From b000857bb4aad2ed0e2639c35136ebc9f3d92cbb Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Juhani=20Krekel=C3=A4?= <juhani@krekelä.fi>
Date: Thu, 24 Feb 2022 21:14:37 +0200
Subject: [PATCH] Implement the RNG for word selection

---
 dosdl.asm         | 281 +++++++++++++++++++++++++++++++++++++++++++++-
 word_selection.py |  28 +++++
 2 files changed, 305 insertions(+), 4 deletions(-)
 create mode 100644 word_selection.py

diff --git a/dosdl.asm b/dosdl.asm
index af58eb4..41a9b6d 100644
--- a/dosdl.asm
+++ b/dosdl.asm
@@ -5,11 +5,36 @@ org 0x100
 
 section .code
 
-select_target:
-	; TODO: Implement target word selection
-	; https://github.com/lynn/hello-wordl/blob/7da40c1f067eb1ec157d4c5b7a9bd8257ed39342/src/util.ts#L14
-	mov si, 1397
+seed equ 20220224
 
+; Seed the rng
+mov ax, seed & 0xffff
+mov dx, seed >> 16
+mov bp, rng_seed
+call store32
+
+select_target:
+	; Call rng to get a number
+	call rng
+	mov bp, rng_output
+	call store32
+
+	; Multiply rng output by number of targer words
+	xor si, si
+	xor dx, dx
+	xor ax, ax
+	mov cx, targets_len
+	.multiply_loop:
+		add ax, [rng_output + 0]
+		adc dx, [rng_output + 2]
+		adc si, 0
+		loop .multiply_loop
+
+	mov ax, si
+	call hexprint16
+	call newline
+
+load_target:
 	shl si, 1
 	shl si, 1
 	add si, targets
@@ -18,6 +43,11 @@ select_target:
 
 	call unpack_target
 
+	; Figure out whether this is a removed word
+	; If so, reroll target word
+	cmp byte [target], '{'
+	je select_target
+
 read_guess:
 	; Number of thus far guessed letters stored in bx
 	xor bx, bx
@@ -428,6 +458,237 @@ unpack_target:
 
 	ret
 
+; out:
+;  dx:ax = result
+rng:
+	; https://github.com/lynn/hello-wordl/blob/7da40c1f067eb1ec157d4c5b7a9bd8257ed39342/src/util.ts#L14
+
+	; rng_t = rng_seed += rng_seed + 0x6d2b79f5
+	mov bp, rng_seed
+	call load32
+	add ax, 0x79f5
+	adc dx, 0x6d2b
+	call store32
+	mov bp, rng_t
+	call store32
+
+	; xorl32 = rng_t
+	mov bp, xorl32
+	call store32
+
+	; xorr32 = rng_t >> 15
+	mov cx, 15
+	call rshift32
+	mov bp, xorr32
+	call store32
+
+	; mull32 = xorl32 ^ xorr32 = rng_t ^ (rng_t >> 15)
+	call xor32
+	mov bp, mull32
+	call store32
+
+	; mulr32 = rng_t | 1
+	mov bp, rng_t
+	call load32
+	or ax, 1
+	mov bp, mulr32
+	call store32
+
+	; rng_t = mull32 * mulr32 = (rng_t ^ (rng_t >> 15)) * (rng_t | 1)
+	call mul32
+	mov bp, rng_t
+	call store32
+
+	; xorl32 = rng_t
+	mov bp, xorl32
+	call store32
+
+	; xorr32 = rng_t >> 7
+	mov cx, 7
+	call rshift32
+	mov bp, xorr32
+	call store32
+
+	; mull32 = xorl32 ^ xorr32 = rng_t ^ (rng_t >> 7)
+	call xor32
+	mov bp, mull32
+	call store32
+
+	; mulr32 = rng_t | 61
+	mov bp, rng_t
+	call load32
+	or ax, 61
+	mov bp, mulr32
+	call store32
+
+	; addl32 = rng_t
+	mov bp, rng_t
+	call load32
+	mov bp, addl32
+	call store32
+
+	; addr32 = mull32 * mulr32 = (rng_t ^ (rng_t >> 7)) * (rng_t | 61)
+	call mul32
+	mov bp, addr32
+	call store32
+
+	; xorl32 = rng_t
+	mov bp, rng_t
+	call load32
+	mov bp, xorl32
+	call store32
+
+	; xorr32 = addl32 + addr32 = rng_t + (rng_t ^ (rng_t >> 7)) * (rng_t | 61)
+	call add32
+	mov bp, xorr32
+	call store32
+
+	; rng_t = xorl32 ^ xorr32 = rng_t ^ (rng_t + (rng_t ^ (rng_t >> 7)) * (rng_t | 61))
+	call xor32
+	mov bp, rng_t
+	call store32
+
+	; xorl32 = rng_t
+	mov bp, xorl32
+	call store32
+
+	; xorr32 = rng_t >> 14
+	mov cx, 14
+	call rshift32
+	mov bp, xorr32
+	call store32
+
+	; result = xorl32 ^ xorr32 = rng_t ^ (rng_t >> 14)
+	call xor32
+
+	ret
+
+; in:
+;  bp = address of uint32
+; out:
+;  dx = high 16 bits
+;  ax = low 16 bits
+load32:
+	mov ax, [bp + 0]
+	mov dx, [bp + 2]
+	ret
+
+; in:
+;  bp = address of uint32
+;  dx = high 16 bits
+;  ax = low 16 bits
+store32:
+	mov [bp + 0], ax
+	mov [bp + 2], dx
+	ret
+
+; in:
+;  dx:ax = uint32 to be shifted
+;  cx = shift amount (must not be 0)
+; out:
+;  dx:ax = result
+rshift32:
+	push cx
+	.loop:
+		shr dx, 1
+		rcr ax, 1
+		loop .loop
+	pop cx
+	ret
+
+; in:
+;  [xorl32], [xorr32] = operands
+; out:
+;  dx:ax = result
+xor32:
+	mov ax, [xorl32 + 0]
+	xor ax, [xorr32 + 0]
+	mov dx, [xorl32 + 2]
+	xor dx, [xorr32 + 2]
+	ret
+
+; in:
+;  [mull32], [mulr32] = operands
+; out:
+;  dx:ax = result
+mul32:
+	push bx
+	push cx
+	; c = 1<<16
+	; l = l₁<<16 + l₀
+	; r = r₁<<16 + r₀
+	; l·r = (l₁<<16 + l₀)·(r₁<<16 + r₀)
+	;     = l₁<<16·r₁<<16 + l₁<<16·r₀ + l₀·r₁<<16 + l₀·r₀
+	;     = l₁·r₁<<32 + l₁·r₀<<16 + l₀·r₁<<16 + l₀·r₀      ||1<<32 = 0 (mod 1<<32)
+	;     = l₁·r₀<<16 + l₀·r₁<<16 + l₀·r₀
+
+	; l₀·r₀
+	mov ax, [mull32 + 0]
+	mul word [mulr32 + 0]
+	mov bx, ax
+	mov cx, dx
+
+	; l₀·r₁<<16
+	mov ax, [mull32 + 0]
+	mul word [mulr32 + 2]
+	add cx, ax
+
+	; l₁·r₀<<16
+	mov ax, [mull32 + 2]
+	mul word [mulr32 + 0]
+	add cx, ax
+
+	mov dx, cx
+	mov ax, bx
+
+	pop cx
+	pop bx
+
+	ret
+
+; in:
+;  [addl32], [addr32] = operands
+; out:
+;  dx:ax = result
+add32:
+	mov ax, [addl32 + 0]
+	mov dx, [addl32 + 2]
+	add ax, [addr32 + 0]
+	adc dx, [addr32 + 2]
+	ret
+
+hexprint32:
+	xchg ax, dx
+	call hexprint16
+	xchg ax, dx
+hexprint16:
+	xchg ah, al
+	call hexprint8
+	xchg ah, al
+hexprint8:
+%rep 4
+	rol al, 1
+%endrep
+	call hexprint4
+%rep 4
+	rol al, 1
+%endrep
+hexprint4:
+	push ax
+	and al, 0xf
+	cmp al, 10
+	jb .digit09
+
+	add al, 'a' - 10 - '0'
+
+	.digit09:
+	add al, '0'
+	mov ah, 0xe
+	int 0x10
+
+	pop ax
+	ret
+
 section .data
 target times 5 db 0
 guess times 5 db 0
@@ -439,6 +700,18 @@ guesses db 0
 
 dictionary_entry times 5 db 0
 
+rng_seed dd 0
+rng_t dd 0
+
+rng_output dd 0
+
+xorl32 dd 0
+xorr32 dd 0
+mull32 dd 0
+mulr32 dd 0
+addl32 dd 0
+addr32 dd 0
+
 section .rodata
 
 not_found_str db ' - word not found$'
diff --git a/word_selection.py b/word_selection.py
new file mode 100644
index 0000000..9f0affe
--- /dev/null
+++ b/word_selection.py
@@ -0,0 +1,28 @@
+import json
+import sys
+
+# https://github.com/lynn/hello-wordl/blob/7da40c1f067eb1ec157d4c5b7a9bd8257ed39342/src/util.ts#L14
+def rshift(n, a): return (n & 0xffff_ffff) >> a
+
+def mulberry32(rng_state):
+	def worker():
+		nonlocal rng_state
+		rng_state += 0x6d2b79f5;
+		t = rng_state
+		t = (t ^ rshift(t, 15)) * (t | 1) & 0xffff_ffff
+		t ^= t + (t ^ rshift(t, 7)) * (t | 61) & 0xffff_ffff
+		return (t ^ (t >> 14))
+	return worker
+
+with open('targets.json', 'r') as f:
+	targets = json.load(f)
+
+targets = targets[:targets.index('murky') + 1]
+targets = [i for i in targets if len(i) == 5]
+
+rng = mulberry32(int(sys.argv[1]))
+while True:
+	num = rng()
+	index = num * len(targets) >> 32
+	print(f'{num:08x}', index, targets[index])
+	if targets[index] != '*****': break