feat(arm64): whole-vector moves, bookkeeping ops and truncating-move lowering

Assisted-by: GLM 5.3 Flash
2026-09-20 21:17:20 +02:00
parent 81e2673923
commit b0f9071bf5
6 changed files with 318 additions and 30 deletions
@@ -0,0 +1,31 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// Differential kernel for the arm64 bookkeeping statements: the funcdata.h
+// pseudo-directives (GO_ARGS, NO_LOCAL_POINTERS, FUNCDATA, PCDATA) contribute
+// no instruction bytes, and every function is byte-compared against
+// go tool asm.
+
+#include "textflag.h"
+#include "funcdata.h"
+
+// func bookkeep()
+TEXT ·bookkeep(SB), NOSPLIT, $8-0
+	GO_ARGS
+	FUNCDATA $3, inline_tree(SB)
+	PCDATA   $1, $2
+	MOVD     R1, 0(RSP)
+	RET
+
+// func bookkeepNoLocals()
+TEXT ·bookkeepNoLocals(SB), NOSPLIT, $16-0
+	NO_LOCAL_POINTERS
+	PCDATA            $0, $0
+	PCDATA            $1, $1
+	MOVD              R2, 8(RSP)
+	RET
+
+// func bookkeepPlain()
+TEXT ·bookkeepPlain(SB), NOSPLIT, $0-0
+	MOVD R3, R4
+	RET
@@ -0,0 +1,57 @@
+// Copyright (c) 2026 Petr Balvín <opensource@petrbalvin.org> (https://petrbalvin.org)
+// SPDX-License-Identifier: BSD-3-Clause
+
+// Differential kernel for the arm64 whole-vector moves between a general
+// register and an arranged vector (VMOV/VDUP Rs, Vd.<T>), the two-operand
+// accumulate spellings VADD/VSUB Vm, Vn, and the toolchain-reserved
+// R18_PLATFORM register name.  Every function is byte-compared against
+// go tool asm.
+
+#include "textflag.h"
+
+// func gpIntoVector()
+TEXT ·gpIntoVector(SB), NOSPLIT, $0-0
+	VMOV R1, V2.B8
+	VMOV R3, V4.B16
+	VMOV R5, V6.H4
+	VMOV R7, V8.H8
+	VMOV R9, V10.S2
+	VMOV R11, V12.S4
+	VMOV R13, V14.D2
+	VDUP R15, V16.B8
+	VDUP R17, V18.B16
+	VDUP R19, V20.H8
+	VDUP R21, V22.S4
+	VDUP R23, V24.D2
+	RET
+
+// func simdAccumulate()
+TEXT ·simdAccumulate(SB), NOSPLIT, $0-0
+	VADD V7, V8
+	VSUB V7, V8
+	VADD V1, V2
+	VSUB V30, V31
+	VADD V0.B16, V1.B16, V2.B16
+	VSUB V0.S4, V1.S4, V2.S4
+	RET
+
+// func truncMove()
+TEXT ·truncMove(SB), NOSPLIT, $0-0
+	MOVB  R3, R4
+	MOVH  R5, R6
+	MOVW  R9, R10
+	MOVBU R3, R4
+	MOVHU R3, R4
+	MOVWU R3, R4
+	MOVD  R3, R4
+	RET
+
+// func platformRegister()
+TEXT ·platformRegister(SB), NOSPLIT, $0-0
+	MOVD R18_PLATFORM, R3
+	MOVW R18_PLATFORM, R4
+	MOVD R3, R18_PLATFORM
+	MOVD 0x68(R18_PLATFORM), R5
+	MOVD R5, 0x68(R18_PLATFORM)
+	MOVW 8(R18_PLATFORM), R6
+	RET