fx.reflection: optimise darwin dynamic function calling

fx.reflection/sys/darwin/arm64/callvm.S

@@ -8,155 +8,69 @@ _callvm_invoke_i:
 	stp x29, x30, [sp, #-16]!
 	mov x29, sp
 
-	# store function pointer for later
-	mov x16, x0
-	mov x15, x1
+	# preserve callee-saved registers
+	stp x19, x20, [sp, #-16]!
 
-	# First, set up the fixed arguments
-	# x8 = fixed arg count
-	ldr x8, [x15, #0]
+	# store function pointer and callvm context for later
+	mov x19, x0
+	mov x20, x1
 
-	# arg[0]
-	cmp x8, #0
-	ble .iv.idone
-	ldr x0, [x15, #48]
-
-	# arg[1]
-	cmp x8, #1
-	ble .iv.idone
-	ldr x1, [x15, #56]
-
-	# arg[2]
-	cmp x8, #2
-	ble .iv.idone
-	ldr x2, [x15, #64]
-
-	# arg[3]
-	cmp x8, #3
-	ble .iv.idone
-	ldr x3, [x15, #72]
-
-	# arg[4]
-	cmp x8, #4
-	ble .iv.idone
-	ldr x4, [x15, #80]
-
-	# arg[5]
-	cmp x8, #5
-	ble .iv.idone
-	ldr x5, [x15, #88]
-
-	# arg[6]
-	cmp x8, #6
-	ble .iv.idone
-	ldr x6, [x15, #96]
-
-	# arg[7]
-	cmp x8, #7
-	ble .iv.idone
-	ldr x7, [x15, #104]
-
-.iv.idone:
-	# Next, set up the fixed double arguments
-	# x8 = fixed double arg count
-	ldr x8, [x15, #8]
-
-	# arg[0]
-	cmp x8, #0
-	ble .iv.vdone
-	ldr d0, [x15, #112]
-
-	# arg[1]
-	cmp x8, #1
-	ble .iv.vdone
-	ldr d1, [x15, #120]
-
-	# arg[2]
-	cmp x8, #2
-	ble .iv.vdone
-	ldr d2, [x15, #128]
-
-	# arg[3]
-	cmp x8, #3
-	ble .iv.vdone
-	ldr d3, [x15, #136]
-
-	# arg[4]
-	cmp x8, #4
-	ble .iv.vdone
-	ldr d4, [x15, #144]
-
-	# arg[5]
-	cmp x8, #5
-	ble .iv.vdone
-	ldr d5, [x15, #152]
-
-	# arg[6]
-	cmp x8, #6
-	ble .iv.vdone
-	ldr d6, [x15, #160]
-
-	# arg[7]
-	cmp x8, #7
-	ble .iv.vdone
-	ldr d7, [x15, #168]
-
-.iv.vdone:
-	# Finally, set up the variable arguments
+	# First, set up the variable arguments
 
+	### calculate the amount of stack space needed for the varargs
+	# store the original stack pointer, so it can be restored later
+	mov x0, sp
+	str x0, [sp, #-16]!
 	# x8 = excess arg count
-	ldr x8, [x15, #32]
-
-	# calculate the amount of stack space needed for the varargs
-	mov x10, #0x10
-	mul x12, x8, x10
-
-	# x19 is caller-saved, and we'll need x12 later
-	str x19, [sp, #-16]!
-	mov x19, x12
-
+	ldr x8, [x20, #32]
+	# x8 = excess arg buffer size (8 bytes per value)
+	lsl x8, x8, #3
 	# allocate the stack space
-	sub sp, sp, x19
-	mov x13, sp
+	sub sp, sp, x8
 
-	# get a pointer to the end of the stack vararg buffer
-	#mov x14, #8
-	#mul x14, x14, x11
-	#add x13, x13, x14
+	# re-align the stack, and take a pointer to the newly-allocated buffer
+	mov x0, sp
+	and x0, x0, #0xFFFFFFFFFFFFFFF0
+	mov sp, x0
 
-	# x8: number of var args (decrements with every iteration)
-	# x9: arg src buffer (increments with every iteration)
-	# x13: var arg dest pointer (starts at the end, decrements with every iteration)
-	ldr x9, [x15, #176]
+	ldr x1, [x20, #176]
+	mov x2, x8
 
-.iv.loop:
-	cmp x8, #0
-	beq .iv.loop.end
+	# memcpy(stack vararg buffer, callvm excess buffer, arg size)
+	bl _memcpy
 
-	# read the arg value from the src pointer
-	ldr x14, [x9]
+	# Next, set up the fixed integer arguments
+	ldr x0, [x20, #48]
+	ldr x1, [x20, #56]
+	ldr x2, [x20, #64]
+	ldr x3, [x20, #72]
+	ldr x4, [x20, #80]
+	ldr x5, [x20, #88]
+	ldr x6, [x20, #96]
+	ldr x7, [x20, #104]
 
-	# write it to the stack, and increment the dest pointer
-	str x14, [x13]
-	add x13, x13, #8
-
-	# increment the src pointer, decrement the arg count
-	sub x8, x8, #1
-	add x9, x9, #8
-	b .iv.loop
-.iv.loop.end:
+	# Finally, set up the fixed double arguments
+	ldr d0, [x20, #112]
+	ldr d1, [x20, #120]
+	ldr d2, [x20, #128]
+	ldr d3, [x20, #136]
+	ldr d4, [x20, #144]
+	ldr d5, [x20, #152]
+	ldr d6, [x20, #160]
+	ldr d7, [x20, #168]
 
 	# call the function implementation
-	blr x16
+	blr x19
 
-	# de-allocate the stack varargs buffer (the size is now stored in x19)
-	add sp, sp, x19
+	# restore the saved stack pointer, and de-allocate the stack varargs buffer
+	ldr x9, [x29, #-32]
+	mov sp, x9
 
-	# restore the saved value of x19,
-	ldr x19, [sp], #16
+	# restore callee-saved registers
+	ldp x19, x20, [sp], #16
 
 	# restore the saved stack frame and link pointer
-	ldp x29,x30, [sp], #16
+	ldp x29, x30, [sp], #16
 	ret