception (3 specifically when that variable value is out of range):

// Console.WriteLine(Sve2.ShiftLeftAndInsert(u, v, 35));

System.TypeInitializationException: The type initializer for 'JIT.HardwareIntrinsics.Arm._Sve2.Program' threw an exception.
 ---> System.ArgumentOutOfRangeException: Specified argument was out of the range of valid values.
   at System.Runtime.Intrinsics.Arm.Sve2.ShiftLeftAndInsert(Vector`1 left, Vector`1 right, Byte shift)

I can't see the exact code being generated for this, because the intrinsic isn't being generated inline anymore. The function contains a call to the intrinsic which I'm guessing contains the jump to throw.

3 generates this instruction when the variable is in range:

//  byte a = 10;
//  Console.WriteLine(Sve2.ShiftLeftAndInsert(u, v, a));
...
IN0012: 000050      sli     z16.s, z8.s, #10
...

The rest of the C# for reference (I dumped these by adding them to the test suite):

        public static void Dump_Test_1()
        {
            var u = new Vector<int>();
            var v = new Vector<int>();
            Console.WriteLine(Sve2.ShiftLeftAndInsert(u, v, 21));
        }

        public static void Dump_Test_2()
        {
            var u = new Vector<int>();
            var v = new Vector<int>();
            Console.WriteLine(Sve2.ShiftLeftAndInsert(u, v, 35));
        }

        public static void Dump_Test_3()
        {
            var u = new Vector<int>();
            var v = new Vector<int>();
            var rand = new Random();
            byte a = (byte)rand.Next();
            Console.WriteLine(Sve2.ShiftLeftAndInsert(u, v, a));
        }

[kunalspathak]

can you try this?

       [MethodImpl(MethodImplOptions.NoInlining)]
        public static Vector<int> Dump_Test_2()
        {
             return Sve2.ShiftLeftAndInsert(u, v, 35);
        }

       [MethodImpl(MethodImplOptions.NoInlining)]
        public static Vector<int> Dump_Test_3(int a)
        {
             return Sve2.ShiftLeftAndInsert(u, v, a);
        }

[snickolls-arm]

can you try this?

       [MethodImpl(MethodImplOptions.NoInlining)]
        public static Vector<int> Dump_Test_2()
        {
             return Sve2.ShiftLeftAndInsert(u, v, 35);
        }

       [MethodImpl(MethodImplOptions.NoInlining)]
        public static Vector<int> Dump_Test_3(int a)
        {
             return Sve2.ShiftLeftAndInsert(u, v, a);
        }

The top one is a function that just branches straight to CORINFO_HELP_THROW_ARGUMENTOUTOFRANGEEXCEPTION.

The second one has this:

; Total bytes of code 64, prolog size 8, PerfScore 12.00, instruction count 16, allocated bytes for cod
e 64 (MethodHash=131eba93) for method JIT.HardwareIntrinsics.Arm.Program:Dump_Test_3(ubyte):System.Nume
rics.Vector`1[int] (FullOpts)
; ============================================================

*************** After end code gen, before unwindEmit()
G_M17772_IG01:        ; func=00, offs=0x000000, size=0x0008, bbWeight=1, PerfScore 1.50, gcrefRegs=0000
 {}, byrefRegs=0000 {}, byref, nogc <-- Prolog IG

IN000d: 000000      stp     fp, lr, [sp, #-0x10]!
IN000e: 000004      mov     fp, sp

G_M17772_IG02:        ; offs=0x000008, size=0x0028, bbWeight=1, PerfScore 8.50, gcrefRegs=0000 {}, byre
fRegs=0000 {}, BB01 [0000], byref, isz

IN0001: 000008      uxtb    w0, w0
IN0002: 00000C      cmp     w0, #32
IN0003: 000010      bhs     G_M17772_IG04
IN0004: 000014      movi    v1.4s, #0
IN0005: 000018      movi    v0.4s, #0
IN0006: 00001C      movz    x1, #0xD3E0      // code for System.Runtime.Intrinsics.Arm.Sve2:ShiftLeftAn
dInsert(System.Numerics.Vector`1[int],System.Numerics.Vector`1[int],ubyte):System.Numerics.Vector`1[int
]
IN0007: 000020      movk    x1, #0x3E6F LSL #16
IN0008: 000024      movk    x1, #0xE2E7 LSL #32
IN0009: 000028      ldr     x1, [x1]
IN000a: 00002C      blr     x1

G_M17772_IG03:        ; offs=0x000030, size=0x0008, bbWeight=1, PerfScore 2.00, epilog, nogc, extend

IN000f: 000030      ldp     fp, lr, [sp], #0x10
IN0010: 000034      ret     lr

G_M17772_IG04:        ; offs=0x000038, size=0x0008, bbWeight=0, PerfScore 0.00, gcVars=0000000000000000
 {}, gcrefRegs=0000 {}, byrefRegs=0000 {}, BB02 [0001], gcvars, byref

IN000b: 000038      bl      CORINFO_HELP_THROW_ARGUMENTOUTOFRANGEEXCEPTION
IN000c: 00003C      brk     #0

and this really big function is called for ShiftLeftAndInsert:

; Total bytes of code 300, prolog size 8, PerfScore 71.00, instruction count 75, allocated bytes for co
de 300 (MethodHash=0b353a25) for method System.Runtime.Intrinsics.Arm.Sve2:ShiftLeftAndInsert(System.Nu
merics.Vector`1[int],System.Numerics.Vector`1[int],ubyte):System.Numerics.Vector`1[int] (FullOpts)
; ============================================================

*************** After end code gen, before unwindEmit()
G_M50650_IG01:        ; func=00, offs=0x000000, size=0x0008, bbWeight=1, PerfScore 1.50, gcrefRegs=0000
 {}, byrefRegs=0000 {}, byref, nogc <-- Prolog IG

IN0048: 000000      stp     fp, lr, [sp, #-0x10]!
IN0049: 000004      mov     fp, sp

G_M50650_IG02:        ; offs=0x000008, size=0x0018, bbWeight=1, PerfScore 4.50, gcrefRegs=0000 {}, byre
fRegs=0000 {}, BB01 [0000], byref, isz

IN0001: 000008      uxtb    w0, w0
IN0002: 00000C      cmp     w0, #32
IN0003: 000010      bhs     G_M50650_IG36
IN0004: 000014      adr     x1, [G_M50650_IG03]
IN0005: 000018      add     x1, x1, x0,  LSL #3
IN0006: 00001C      br      x1

G_M50650_IG03:        ; offs=0x000020, size=0x0008, bbWeight=1, PerfScore 2.00, BB01 [0000], extend

IN0007: 000020      sli     z0.s, z1.s, #0
IN0008: 000024      b       G_M50650_IG35

G_M50650_IG04:        ; offs=0x000028, size=0x0008, bbWeight=1, PerfScore 2.00, BB01 [0000], extend

IN0009: 000028      sli     z0.s, z1.s, #1
IN000a: 00002C      b       G_M50650_IG35

G_M50650_IG05:        ; offs=0x000030, size=0x0008, bbWeight=1, PerfScore 2.00, BB01 [0000], extend

IN000b: 000030      sli     z0.s, z1.s, #2
IN000c: 000034      b       G_M50650_IG35

G_M50650_IG06:        ; offs=0x000038, size=0x0008, bbWeight=1, PerfScore 2.00, BB01 [0000], extend

IN000d: 000038      sli     z0.s, z1.s, #3
IN000e: 00003C      b       G_M50650_IG35

G_M50650_IG07:        ; offs=0x000040, size=0x0008, bbWeight=1, PerfScore 2.00, BB01 [0000], extend

IN000f: 000040      sli     z0.s, z1.s, #4
IN0010: 000044      b       G_M50650_IG35

... and so on up to #31 ...

G_M50650_IG35:        ; offs=0x00011C, size=0x0008, bbWeight=1, PerfScore 2.00, epilog, nogc, extend

IN004a: 00011C      ldp     fp, lr, [sp], #0x10
IN004b: 000120      ret     lr

G_M50650_IG36:        ; offs=0x000124, size=0x0008, bbWeight=0, PerfScore 0.00, gcVars=0000000000000000 {}, gcrefRegs=0000 {}, byrefRegs=0000 {}, BB02 [0001], gcvars, byref

IN0046: 000124      bl      CORINFO_HELP_THROW_ARGUMENTOUTOFRANGEEXCEPTION
IN0047: 000128      brk     #0

[kunalspathak]

LGTM