From cc2685013551ac1e2b3e08e3eef49bf8e4b728fc Mon Sep 17 00:00:00 2001 From: Michael Sevakis Date: Sat, 14 Jan 2012 16:16:05 -0500 Subject: Cut a few instructions from lcd_copy_buffer_rect. Change-Id: I5159534f90b041c6ffefc00c8f91abc68ca6eb42 --- firmware/target/arm/lcd-as-memframe.S | 30 +++++++++++++----------------- 1 file changed, 13 insertions(+), 17 deletions(-) diff --git a/firmware/target/arm/lcd-as-memframe.S b/firmware/target/arm/lcd-as-memframe.S index 163eacefe6..52ab0447c2 100644 --- a/firmware/target/arm/lcd-as-memframe.S +++ b/firmware/target/arm/lcd-as-memframe.S @@ -51,46 +51,42 @@ lcd_copy_buffer_rect: @ ble 40f @ finish line @ next line if lt or finish @ trailing halfword if eq 20: @ rem copy @ - add r14, r2, #1 @ get remaining width mod 16 after word + sub r14, r2, #1 @ wrap remaining width mod 16 after word @ align (rw) - and r14, r14, #0xe @ r14 = 0 (16), 2, 4, 6, 8, 10, 12, 14 + and r14, r14, #0xe @ r14 = 14, 0, 2, 4, 6, 8, 10, 12 add pc, pc, r14, lsl #3 @ branch to 32-byte align nop @ - b 30f @ rw % 16 = 0 or 1? use octword loop - nop @ - nop @ - nop @ - ldr r6, [r1], #4 @ rw % 16 = 2 or 3 + ldr r6, [r1], #4 @ rw % 16 = 2 or 3 (0) subs r2, r2, #2 @ str r6, [r0], #4 @ b 25f @ copy up done @ - ldmia r1!, { r6-r7 } @ rw % 16 = 4 or 5 + ldmia r1!, { r6-r7 } @ rw % 16 = 4 or 5 (2) subs r2, r2, #4 @ stmia r0!, { r6-r7 } @ b 25f @ copy up done @ - ldmia r1!, { r6-r8 } @ rw % 16 = 6 or 7 + ldmia r1!, { r6-r8 } @ rw % 16 = 6 or 7 (4) subs r2, r2, #6 @ stmia r0!, { r6-r8 } @ b 25f @ copy up done @ - ldmia r1!, { r6-r9 } @ rw % 16 = 8 or 9 + ldmia r1!, { r6-r9 } @ rw % 16 = 8 or 9 (6) subs r2, r2, #8 @ stmia r0!, { r6-r9 } @ b 25f @ copy up done @ - ldmia r1!, { r6-r10 } @ rw % 16 = 10 or 11 + ldmia r1!, { r6-r10 } @ rw % 16 = 10 or 11 (8) subs r2, r2, #10 @ stmia r0!, { r6-r10 } @ b 25f @ copy up done @ - ldmia r1!, { r6-r11 } @ rw % 16 = 12 or 13 + ldmia r1!, { r6-r11 } @ rw % 16 = 12 or 13 (10) subs r2, r2, #12 @ stmia r0!, { r6-r11 } @ b 25f @ copy up done @ - ldmia r1!, { r6-r12 } @ rw % 16 = 14 or 15 + ldmia r1!, { r6-r12 } @ rw % 16 = 14 or 15 (12) subs r2, r2, #14 @ stmia r0!, { r6-r12 } @ 25: @ copy up done @ - ble 40f @ finish line @ no 32-byte segments remaining? -30: @ octword loop @ copy 16 pixels per loop - ldmia r1!, { r6-r12, r14 } @ + ble 40f @ finish line @ +30: @ octword loop @ rw % 16 = 0 or 1 (14) + ldmia r1!, { r6-r12, r14 } @ copy 16 pixels per loop subs r2, r2, #16 @ stmia r0!, { r6-r12, r14 } @ bgt 30b @ octword loop @ @@ -102,9 +98,9 @@ lcd_copy_buffer_rect: @ subs r3, r3, #1 @ next line bgt 10b @ copy line @ ldmpc regs=r4-r11 @ restore regs and return - .ltorg @ dump constant pool .size lcd_copy_buffer_rect, .-lcd_copy_buffer_rect + /**************************************************************************** * void lcd_write_yuv420_lines(fb_data *dst, * unsigned char const * const src[3], -- cgit v1.2.3