ARM64-SVE: Optimise away instruction when mask is all false

[a74nh]

Consider:

static void falsecndselect1(Vector<int> op1, Vector<int> op2) {
  var result1 = Sve.ConditionalSelect(Sve.CreateFalseMaskInt32(), Sve.Add(op1, op2), op1);
  Consume(result1);
}

G_M42078_IG01:  ;; offset=0x0000
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
						;; size=8 bbWeight=1 PerfScore 1.50
G_M42078_IG02:  ;; offset=0x0008
            pfalse  p0.b
            add     z0.s, p0/m, z0.s, z1.s
            movz    x0, #0x72B8      // code for CSharpTutorials.Program:Consume[System.Numerics.Vector`1[int]](System.Numerics.Vector`1[int])
            movk    x0, #0x2218 LSL #16
            movk    x0, #0xE088 LSL #32
            ldr     x0, [x0]
            blr     x0
						;; size=28 bbWeight=1 PerfScore 8.50
G_M42078_IG03:  ;; offset=0x0024
            ldp     fp, lr, [sp], #0x10
            ret     lr
						;; size=8 bbWeight=1 PerfScore 2.00

static void falsecndselect2(Vector<int> op1, Vector<int> op2) {
  var result2 = Sve.ConditionalSelect(Vector<int>.Zero, Sve.Add(op1, op2), op1);
  Consume(result2);
}

G_M39997_IG01:  ;; offset=0x0000
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
						;; size=8 bbWeight=1 PerfScore 1.50
G_M39997_IG02:  ;; offset=0x0008
            ptrue   p0.s
            movi    v16.4s, #0
            cmpne   p0.s, p0/z, z16.s, #0
            add     z0.s, p0/m, z0.s, z1.s
            movz    x0, #0x72B8      // code for CSharpTutorials.Program:Consume[System.Numerics.Vector`1[int]](System.Numerics.Vector`1[int])
            movk    x0, #0x2218 LSL #16
            movk    x0, #0xE088 LSL #32
            ldr     x0, [x0]
            blr     x0
						;; size=36 bbWeight=1 PerfScore 12.00
G_M39997_IG03:  ;; offset=0x002C
            ldp     fp, lr, [sp], #0x10
            ret     lr
						;; size=8 bbWeight=1 PerfScore 2.00

For both of these, because the mask is all false, the ADD will do nothing. This and the mask can be optimised away. Becoming:

            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
            movz    x0, #0x72B8      // code for CSharpTutorials.Program:Consume[System.Numerics.Vector`1[int]](System.Numerics.Vector`1[int])
            movk    x0, #0x2218 LSL #16
            movk    x0, #0xE088 LSL #32
            ldr     x0, [x0]
            blr     x0
            ldp     fp, lr, [sp], #0x10
            ret     lr

This can potentially be applied to most predicated instructions. However, we should check each individually for any possible side effects such as exceptions.

Full list of SVE1 APIs with HW_Flag_ExplicitMaskedOperation:

Abs,
AbsoluteCompareGreaterThan,
AbsoluteCompareGreaterThanOrEqual,
AbsoluteCompareLessThan,
AbsoluteCompareLessThanOrEqual,
AbsoluteDifference,
AddAcross,
AddRotateComplex,
AddSequentialAcross,
AndAcross,
BooleanNot,
CompareEqual,
CompareGreaterThan,
CompareGreaterThanOrEqual,
CompareLessThan,
CompareLessThanOrEqual,
CompareNotEqualTo,
CompareUnordered,
ConvertToDouble,
ConvertToInt32,
ConvertToInt64,
ConvertToSingle,
ConvertToUInt32,
ConvertToUInt64,
CreateBreakPropagateMask,
Divide,
FusedMultiplyAdd,
FusedMultiplyAddNegated,
FusedMultiplySubtract,
FusedMultiplySubtractNegated,
LeadingSignCount,
LeadingZeroCount,
LoadVectorByteNonFaultingZeroExtendToInt16,
LoadVectorByteNonFaultingZeroExtendToInt32,
LoadVectorByteNonFaultingZeroExtendToInt64,
LoadVectorByteNonFaultingZeroExtendToUInt16,
LoadVectorByteNonFaultingZeroExtendToUInt32,
LoadVectorByteNonFaultingZeroExtendToUInt64,
LoadVectorInt16NonFaultingSignExtendToInt32,
LoadVectorInt16NonFaultingSignExtendToInt64,
LoadVectorInt16NonFaultingSignExtendToUInt32,
LoadVectorInt16NonFaultingSignExtendToUInt64,
LoadVectorInt32NonFaultingSignExtendToInt64,
LoadVectorInt32NonFaultingSignExtendToUInt64,
LoadVectorNonFaulting,
LoadVectorSByteNonFaultingSignExtendToInt16,
LoadVectorSByteNonFaultingSignExtendToInt32,
LoadVectorSByteNonFaultingSignExtendToInt64,
LoadVectorSByteNonFaultingSignExtendToUInt16,
LoadVectorSByteNonFaultingSignExtendToUInt32,
LoadVectorSByteNonFaultingSignExtendToUInt64,
LoadVectorUInt16NonFaultingZeroExtendToInt32,
LoadVectorUInt16NonFaultingZeroExtendToInt64,
LoadVectorUInt16NonFaultingZeroExtendToUInt32,
LoadVectorUInt16NonFaultingZeroExtendToUInt64,
LoadVectorUInt32NonFaultingZeroExtendToInt64,
LoadVectorUInt32NonFaultingZeroExtendToUInt64,
Max,
MaxAcross,
MaxNumber,
MaxNumberAcross,
Min,
MinAcross,
MinNumber,
MinNumberAcross,
Multiply,
MultiplyAdd,
MultiplyAddRotateComplex,
MultiplyExtended,
MultiplySubtract,
Negate,
Not,
OrAcross,
PopCount,
ReciprocalExponent,
ReverseBits,
ReverseElement16,
ReverseElement32,
ReverseElement8,
RoundAwayFromZero,
RoundToNearest,
RoundToNegativeInfinity,
RoundToPositiveInfinity,
RoundToZero,
Scale,
ShiftLeftLogical,
ShiftRightArithmetic,
ShiftRightArithmeticForDivide,
ShiftRightLogical,
SignExtend16,
SignExtend32,
SignExtend8,
Sqrt,
XorAcross,
ZeroExtend16,
ZeroExtend32,
ZeroExtend8,

[dotnet-policy-service]

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

[JulieLeeMSFT]

@a74nh, do we plan to work on this item in .NET 10?

[a74nh]

@a74nh, do we plan to work on this item in .NET 10?

I think this should be part of the list of SVE issues we are looking at for .NET 10

[a74nh]

This will be fixed as part of #115566

❯ $CORE_ROOT/corerun ./bin/release/net10.0/lerp.dll
; Assembly listing for method CSharpTutorials.Program:falsecndselect1(System.Numerics.Vector`1[int],System.Numerics.Vector`1[int]) (FullOpts)
; Emitting BLENDED_CODE for generic ARM64 - Unix
; FullOpts code
; optimized code
; fp based frame
; partially interruptible
; No PGO data
; Final local variable assignments
;
;  V00 arg0         [V00,T00] (  3,  3   )  simd16  ->   d0         single-def <System.Numerics.Vector`1[int]>
;* V01 arg1         [V01    ] (  0,  0   )  simd16  ->  zero-ref    single-def <System.Numerics.Vector`1[int]>
;# V02 OutArgs      [V02    ] (  1,  1   )  struct ( 0) [sp+0x00]   do-not-enreg[XS] addr-exposed "OutgoingArgSpace" <Empty>
;
; Lcl frame size = 0

G_M42078_IG01:  ;; offset=0x0000
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
						;; size=8 bbWeight=1 PerfScore 1.50
G_M42078_IG02:  ;; offset=0x0008
            movz    x0, #0x44B0      // code for CSharpTutorials.Program:Consume[int](System.Numerics.Vector`1[int])
            movk    x0, #0xE5F6 LSL #16
            movk    x0, #0xF378 LSL #32
            ldr     x0, [x0]
            blr     x0
						;; size=20 bbWeight=1 PerfScore 5.50
G_M42078_IG03:  ;; offset=0x001C
            ldp     fp, lr, [sp], #0x10
            ret     lr
						;; size=8 bbWeight=1 PerfScore 2.00

; Total bytes of code 36, prolog size 8, PerfScore 9.00, instruction count 9, allocated bytes for code 36 (MethodHash=581a5ba1) for method CSharpTutorials.Program:falsecndselect1(System.Numerics.Vector`1[int],System.Numerics.Vector`1[int]) (FullOpts)
; ============================================================

; Assembly listing for method CSharpTutorials.Program:falsecndselect2(System.Numerics.Vector`1[int],System.Numerics.Vector`1[int]) (FullOpts)
; Emitting BLENDED_CODE for generic ARM64 - Unix
; FullOpts code
; optimized code
; fp based frame
; partially interruptible
; No PGO data
; Final local variable assignments
;
;  V00 arg0         [V00,T00] (  3,  3   )  simd16  ->   d0         single-def <System.Numerics.Vector`1[int]>
;* V01 arg1         [V01    ] (  0,  0   )  simd16  ->  zero-ref    single-def <System.Numerics.Vector`1[int]>
;# V02 OutArgs      [V02    ] (  1,  1   )  struct ( 0) [sp+0x00]   do-not-enreg[XS] addr-exposed "OutgoingArgSpace" <Empty>
;
; Lcl frame size = 0

G_M39997_IG01:  ;; offset=0x0000
            stp     fp, lr, [sp, #-0x10]!
            mov     fp, sp
						;; size=8 bbWeight=1 PerfScore 1.50
G_M39997_IG02:  ;; offset=0x0008
            movz    x0, #0x44B0      // code for CSharpTutorials.Program:Consume[int](System.Numerics.Vector`1[int])
            movk    x0, #0xE5F6 LSL #16
            movk    x0, #0xF378 LSL #32
            ldr     x0, [x0]
            blr     x0
						;; size=20 bbWeight=1 PerfScore 5.50
G_M39997_IG03:  ;; offset=0x001C
            ldp     fp, lr, [sp], #0x10
            ret     lr
						;; size=8 bbWeight=1 PerfScore 2.00

; Total bytes of code 36, prolog size 8, PerfScore 9.00, instruction count 9, allocated bytes for code 36 (MethodHash=ad7263c2) for method CSharpTutorials.Program:falsecndselect2(System.Numerics.Vector`1[int],System.Numerics.Vector`1[int]) (FullOpts)
; ============================================================