Arm64/Sve: Add CreateTrue*_Mask version that works when immediate are constants

[kunalspathak]

Also fixed a minor issue where we should be allocating vector registers instead of predicate. I found this out when testing with the "full predicate registeR" support.

All tests pass: https://gist.github.com/kunalspathak/86d75283326949a6e0871a2fecf64df3

[kunalspathak]

@dotnet/arm64-contrib

[dotnet-policy-service]

Tagging subscribers to this area: @JulieLeeMSFT, @jakobbotsch
See info in area-owners.md if you want to be subscribed.

[TIHan]

LGTM!

My only question is why we need to add these _Mask versions? Is it that the non-_Mask versions need to actually return a vector and not a mask, therefore we need to introduce these _Mask versions?

[kunalspathak]

LGTM!

My only question is why we need to add these _Mask versions? Is it that the non-_Mask versions need to actually return a vector and not a mask, therefore we need to introduce these _Mask versions?

That's correct.

[tannergooding]

We shouldn't need these with the PR that went in last night to support rewriting TYP_MASK in rationalization and we should already be getting good codegen without it being rewritten to a call in the case that the imm is an in-bounds constant at import -or- becomes a constant post import.

You can see that the following:

public static Vector<byte> M1() => Sve.CreateTrueMaskByte(SveMaskPattern.LargestMultipleOf4);

public static Vector<byte> M2()
{
    var pattern = SveMaskPattern.LargestMultipleOf4;
    return Sve.CreateTrueMaskByte(pattern);
}

Generates:

; Method Program:M1():System.Numerics.Vector`1[ubyte] (FullOpts)
G_M31775_IG01:  ;; offset=0x0000
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
						;; size=8 bbWeight=1 PerfScore 1.50

G_M31775_IG02:  ;; offset=0x0008
            ptrue   p0.b, mul4
            mov     z0.b, p0/z, #1
						;; size=8 bbWeight=1 PerfScore 4.00

G_M31775_IG03:  ;; offset=0x0010
            ldp     fp, lr, [sp], #0x10
            ret     lr
						;; size=8 bbWeight=1 PerfScore 2.00
; Total bytes of code: 24

---

There does notably look to be an issue still with:

[MethodImpl(MethodImplOptions.NoInlining)]
public static Vector<byte> M3(SveMaskPattern pattern)
{
    return Sve.CreateTrueMaskByte(pattern);
}

I've put up a PR (#103356) to handle the node insertion order issue, which fixes it to generate:

; Method Program:M3(ubyte):System.Numerics.Vector`1[ubyte] (FullOpts)
G_M15394_IG01:  ;; offset=0x0000
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
						;; size=8 bbWeight=1 PerfScore 1.50

G_M15394_IG02:  ;; offset=0x0008
            uxtb    w0, w0
            movz    x1, #0x6E98      // code for System.Runtime.Intrinsics.Arm.Sve:CreateTrueMaskByte(ubyte):System.Numerics.Vector`1[ubyte]
            movk    x1, #0xE7A8 LSL #16
            movk    x1, #0x7FFC LSL #32
            ldr     x1, [x1]
            blr     x1
            ptrue   p7.b
            cmpne   p0.b, p7/z, z0.b, #0
            mov     z0.b, p0/z, #1
						;; size=36 bbWeight=1 PerfScore 13.00

G_M15394_IG03:  ;; offset=0x002C
            ldp     fp, lr, [sp], #0x10
            ret     lr
						;; size=8 bbWeight=1 PerfScore 2.00
; Total bytes of code: 52

Which is to say the IR:

N005 (  6,  7) [000003] --C--------                         *  RETURN    simd16 $VN.Void
N004 (  5,  6) [000002] --C--------                         \--*  HWINTRINSIC simd16 ubyte ConvertMaskToVector $200
N003 (  4,  5) [000001] --C--------                            \--*  HWINTRINSIC mask   ubyte CreateTrueMaskByte $140
N002 (  3,  4) [000004] -----------                               \--*  CAST      int <- ubyte <- int $100
N001 (  2,  2) [000000] -----------                                  \--*  LCL_VAR   int    V00 arg0         u:1 (last use) $80

rationalizes into:

               [000005] -----------                            IL_OFFSET void   INLRT @ 0x000[E-]
N001 (  2,  2) [000000] -----------                    t0 =    LCL_VAR   int    V00 arg0         u:1 (last use) $80
                                                            /--*  t0     int    
N002 (  3,  4) [000004] -----------                    t4 = *  CAST      int <- ubyte <- int $100
                                                            /--*  t4     int    arg0 in x0
N003 ( 17,  7) [000006] --C-G------                    t6 = *  CALL      simd16 System.Runtime.Intrinsics.Arm.Sve:CreateTrueMaskByte(ubyte):System.Numerics.Vector`1[ubyte]
N004 (  1,  1) [000007] -----------                    t7 =    HWINTRINSIC mask   ubyte CreateTrueMaskAll
                                                            /--*  t7     mask   
                                                            +--*  t6     simd16 
N005 ( 19,  9) [000008] --C-G------                    t8 = *  HWINTRINSIC mask   ubyte ConvertVectorToMask
                                                            /--*  t8     mask   
N004 (  5,  6) [000002] --C-G------                    t2 = *  HWINTRINSIC simd16 ubyte ConvertMaskToVector $200
                                                            /--*  t2     simd16 
N005 (  6,  7) [000003] --C-G------                         *  RETURN    simd16 $VN.Void

-- There is however still some LSRA issue where referant is nullptr:

>	clrjit_universal_arm64_x64.dll!RefPosition::getRegisterType() Line 2665	C++
 	clrjit_universal_arm64_x64.dll!RefPosition::dump(LinearScan * linearScan) Line 10451	C++
 	clrjit_universal_arm64_x64.dll!LinearScan::newRefPosition(Interval * theInterval, unsigned int theLocation, RefType theRefType, GenTree * theTreeNode, unsigned __int64 mask, unsigned int multiRegIdx) Line 662	C++
 	clrjit_universal_arm64_x64.dll!LinearScan::addKillForRegs(regMaskTP mask, unsigned int currentLoc) Line 722	C++
 	clrjit_universal_arm64_x64.dll!LinearScan::buildKillPositionsForNode(GenTree * tree, unsigned int currentLoc, regMaskTP killMask) Line 1141	C++
 	clrjit_universal_arm64_x64.dll!LinearScan::BuildKills(GenTree * tree, regMaskTP killMask) Line 3257	C++
 	clrjit_universal_arm64_x64.dll!LinearScan::BuildDefWithKills(GenTree * tree, unsigned __int64 dstCandidates, regMaskTP killMask) Line 3304	C++
 	clrjit_universal_arm64_x64.dll!LinearScan::BuildCall(GenTreeCall * call) Line 413	C++
 	clrjit_universal_arm64_x64.dll!LinearScan::BuildNode(GenTree * tree) Line 1076	C++
 	clrjit_universal_arm64_x64.dll!LinearScan::buildRefPositionsForNode(GenTree * tree, unsigned int currentLoc) Line 1787	C++
 	clrjit_universal_arm64_x64.dll!LinearScan::buildIntervals<1>() Line 2589	C++
 	clrjit_universal_arm64_x64.dll!LinearScan::doLinearScan() Line 1494	C++
 	clrjit_universal_arm64_x64.dll!Compiler::compCompile::__l2::<lambda_5>::operator()() Line 5322	C++
 	clrjit_universal_arm64_x64.dll!ActionPhase<`Compiler::compCompile'::`2'::<lambda_5>>::DoPhase() Line 70	C++
 	clrjit_universal_arm64_x64.dll!Phase::Run() Line 61	C++
 	clrjit_universal_arm64_x64.dll!DoPhase<`Compiler::compCompile'::`2'::<lambda_5>>(Compiler * _compiler, Phases _phase, Compiler::compCompile::__l2::<lambda_5> _action) Line 83	C++
 	clrjit_universal_arm64_x64.dll!Compiler::compCompile(void * * methodCodePtr, unsigned int * methodCodeSize, JitFlags * compileFlags) Line 5327	C++
 	clrjit_universal_arm64_x64.dll!Compiler::compCompileHelper(CORINFO_MODULE_STRUCT_ * classPtr, ICorJitInfo * compHnd, CORINFO_METHOD_INFO * methodInfo, void * * methodCodePtr, unsigned int * methodCodeSize, JitFlags * compileFlags) Line 7396	C++
 	clrjit_universal_arm64_x64.dll!`Compiler::compCompile'::`127'::__Body::Run(Compiler::compCompile::__l2::__JITParam * __JITpParam) Line 6532	C++
 	clrjit_universal_arm64_x64.dll!Compiler::compCompile(CORINFO_MODULE_STRUCT_ * classPtr, void * * methodCodePtr, unsigned int * methodCodeSize, JitFlags * compileFlags) Line 6536	C++
 	clrjit_universal_arm64_x64.dll!``jitNativeCode'::`8'::__Body::Run'::`6'::__Body::Run(jitNativeCode::__l8::__Body::Run::__l5::__JITParam * __JITpParam) Line 8035	C++
 	clrjit_universal_arm64_x64.dll!`jitNativeCode'::`8'::__Body::Run(jitNativeCode::__l2::__JITParam * __JITpParam) Line 8038	C++
 	clrjit_universal_arm64_x64.dll!jitNativeCode(CORINFO_METHOD_STRUCT_ * methodHnd, CORINFO_MODULE_STRUCT_ * classPtr, ICorJitInfo * compHnd, CORINFO_METHOD_INFO * methodInfo, void * * methodCodePtr, unsigned int * methodCodeSize, JitFlags * compileFlags, void * inlineInfoPtr) Line 8062	C++
 	clrjit_universal_arm64_x64.dll!CILJit::compileMethod(ICorJitInfo * compHnd, CORINFO_METHOD_INFO * methodInfo, unsigned int flags, unsigned char * * entryAddress, unsigned int * nativeSizeOfCode) Line 291	C++

[kunalspathak]

with #103356, we don't need this anymore. I verified the tests pass with the latest main.