More ARM code thresholding work.

Fix do_threshold_1 and implement do_threshold_4.

do_threshold_1 checks for white and shortcuts the work if it
can.

There are 2 ARM variants of do_threshold_4. One for ARMs that
support unaligned loads, and one for the (rare configurations)
that don't. The former checks for white and shortcuts the work.

1 files changed, 250 insertions, 49 deletions

diff --git a/source/fitz/halftone.c b/source/fitz/halftone.c
index 2df8e3e7..186c0aa6 100644
--- a/source/fitz/halftone.c
+++ b/source/fitz/halftone.c
@@ -147,7 +147,6 @@ static void make_ht_line(unsigned char *buf, fz_halftone *ht, int x, int y, int
 typedef void (threshold_fn)(const unsigned char *ht_line, const unsigned char *pixmap, unsigned char *out, int w, int ht_len);
 
 #ifdef ARCH_ARM
-
 static void
 do_threshold_1(const unsigned char * restrict ht_line, const unsigned char * restrict pixmap, unsigned char *restrict out, int w, int ht_len)
 __attribute__((naked));
@@ -166,95 +165,110 @@ do_threshold_1(const unsigned char * restrict ht_line, const unsigned char * res
 	"@ <> = ht_len						\n"
 	"ldr	r9, [r13,#6*4]		@ r9 = ht_len		\n"
 	"subs	r3, r3, #7		@ r3 = w -= 7		\n"
-	"blt	2f			@ while (w > 0) {	\n"
+	"ble	2f			@ while (w > 0) {	\n"
 	"mov	r12,r9			@ r12= l = ht_len	\n"
+	"b	1f						\n"
+	"9:							\n"
+	"strb	r14,[r2], #1		@ *out++ = 0		\n"
+	"subs	r12,r12,#8		@ r12 = l -= 8		\n"
+	"moveq	r12,r9			@ if(l==0) l = ht_len	\n"
+	"subeq	r0, r0, r9		@          ht_line -= l	\n"
+	"subs	r3, r3, #8		@ w -= 8		\n"
+	"ble	2f			@ }			\n"
 	"1:							\n"
+	"ldr	r14,[r1], #4		@ r14= pixmap[0..3]	\n"
+	"ldr	r5, [r1], #4		@ r5 = pixmap[4..7]	\n"
+	"ldr	r6, [r1], #4		@ r6 = pixmap[8..11]	\n"
+	"ldr	r7, [r1], #4		@ r7 = pixmap[12..15]	\n"
+	"ldrb	r4, [r0], #8		@ r0 = ht_line += 8	\n"
+	"eors	r14,r14,r5, ROR #8	@ if (white)		\n"
+	"teqeq	r6, r7, ROR #8		@			\n"
+	"beq	9b			@	white		\n"
+	"ldrb	r5, [r1, #-16]		@ r5 = pixmap[0]	\n"
+	"ldrb	r6, [r0, #-7]		@ r6 = ht_line[1]	\n"
+	"ldrb	r7, [r1, #-14]		@ r7 = pixmap[2]	\n"
 	"mov	r14,#0			@ r14= h = 0		\n"
-	"ldrb	r4, [r0], #1		@ r4 = ht_line[0]	\n"
-	"ldrb	r5, [r1], #2		@ r5 = pixmap[0]	\n"
-	"ldrb	r6, [r0], #1		@ r6 = ht_line[1]	\n"
-	"ldrb	r7, [r1], #2		@ r7 = pixmap[2]	\n"
 	"cmp	r5, r4			@ if (r5 < r4)		\n"
 	"orrlt	r14,r14,#0x80		@	h |= 0x80	\n"
-	"ldrb	r4, [r0], #1		@ r4 = ht_line[2]	\n"
-	"ldrb	r5, [r1], #2		@ r5 = pixmap[4]	\n"
+	"ldrb	r4, [r0, #-6]		@ r4 = ht_line[2]	\n"
+	"ldrb	r5, [r1, #-12]		@ r5 = pixmap[4]	\n"
 	"cmp	r7, r6			@ if (r7 < r6)		\n"
 	"orrlt	r14,r14,#0x40		@	h |= 0x40	\n"
-	"ldrb	r6, [r0], #1		@ r6 = ht_line[3]	\n"
-	"ldrb	r7, [r1], #2		@ r7 = pixmap[6]	\n"
+	"ldrb	r6, [r0, #-5]		@ r6 = ht_line[3]	\n"
+	"ldrb	r7, [r1, #-10]		@ r7 = pixmap[6]	\n"
 	"cmp	r5, r4			@ if (r5 < r4)		\n"
 	"orrlt	r14,r14,#0x20		@	h |= 0x20	\n"
-	"ldrb	r4, [r0], #1		@ r4 = ht_line[2]	\n"
-	"ldrb	r5, [r1], #2		@ r5 = pixmap[4]	\n"
+	"ldrb	r4, [r0, #-4]		@ r4 = ht_line[4]	\n"
+	"ldrb	r5, [r1, #-8]		@ r5 = pixmap[8]	\n"
 	"cmp	r7, r6			@ if (r7 < r6)		\n"
 	"orrlt	r14,r14,#0x10		@	h |= 0x10	\n"
-	"ldrb	r6, [r0], #1		@ r6 = ht_line[3]	\n"
-	"ldrb	r7, [r1], #2		@ r7 = pixmap[6]	\n"
+	"ldrb	r6, [r0, #-3]		@ r6 = ht_line[5]	\n"
+	"ldrb	r7, [r1, #-6]		@ r7 = pixmap[10]	\n"
 	"cmp	r5, r4			@ if (r5 < r4)		\n"
 	"orrlt	r14,r14,#0x08		@	h |= 0x08	\n"
-	"ldrb	r4, [r0], #1		@ r4 = ht_line[2]	\n"
-	"ldrb	r5, [r1], #2		@ r5 = pixmap[4]	\n"
+	"ldrb	r4, [r0, #-2]		@ r4 = ht_line[6]	\n"
+	"ldrb	r5, [r1, #-4]		@ r5 = pixmap[12]	\n"
 	"cmp	r7, r6			@ if (r7 < r6)		\n"
 	"orrlt	r14,r14,#0x04		@	h |= 0x04	\n"
-	"ldrb	r6, [r0], #1		@ r6 = ht_line[3]	\n"
-	"ldrb	r7, [r1], #2		@ r7 = pixmap[6]	\n"
+	"ldrb	r6, [r0, #-1]		@ r6 = ht_line[7]	\n"
+	"ldrb	r7, [r1, #-2]		@ r7 = pixmap[14]	\n"
 	"cmp	r5, r4			@ if (r5 < r4)		\n"
 	"orrlt	r14,r14,#0x02		@	h |= 0x02	\n"
 	"cmp	r7, r6			@ if (r7 < r6)		\n"
 	"orrlt	r14,r14,#0x01		@	h |= 0x01	\n"
 	"subs	r12,r12,#8		@ r12 = l -= 8		\n"
+	"strb	r14,[r2], #1		@ *out++ = h		\n"
 	"moveq	r12,r9			@ if(l==0) l = ht_len	\n"
-	"subeq	r6, r6, r12,LSL #2	@          ht_line -= l	\n"
+	"subeq	r0, r0, r9		@          ht_line -= l	\n"
 	"subs	r3, r3, #8		@ w -= 8		\n"
-	"strb	r14,[r2], #1		@ *out++ = h		\n"
 	"bgt	1b			@ }			\n"
-	"1:							\n"
-	"adds	r3, r3, #6		@ w += 6		\n"
-	"blt	3f			@ if (w < 0) {		\n"
+	"2:							\n"
+	"adds	r3, r3, #7		@ w += 7		\n"
+	"ble	4f			@ if (w >= 0) {		\n"
 	"ldrb	r4, [r0], #1		@ r4 = ht_line[0]	\n"
 	"ldrb	r5, [r1], #2		@ r5 = pixmap[0]	\n"
 	"mov	r14, #0			@ r14= h = 0		\n"
 	"cmp	r5, r4			@ if (r5 < r4)		\n"
 	"orrlt	r14,r14,#0x80		@	h |= 0x80	\n"
-	"teq	r3, #0			@			\n"
-	"ldrneb	r4, [r0], #1		@ r6 = ht_line[1]	\n"
-	"ldrneb	r5, [r1], #2		@ r7 = pixmap[2]	\n"
-	"beq	2f			@			\n"
+	"cmp	r3, #1			@			\n"
+	"ldrgtb	r4, [r0], #1		@ r6 = ht_line[1]	\n"
+	"ldrgtb	r5, [r1], #2		@ r7 = pixmap[2]	\n"
+	"ble	3f			@			\n"
 	"cmp	r5, r4			@ if (r5 < r4)		\n"
 	"orrlt	r14,r14,#0x40		@	h |= 0x40	\n"
-	"teq	r3, #1			@			\n"
-	"ldrneb	r4, [r0], #1		@ r6 = ht_line[1]	\n"
-	"ldrneb	r5, [r1], #2		@ r7 = pixmap[2]	\n"
-	"beq	2f			@			\n"
+	"cmp	r3, #2			@			\n"
+	"ldrgtb	r4, [r0], #1		@ r6 = ht_line[2]	\n"
+	"ldrgtb	r5, [r1], #2		@ r7 = pixmap[4]	\n"
+	"ble	3f			@			\n"
 	"cmp	r5, r4			@ if (r5 < r4)		\n"
 	"orrlt	r14,r14,#0x20		@	h |= 0x20	\n"
-	"teq	r3, #2			@			\n"
-	"ldrneb	r4, [r0], #1		@ r6 = ht_line[1]	\n"
-	"ldrneb	r5, [r1], #2		@ r7 = pixmap[2]	\n"
-	"beq	2f			@			\n"
+	"cmp	r3, #3			@			\n"
+	"ldrgtb	r4, [r0], #1		@ r6 = ht_line[3]	\n"
+	"ldrgtb	r5, [r1], #2		@ r7 = pixmap[6]	\n"
+	"ble	3f			@			\n"
 	"cmp	r5, r4			@ if (r5 < r4)		\n"
 	"orrlt	r14,r14,#0x10		@	h |= 0x10	\n"
-	"teq	r3, #3			@			\n"
-	"ldrneb	r4, [r0], #1		@ r6 = ht_line[1]	\n"
-	"ldrneb	r5, [r1], #2		@ r7 = pixmap[2]	\n"
-	"beq	2f			@			\n"
+	"cmp	r3, #4			@			\n"
+	"ldrgtb	r4, [r0], #1		@ r6 = ht_line[4]	\n"
+	"ldrgtb	r5, [r1], #2		@ r7 = pixmap[8]	\n"
+	"ble	3f			@			\n"
 	"cmp	r5, r4			@ if (r5 < r4)		\n"
 	"orrlt	r14,r14,#0x08		@	h |= 0x08	\n"
-	"teq	r3, #4			@			\n"
-	"ldrneb	r4, [r0], #1		@ r6 = ht_line[1]	\n"
-	"ldrneb	r5, [r1], #2		@ r7 = pixmap[2]	\n"
-	"beq	2f			@			\n"
+	"cmp	r3, #5			@			\n"
+	"ldrgtb	r4, [r0], #1		@ r6 = ht_line[5]	\n"
+	"ldrgtb	r5, [r1], #2		@ r7 = pixmap[10]	\n"
+	"ble	3f			@			\n"
 	"cmp	r5, r4			@ if (r5 < r4)		\n"
 	"orrlt	r14,r14,#0x04		@	h |= 0x04	\n"
-	"teq	r3, #5			@			\n"
-	"ldrneb	r4, [r0], #1		@ r6 = ht_line[1]	\n"
-	"ldrneb	r5, [r1], #2		@ r7 = pixmap[2]	\n"
-	"beq	2f			@			\n"
+	"cmp	r3, #6			@			\n"
+	"ldrgtb	r4, [r0], #1		@ r6 = ht_line[6]	\n"
+	"ldrgtb	r5, [r1], #2		@ r7 = pixmap[12]	\n"
+	"ble	3f			@			\n"
 	"cmp	r5, r4			@ if (r5 < r4)		\n"
 	"orrlt	r14,r14,#0x02		@	h |= 0x02	\n"
-	"2:							\n"
-	"strb	r14,[r2]		@ *out = h		\n"
 	"3:							\n"
+	"strb	r14,[r2]		@ *out = h		\n"
+	"4:							\n"
 	"ldmfd	r13!,{r4-r7,r9,PC}	@ pop, return to thumb	\n"
 	ENTER_THUMB
 	);
@@ -325,6 +339,192 @@ static void do_threshold_1(const unsigned char * restrict ht_line, const unsigne
 	white = 0xFF. Reversing these tests enables us to maintain that
 	BlackIs1 in bitmaps.
 */
+#ifdef ARCH_ARM
+static void
+do_threshold_4(const unsigned char * restrict ht_line, const unsigned char * restrict pixmap, unsigned char *restrict out, int w, int ht_len)
+__attribute__((naked));
+
+#ifdef ARCH_ARM_CAN_LOAD_UNALIGNED
+static void
+do_threshold_4(const unsigned char * restrict ht_line, const unsigned char * restrict pixmap, unsigned char *restrict out, int w, int ht_len)
+{
+	asm volatile(
+	ENTER_ARM
+	// Store one more reg that required to keep double stack alignment
+	"stmfd	r13!,{r4-r7,r9,r14}				\n"
+	"@ r0 = ht_line						\n"
+	"@ r1 = pixmap						\n"
+	"@ r2 = out						\n"
+	"@ r3 = w						\n"
+	"@ <> = ht_len						\n"
+	"ldr	r9, [r13,#6*4]		@ r9 = ht_len		\n"
+	"subs	r3, r3, #1		@ r3 = w -= 1		\n"
+	"ble	2f			@ while (w > 0) {	\n"
+	"mov	r12,r9			@ r12= l = ht_len	\n"
+	"b	1f			@			\n"
+	"9:				@			\n"
+	"strb	r14,[r2], #1		@ *out++ = h		\n"
+	"subs	r12,r12,#2		@ r12 = l -= 2		\n"
+	"moveq	r12,r9			@ if(l==0) l = ht_len	\n"
+	"subeq	r0, r0, r9, LSL #2	@          ht_line -= l	\n"
+	"subs	r3, r3, #2		@ w -= 2		\n"
+	"beq	2f			@ }			\n"
+	"blt	3f			@			\n"
+	"1:							\n"
+	"ldr	r5, [r1], #5		@ r5 = pixmap[0..3]	\n"
+	"ldr	r7, [r1], #5		@ r7 = pixmap[5..8]	\n"
+	"add	r0, r0, #8		@ r0 = ht_line += 8	\n"
+	"mov	r14,#0			@ r14= h = 0		\n"
+	"orrs	r5, r5, r7		@ if (r5 | r7 == 0)	\n"
+	"beq	9b			@	white		\n"
+	"ldrb	r4, [r0, #-8]		@ r4 = ht_line[0]	\n"
+	"ldrb	r5, [r1, #-10]		@ r5 = pixmap[0]	\n"
+	"ldrb	r6, [r0, #-7]		@ r6 = ht_line[1]	\n"
+	"ldrb	r7, [r1, #-9]		@ r7 = pixmap[1]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x80		@	h |= 0x80	\n"
+	"ldrb	r4, [r0, #-6]		@ r4 = ht_line[2]	\n"
+	"ldrb	r5, [r1, #-8]		@ r5 = pixmap[2]	\n"
+	"cmp	r6, r7			@ if (r6 < r7)		\n"
+	"orrle	r14,r14,#0x40		@	h |= 0x40	\n"
+	"ldrb	r6, [r0, #-5]		@ r6 = ht_line[3]	\n"
+	"ldrb	r7, [r1, #-7]		@ r7 = pixmap[3]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x20		@	h |= 0x20	\n"
+	"ldrb	r4, [r0, #-4]		@ r4 = ht_line[4]	\n"
+	"ldrb	r5, [r1, #-5]		@ r5 = pixmap[5]	\n"
+	"cmp	r6, r7			@ if (r6 < r7)		\n"
+	"orrle	r14,r14,#0x10		@	h |= 0x10	\n"
+	"ldrb	r6, [r0, #-3]		@ r6 = ht_line[5]	\n"
+	"ldrb	r7, [r1, #-4]		@ r7 = pixmap[6]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x08		@	h |= 0x08	\n"
+	"ldrb	r4, [r0, #-2]		@ r4 = ht_line[6]	\n"
+	"ldrb	r5, [r1, #-3]		@ r5 = pixmap[7]	\n"
+	"cmp	r6, r7			@ if (r6 < r7)		\n"
+	"orrle	r14,r14,#0x04		@	h |= 0x04	\n"
+	"ldrb	r6, [r0, #-1]		@ r6 = ht_line[7]	\n"
+	"ldrb	r7, [r1, #-2]		@ r7 = pixmap[8]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x02		@	h |= 0x02	\n"
+	"cmp	r6, r7			@ if (r7 < r6)		\n"
+	"orrle	r14,r14,#0x01		@	h |= 0x01	\n"
+	"subs	r12,r12,#2		@ r12 = l -= 2		\n"
+	"strb	r14,[r2], #1		@ *out++ = h		\n"
+	"moveq	r12,r9			@ if(l==0) l = ht_len	\n"
+	"subeq	r0, r0, r9, LSL #2	@          ht_line -= l	\n"
+	"subs	r3, r3, #2		@ w -= 2		\n"
+	"bgt	1b			@ }			\n"
+	"blt	3f			@			\n"
+	"2:							\n"
+	"ldrb	r4, [r0], #1		@ r4 = ht_line[0]	\n"
+	"ldrb	r5, [r1], #1		@ r5 = pixmap[0]	\n"
+	"mov	r14, #0			@ r14= h = 0		\n"
+	"ldrb	r6, [r0], #1		@ r6 = ht_line[1]	\n"
+	"ldrb	r7, [r1], #1		@ r7 = pixmap[1]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x80		@	h |= 0x80	\n"
+	"ldrb	r4, [r0], #1		@ r6 = ht_line[2]	\n"
+	"ldrb	r5, [r1], #1		@ r7 = pixmap[2]	\n"
+	"cmp	r6, r7			@ if (r6 < r7)		\n"
+	"orrle	r14,r14,#0x40		@	h |= 0x40	\n"
+	"ldrb	r6, [r0], #1		@ r6 = ht_line[1]	\n"
+	"ldrb	r7, [r1], #2		@ r7 = pixmap[2]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x20		@	h |= 0x20	\n"
+	"cmp	r6, r7			@ if (r6 < r7)		\n"
+	"orrle	r14,r14,#0x10		@	h |= 0x10	\n"
+	"strb	r14,[r2]		@ *out = h		\n"
+	"3:							\n"
+	"ldmfd	r13!,{r4-r7,r9,PC}	@ pop, return to thumb	\n"
+	ENTER_THUMB
+	);
+}
+#else
+/* Vanilla version, should work on all ARMs */
+static void
+do_threshold_4(const unsigned char * restrict ht_line, const unsigned char * restrict pixmap, unsigned char *restrict out, int w, int ht_len)
+{
+	asm volatile(
+	ENTER_ARM
+	// Store one more reg that required to keep double stack alignment
+	"stmfd	r13!,{r4-r7,r9,r14}				\n"
+	"@ r0 = ht_line						\n"
+	"@ r1 = pixmap						\n"
+	"@ r2 = out						\n"
+	"@ r3 = w						\n"
+	"@ <> = ht_len						\n"
+	"ldr	r9, [r13,#6*4]		@ r9 = ht_len		\n"
+	"subs	r3, r3, #1		@ r3 = w -= 1		\n"
+	"ble	2f			@ while (w > 0) {	\n"
+	"mov	r12,r9			@ r12= l = ht_len	\n"
+	"1:							\n"
+	"mov	r14,#0			@ r14= h = 0		\n"
+	"ldrb	r4, [r0], #1		@ r4 = ht_line[0]	\n"
+	"ldrb	r5, [r1], #1		@ r5 = pixmap[0]	\n"
+	"ldrb	r6, [r0], #1		@ r6 = ht_line[1]	\n"
+	"ldrb	r7, [r1], #1		@ r7 = pixmap[1]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x80		@	h |= 0x80	\n"
+	"ldrb	r4, [r0], #1		@ r4 = ht_line[2]	\n"
+	"ldrb	r5, [r1], #1		@ r5 = pixmap[2]	\n"
+	"cmp	r6, r7			@ if (r6 < r7)		\n"
+	"orrle	r14,r14,#0x40		@	h |= 0x40	\n"
+	"ldrb	r6, [r0], #1		@ r6 = ht_line[3]	\n"
+	"ldrb	r7, [r1], #2		@ r7 = pixmap[3]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x20		@	h |= 0x20	\n"
+	"ldrb	r4, [r0], #1		@ r4 = ht_line[4]	\n"
+	"ldrb	r5, [r1], #1		@ r5 = pixmap[4]	\n"
+	"cmp	r6, r7			@ if (r6 < r7)		\n"
+	"orrle	r14,r14,#0x10		@	h |= 0x10	\n"
+	"ldrb	r6, [r0], #1		@ r6 = ht_line[5]	\n"
+	"ldrb	r7, [r1], #1		@ r7 = pixmap[6]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x08		@	h |= 0x08	\n"
+	"ldrb	r4, [r0], #1		@ r4 = ht_line[6]	\n"
+	"ldrb	r5, [r1], #1		@ r5 = pixmap[7]	\n"
+	"cmp	r6, r7			@ if (r6 < r7)		\n"
+	"orrle	r14,r14,#0x04		@	h |= 0x04	\n"
+	"ldrb	r6, [r0], #1		@ r6 = ht_line[7]	\n"
+	"ldrb	r7, [r1], #2		@ r7 = pixmap[8]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x02		@	h |= 0x02	\n"
+	"cmp	r6, r7			@ if (r7 < r6)		\n"
+	"orrle	r14,r14,#0x01		@	h |= 0x01	\n"
+	"subs	r12,r12,#2		@ r12 = l -= 2		\n"
+	"strb	r14,[r2], #1		@ *out++ = h		\n"
+	"moveq	r12,r9			@ if(l==0) l = ht_len	\n"
+	"subeq	r0, r0, r9, LSL #2	@          ht_line -= l	\n"
+	"subs	r3, r3, #2		@ w -= 2		\n"
+	"bgt	1b			@ }			\n"
+	"blt	3f			@			\n"
+	"2:							\n"
+	"ldrb	r4, [r0], #1		@ r4 = ht_line[0]	\n"
+	"ldrb	r5, [r1], #1		@ r5 = pixmap[0]	\n"
+	"mov	r14, #0			@ r14= h = 0		\n"
+	"ldrb	r6, [r0], #1		@ r6 = ht_line[1]	\n"
+	"ldrb	r7, [r1], #1		@ r7 = pixmap[1]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x80		@	h |= 0x80	\n"
+	"ldrb	r4, [r0], #1		@ r6 = ht_line[2]	\n"
+	"ldrb	r5, [r1], #1		@ r7 = pixmap[2]	\n"
+	"cmp	r6, r7			@ if (r6 < r7)		\n"
+	"orrle	r14,r14,#0x40		@	h |= 0x40	\n"
+	"ldrb	r6, [r0], #1		@ r6 = ht_line[1]	\n"
+	"ldrb	r7, [r1], #2		@ r7 = pixmap[2]	\n"
+	"cmp	r4, r5			@ if (r4 < r5)		\n"
+	"orrle	r14,r14,#0x20		@	h |= 0x20	\n"
+	"cmp	r6, r7			@ if (r6 < r7)		\n"
+	"orrle	r14,r14,#0x10		@	h |= 0x10	\n"
+	"strb	r14,[r2]		@ *out = h		\n"
+	"3:							\n"
+	"ldmfd	r13!,{r4-r7,r9,PC}	@ pop, return to thumb	\n"
+	ENTER_THUMB
+	);
+}
+#endif /* UNALIGNED */
+#else
 static void do_threshold_4(const unsigned char * restrict ht_line, const unsigned char * restrict pixmap, unsigned char * restrict out, int w, int ht_len)
 {
 	int l = ht_len;
@@ -374,6 +574,7 @@ static void do_threshold_4(const unsigned char * restrict ht_line, const unsigne
 		*out = h;
 	}
 }
+#endif
 
 fz_bitmap *fz_new_bitmap_from_pixmap(fz_context *ctx, fz_pixmap *pix, fz_halftone *ht)
 {