fix: Fix arm64 assembly stack frame manipulation

arrow/compute/internal/kernels/cast_numeric_neon_arm64.s

@@ -10,13 +10,7 @@ TEXT ·_cast_type_numeric_neon(SB), $0-40
     MOVD len+32(FP), R4
 
 
-    // The Go ABI saves the frame pointer register one word below the 
-    // caller's frame. Make room so we don't overwrite it. Needs to stay 
-    // 16-byte aligned 
-    SUB $16, RSP
-    WORD $0xa9bf7bfd // stp    x29, x30, [sp, #-16]!
     WORD $0x7100181f // cmp    w0, #6
-    WORD $0x910003fd // mov    x29, sp
     BGT LBB0_17
 
     WORD $0x71000c1f // cmp    w0, #3
@@ -4450,9 +4444,6 @@ LBB0_892:
     WORD $0xb800452b // str    w11, [x9], #4
     BNE LBB0_892
 LBB0_893:
-    WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-    // Put the stack pointer back where it was 
-    ADD $16, RSP
     RET
 LBB0_894:
     WORD $0x927b6909 // and x9, x8, #0xffffffe0

arrow/math/int64_neon_arm64.s

@@ -10,13 +10,7 @@ TEXT ·_sum_int64_neon(SB), $0-24
 	MOVD	buf+0(FP), R0
 	MOVD	len+8(FP), R1
 	MOVD	res+16(FP), R2
-
-	// The Go ABI saves the frame pointer register one word below the 
-	// caller's frame. Make room so we don't overwrite it. Needs to stay 
-	// 16-byte aligned 
-	SUB $16, RSP
-	WORD $0xa9bf7bfd // stp x29, x30, [sp, #-16]!
-	WORD $0x910003fd // mov x29, sp
+
 	CBZ R1, LBB0_3
 	WORD $0xf1000c3f // cmp x1, #3
 	BHI LBB0_4
@@ -26,9 +20,6 @@ TEXT ·_sum_int64_neon(SB), $0-24
 LBB0_3:
 	WORD $0xaa1f03e9 // mov     x9, xzr
 	WORD $0xf9000049 // str     x9, [x2]
-	WORD $0xa8c17bfd // ldp     x29, x30, [sp], #16
-	// Put the stack pointer back where it was 
-	ADD $16, RSP
 	RET
 LBB0_4:
 	WORD $0x927ef428 // and     x8, x1, #0xfffffffffffffffc
@@ -59,8 +50,5 @@ LBB0_8:
 	BNE LBB0_8
 LBB0_9:
 	WORD $0xf9000049 // str     x9, [x2]
-	WORD $0xa8c17bfd // ldp     x29, x30, [sp], #16
-	// Put the stack pointer back where it was 
-	ADD $16, RSP
 	RET
 

arrow/math/uint64_neon_arm64.s

@@ -10,13 +10,7 @@ TEXT ·_sum_uint64_neon(SB), $0-24
 	MOVD	buf+0(FP), R0
 	MOVD	len+8(FP), R1
 	MOVD	res+16(FP), R2
-
-	// The Go ABI saves the frame pointer register one word below the 
-	// caller's frame. Make room so we don't overwrite it. Needs to stay 
-	// 16-byte aligned 
-	SUB $16, RSP
-	WORD $0xa9bf7bfd // stp x29, x30, [sp, #-16]!
-	WORD $0x910003fd // mov x29, sp
+
 	CBZ R1, LBB0_3
 	WORD $0xf1000c3f // cmp x1, #3
 	BHI LBB0_4
@@ -26,9 +20,6 @@ TEXT ·_sum_uint64_neon(SB), $0-24
 LBB0_3:
 	WORD $0xaa1f03e9 // mov     x9, xzr
 	WORD $0xf9000049 // str     x9, [x2]
-	WORD $0xa8c17bfd // ldp     x29, x30, [sp], #16
-	// Put the stack pointer back where it was 
-	ADD $16, RSP
 	RET
 LBB0_4:
 	WORD $0x927ef428 // and     x8, x1, #0xfffffffffffffffc
@@ -59,8 +50,5 @@ LBB0_8:
 	BNE LBB0_8
 LBB0_9:
 	WORD $0xf9000049 // str     x9, [x2]
-	WORD $0xa8c17bfd // ldp     x29, x30, [sp], #16
-	// Put the stack pointer back where it was 
-	ADD $16, RSP
 	RET
 

arrow/memory/memory_neon_arm64.s

@@ -11,14 +11,8 @@ TEXT ·_memset_neon(SB), $0-24
 	MOVD	len+8(FP), R1
 	MOVD	c+16(FP), R2
 
-	// The Go ABI saves the frame pointer register one word below the 
-	// caller's frame. Make room so we don't overwrite it. Needs to stay 
-	// 16-byte aligned 
-	SUB $16, RSP
-	WORD $0xa9bf7bfd // stp    x29, x30, [sp, #-16]!
 	WORD $0x8b010008 // add    x8, x0, x1
 	WORD $0xeb00011f // cmp    x8, x0
-	WORD $0x910003fd // mov    x29, sp
 	BLS	LBB0_7
 
 	WORD $0xf100803f // cmp    x1, #32
@@ -43,7 +37,4 @@ LBB0_6:
 	WORD $0xeb09011f // cmp    x8, x9
 	BNE	LBB0_6
 LBB0_7:
-	WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-	// Put the stack pointer back where it was 
-	ADD $16, RSP
 	RET

internal/utils/min_max_neon_arm64.s

@@ -12,13 +12,7 @@ TEXT ·_int32_max_min_neon(SB), $0-32
 	MOVD    minout+16(FP), R2
 	MOVD    maxout+24(FP), R3
 
-	// The Go ABI saves the frame pointer register one word below the
-	// caller's frame. Make room so we don't overwrite it. Needs to stay
-	// 16-byte aligned
-	SUB $16, RSP
-	WORD $0xa9bf7bfd // stp    x29, x30, [sp, #-16]!
 	WORD $0x7100043f // cmp    w1, #1
-	WORD $0x910003fd // mov    x29, sp
 	BLT int32_early_exit
 
 	WORD $0x71001c3f // cmp    w1, #7
@@ -34,9 +28,6 @@ int32_early_exit:
 	WORD $0x52b0000b // mov    w11, #-2147483648
 	WORD $0xb900006b // str    w11, [x3]
 	WORD $0xb900004a // str    w10, [x2]
-	WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-	// Put the stack pointer back where it was
-	ADD $16, RSP
 	RET
 int32_neon:
 	WORD $0x927d7109 // and    x9, x8, #0xfffffff8
@@ -78,9 +69,6 @@ int32_scalar_loop:
 int32_done:
 	WORD $0xb900006b // str    w11, [x3]
 	WORD $0xb900004a // str    w10, [x2]
-	WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-	// Put the stack pointer back where it was
-	ADD $16, RSP
 	RET
 
 // func _uint32_max_min_neon(values unsafe.Pointer, length int, minout, maxout unsafe.Pointer)
@@ -91,13 +79,7 @@ TEXT ·_uint32_max_min_neon(SB), $0-32
 	MOVD    minout+16(FP), R2
 	MOVD    maxout+24(FP), R3
 
-	// The Go ABI saves the frame pointer register one word below the
-	// caller's frame. Make room so we don't overwrite it. Needs to stay
-	// 16-byte aligned
-	SUB $16, RSP
-	WORD $0xa9bf7bfd // stp    x29, x30, [sp, #-16]!
 	WORD $0x7100043f // cmp    w1, #1
-	WORD $0x910003fd // mov    x29, sp
 	BLT uint32_early_exit
 
 	WORD $0x71001c3f // cmp    w1, #7
@@ -113,9 +95,6 @@ uint32_early_exit:
 	WORD $0x1280000b // mov    w11, #-1
 	WORD $0xb900006a // str    w10, [x3]
 	WORD $0xb900004b // str    w11, [x2]
-	WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-	// Put the stack pointer back where it was
-	ADD $16, RSP
 	RET
 uint32_neon:
 	WORD $0x927d7109 // and    x9, x8, #0xfffffff8
@@ -157,9 +136,6 @@ uint32_scalar_loop:
 uint32_done:
 	WORD $0xb900006a // str    w10, [x3]
 	WORD $0xb900004b // str    w11, [x2]
-	WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-	// Put the stack pointer back where it was
-	ADD $16, RSP
 	RET
 
 // func _int64_max_min_neon(values unsafe.Pointer, length int, minout, maxout unsafe.Pointer)
@@ -170,13 +146,7 @@ TEXT ·_int64_max_min_neon(SB), $0-32
 	MOVD    minout+16(FP), R2
 	MOVD    maxout+24(FP), R3
 
-	// The Go ABI saves the frame pointer register one word below the
-	// caller's frame. Make room so we don't overwrite it. Needs to stay
-	// 16-byte aligned
-	SUB $16, RSP
-	WORD $0xa9bf7bfd // stp    x29, x30, [sp, #-16]!
 	WORD $0x7100043f // cmp    w1, #1
-	WORD $0x910003fd // mov    x29, sp
 	BLT int64_early_exit
 
 	WORD $0x2a0103e8 // mov    w8, w1
@@ -192,9 +162,6 @@ int64_early_exit:
 	WORD $0xd2f0000b // mov    x11, #-9223372036854775808
 	WORD $0xf900006b // str    x11, [x3]
 	WORD $0xf900004a // str    x10, [x2]
-	WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-	// Put the stack pointer back where it was
-	ADD $16, RSP
 	RET
 int64_neon:
 	WORD $0x927e7509 // and    x9, x8, #0xfffffffc
@@ -246,9 +213,6 @@ int64_scalar_loop:
 int64_done:
 	WORD $0xf900006b // str    x11, [x3]
 	WORD $0xf900004a // str    x10, [x2]
-	WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-	// Put the stack pointer back where it was
-	ADD $16, RSP
 	RET
 
 
@@ -260,13 +224,7 @@ TEXT ·_uint64_max_min_neon(SB), $0-32
 	MOVD    minout+16(FP), R2
 	MOVD    maxout+24(FP), R3
 
-	// The Go ABI saves the frame pointer register one word below the
-	// caller's frame. Make room so we don't overwrite it. Needs to stay
-	// 16-byte aligned
-	SUB $16, RSP
-	WORD $0xa9bf7bfd // stp    x29, x30, [sp, #-16]!
 	WORD $0x7100043f // cmp    w1, #1
-	WORD $0x910003fd // mov    x29, sp
 	BLT uint64_early_exit
 
 	WORD $0x71000c3f // cmp    w1, #3
@@ -282,9 +240,6 @@ uint64_early_exit:
 	WORD $0x9280000b // mov    x11, #-1
 	WORD $0xf900006a // str    x10, [x3]
 	WORD $0xf900004b // str    x11, [x2]
-	WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-	// Put the stack pointer back where it was
-	ADD $16, RSP
 	RET
 uint64_neon:
 	WORD $0x927e7509 // and    x9, x8, #0xfffffffc
@@ -336,7 +291,4 @@ uint64_scalar_loop:
 uint64_done:
 	WORD $0xf900006a // str    x10, [x3]
 	WORD $0xf900004b // str    x11, [x2]
-	WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-	// Put the stack pointer back where it was
-	ADD $16, RSP
 	RET

parquet/internal/bmi/bitmap_neon_arm64.s

@@ -8,13 +8,7 @@ TEXT ·_levels_to_bitmap_neon(SB), $0-32
     MOVD numLevels+8(FP), R1
     MOVD rhs+16(FP), R2
 
-    // The Go ABI saves the frame pointer register one word below the 
-    // caller's frame. Make room so we don't overwrite it. Needs to stay 
-    // 16-byte aligned 
-    SUB $16, RSP
-    WORD $0xa9bf7bfd // stp    x29, x30, [sp, #-16]!
     WORD $0x7100043f // cmp    w1, #1
-    WORD $0x910003fd // mov    x29, sp
     BLT LBB1_3
 
     WORD $0x71000c3f // cmp    w1, #3
@@ -82,9 +76,6 @@ LBB1_7:
     WORD $0xaa080168 // orr    x8, x11, x8
     BNE LBB1_7
 LBB1_8:
-    WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-    // Put the stack pointer back where it was 
-    ADD $16, RSP
     MOVD R8, res+24(FP)
     RET
 

parquet/internal/utils/bit_packing_neon_arm64.s

@@ -272,7 +272,7 @@
 #define LCPI0_196 $0xfffffffcfffffffe
 #define LCPI0_200 $0xfffffffeffffffff
 
-TEXT ·_unpack32_neon(SB), $0-40
+TEXT ·_unpack32_neon(SB), $496-40
 
 	MOVD in+0(FP), R0
 	MOVD out+8(FP), R1
@@ -281,19 +281,14 @@ TEXT ·_unpack32_neon(SB), $0-40
 	// LEAQ LCDATA1<>(SB), BP
 
 	// %bb.0:
-	// The Go ABI saves the frame pointer register one word below the 
-	// caller's frame. Make room so we don't overwrite it. Needs to stay 
-	// 16-byte aligned 
-	SUB $16, RSP
-	WORD $0xa9ba7bfd // stp    x29, x30, [sp, #-96]!
-	WORD $0xd10643e9 // sub    x9, sp, #400
-	WORD $0xa9016ffc // stp    x28, x27, [sp, #16]
-	WORD $0xa90267fa // stp    x26, x25, [sp, #32]
-	WORD $0x910003fd // mov    x29, sp
-	WORD $0xa9035ff8 // stp    x24, x23, [sp, #48]
-	WORD $0xa90457f6 // stp    x22, x21, [sp, #64]
-	WORD $0xa9054ff4 // stp    x20, x19, [sp, #80]
-	WORD $0x927df13f // and    sp, x9, #0xfffffffffffffff8
+	// TODO: do not use x28, x27, or x18 since they are reserved.
+	// It is probably unnecessary to save the rest of these registers since
+	// the ABI doesn't have callee saved registers
+	WORD $0xa91a6ffc // stp    x28, x27, [sp, #416]
+	WORD $0xa91b67fa // stp    x26, x25, [sp, #432]
+	WORD $0xa91c5ff8 // stp    x24, x23, [sp, #448]
+	WORD $0xa91d57f6 // stp    x22, x21, [sp, #464]
+	WORD $0xa91e4ff4 // stp    x20, x19, [sp, #480]
 	WORD $0x11007c48 // add    w8, w2, #31
 	WORD $0x7100005f // cmp    w2, #0
 	WORD $0x1a82b108 // csel    w8, w8, w2, lt
@@ -6919,14 +6914,13 @@ LBB0_158:
 
 LBB0_156:
 	WORD $0x531b6a60    // lsl    w0, w19, #5
-	WORD $0x910003bf    // mov    sp, x29
-	WORD $0xa9454ff4    // ldp    x20, x19, [sp, #80]
-	WORD $0xa94457f6    // ldp    x22, x21, [sp, #64]
-	WORD $0xa9435ff8    // ldp    x24, x23, [sp, #48]
-	WORD $0xa94267fa    // ldp    x26, x25, [sp, #32]
-	WORD $0xa9416ffc    // ldp    x28, x27, [sp, #16]
-	WORD $0xa8c67bfd    // ldp    x29, x30, [sp], #96
-	// Put the stack pointer back where it was 
-	ADD $16, RSP
+	WORD $0xa95e4ff4    // ldp    x20, x19, [sp, #480]
+	WORD $0xa95d57f6    // ldp    x22, x21, [sp, #464]
+	WORD $0xa95c5ff8    // ldp    x24, x23, [sp, #448]
+	WORD $0xa95b67fa    // ldp    x26, x25, [sp, #432]
+	WORD $0xa95a6ffc    // ldp    x28, x27, [sp, #416]
+	// The raw compiler-generated body uses x30 as a scratch register. The
+	// Go-generated prologue saved the return address at the bottom of the frame.
+	MOVD 0(RSP), R30
 	MOVD R0, num+32(FP)
 	RET

parquet/internal/utils/unpack_bool_neon_arm64.s

@@ -11,13 +11,6 @@ TEXT ·_bytes_to_bools_neon(SB), $0-32
     MOVD out+16(FP), R2
     MOVD outlen+24(FP), R3
 
-    // The Go ABI saves the frame pointer register one word below the
-    // caller's frame. Make room so we don't overwrite it. Needs to stay
-    // 16-byte aligned
-    SUB $16, RSP
-    WORD $0xa9bf7bfd // stp    x29, x30, [sp, #-16]!
-    WORD $0x910003fd // mov    x29, sp
-
     WORD $0x7100043f // cmp    w1, #1
     BLT done
 
@@ -145,7 +138,4 @@ scalar_next:
     JMP scalar_loop
 
 done:
-    WORD $0xa8c17bfd // ldp    x29, x30, [sp], #16
-    // Put the stack pointer back where it was
-    ADD $16, RSP
     RET