Btw, the assembly code was quite terrible... Here is a new version of it for people interested (feedback still welcomed):
TEXT ·AddVector16(SB),NOSPLIT,$0-48 MOVD $a+0(FP), R0 VLD1.P 32(R0), [V0.D2, V1.D2] VADD V1.B16, V0.B16, V0.B16 MOVD $ret+32(FP), R0 VST1.P [V0.D2], (R0) RET TEXT ·AddVector32(SB),NOSPLIT,$0-96 MOVD $a+0(FP), R0 VLD1.P 64(R0), [V0.D2, V1.D2, V2.D2, V3.D2] VADD V2.B16, V0.B16, V0.B16 VADD V3.B16, V1.B16, V1.B16 MOVD $ret+64(FP), R0 VST1.P [V0.D2, V1.D2], (R0) RET On Wednesday, December 20, 2023 at 7:40:36 AM UTC+8 Clément Jean wrote: > That's what I thought. When I checked the assembly of a generic function I > could see the go.shape part and that's what made me think about > specialisation. > However, after trying to write that myself it just didn't compile. > > I guess, I'll have to do lot of copy pasting ^^ > > On Wednesday, December 20, 2023 at 2:23:01 AM UTC+8 David Finkel wrote: > >> On Thu, Dec 14, 2023 at 10:43 PM Clément Jean <clemen...@gmail.com> >> wrote: >> >>> Hi, >>> >>> I'm currently learning how to write Go assembly on arm64 and I made the >>> following functions: >>> >>> TEXT ·AddVector16(SB),NOSPLIT,$0-48 >>> MOVD a+0(FP), R0 >>> MOVD a+8(FP), R1 >>> MOVD a+16(FP), R2 >>> MOVD a+24(FP), R3 >>> VMOV R0, V0.D[0] >>> VMOV R1, V0.D[1] >>> VMOV R2, V1.D[0] >>> VMOV R3, V1.D[1] >>> VADD V1.B16, V0.B16, V0.B16 >>> VMOV V0.D[0], R0 >>> VMOV V0.D[1], R1 >>> MOVD R0, ret+32(FP) >>> MOVD R1, ret+40(FP) >>> RET >>> >>> TEXT ·AddVector32(SB),NOSPLIT,$0-96 >>> MOVD a+0(FP), R0 >>> MOVD a+8(FP), R1 >>> MOVD a+16(FP), R2 >>> MOVD a+24(FP), R3 >>> MOVD a+32(FP), R4 >>> MOVD a+40(FP), R5 >>> MOVD a+48(FP), R6 >>> MOVD a+56(FP), R7 >>> VMOV R0, V0.D[0] >>> VMOV R1, V0.D[1] >>> VMOV R2, V1.D[0] >>> VMOV R3, V1.D[1] >>> VMOV R4, V2.D[0] >>> VMOV R5, V2.D[1] >>> VMOV R6, V3.D[0] >>> VMOV R7, V3.D[1] >>> VADD V2.B16, V0.B16, V0.B16 >>> VADD V3.B16, V1.B16, V1.B16 >>> VMOV V0.D[0], R0 >>> VMOV V0.D[1], R1 >>> VMOV V1.D[0], R2 >>> VMOV V1.D[1], R3 >>> MOVD R0, ret+64(FP) >>> MOVD R1, ret+72(FP) >>> MOVD R2, ret+80(FP) >>> MOVD R3, ret+88(FP) >>> RET >>> While these work, I'm forced to add the 16 and 32 suffixes. Is there any >>> way to have only one function name and link the assembly to something like: >>> type Uint8x16 [16]uint8 >>> type Uint8x32 [32]uint8 >>> >>> //go:noescape >>> func AddVector[T Uint8x16 | Uint8x32](a, b T) T >>> If there is let me know, otherwise is there some other alternatives? >>> >>> >> Unfortunately, I think a suffix is the best you can do without jumping >> through some serious (and ugly) hoops. >> The generics implementation opted for a "stenciling" approach, (and >> doesn't support specializations) so although the symbols do encode the GC >> shape a bit. >> Here's the design doc for GC Shape Stenciling >> >> https://go.googlesource.com/proposal/+/refs/heads/master/design/generics-implementation-gcshape.md >> >> >> >>> Btw any feedback on the assembly is welcome, I'm new to it. >>> >>> -- >>> You received this message because you are subscribed to the Google >>> Groups "golang-nuts" group. >>> To unsubscribe from this group and stop receiving emails from it, send >>> an email to golang-nuts...@googlegroups.com. >>> To view this discussion on the web visit >>> https://groups.google.com/d/msgid/golang-nuts/3224bd47-2747-4e1d-bbc5-4a5c4a300f85n%40googlegroups.com >>> >>> <https://groups.google.com/d/msgid/golang-nuts/3224bd47-2747-4e1d-bbc5-4a5c4a300f85n%40googlegroups.com?utm_medium=email&utm_source=footer> >>> . >>> >> -- You received this message because you are subscribed to the Google Groups "golang-nuts" group. To unsubscribe from this group and stop receiving emails from it, send an email to golang-nuts+unsubscr...@googlegroups.com. To view this discussion on the web visit https://groups.google.com/d/msgid/golang-nuts/b19562ac-5b2f-44f2-a6f9-77bf7eb8678en%40googlegroups.com.
