From 8f424d49e839f2a9bd0b4bba42919fc495e16aac Mon Sep 17 00:00:00 2001
From: fef <owo@fef.moe>
Date: Sun, 1 Aug 2021 01:16:13 +0200
Subject: [PATCH] asm: refactor to use thumb2 instructions

---
 arch/at91sam3x8e/irq_pend_sv.S | 42 +++++------------------------
 arch/at91sam3x8e/irq_svc.S     | 29 ++++----------------
 arch/at91sam3x8e/syscall.S     | 48 ++++++++++++----------------------
 3 files changed, 28 insertions(+), 91 deletions(-)

diff --git a/arch/at91sam3x8e/irq_pend_sv.S b/arch/at91sam3x8e/irq_pend_sv.S
index 9466664..ea618ab 100644
--- a/arch/at91sam3x8e/irq_pend_sv.S
+++ b/arch/at91sam3x8e/irq_pend_sv.S
@@ -1,24 +1,13 @@
 /* See the end of this file for copyright, license, and warranty information. */
 
-.syntax unified
-.thumb
+.include "asm.S"
 
 .text
 
 .extern sched_process_switch
 
-.thumb_func
-.global irq_pend_sv
-.type irq_pend_sv, %function
-
 /* void irq_pend_sv(void); */
-irq_pend_sv:
-	/*
-	 * There seems to be a limitation in Thumb which prohibits popping to
-	 * any register > r7.  Hence, we need to split the PUSHes and POPs into
-	 * multiple operations and MOV register values around.
-	 */
-
+func_begin irq_pend_sv
 	/*
 	 * Some registers have already been saved by hardware at this point,
 	 * we only need to take care of r4-r11 and lr (the latter of which is
@@ -26,17 +15,10 @@ irq_pend_sv:
 	 * The stuff we push onto the stack manually looks about like this:
 	 *
 	 * <<< stack grow direction (decreasing addresses) <<<
-	 * r8 r9 r10 r11 lr r4 r5 r6 r7
+	 * r4 r5 r6 r7 r8 r9 r10 r11 lr
 	 */
 
-	push	{r4-r7}			/* r4-r7 */
-
-	mov	r3,	r8
-	mov	r4,	r9
-	mov	r5,	r10
-	mov	r6,	r11
-	mov	r7,	lr
-	push	{r3-r7}			/* r8-r11, lr */
+	push	{r4-r11,lr}
 
 	/*
 	 * Now that our stack is completely saved, we can proceed to call the
@@ -54,23 +36,11 @@ irq_pend_sv:
 	 * just in reverse order.
 	 */
 
-	pop	{r3-r7}			/* r8-r11, lr */
-	mov	lr,	r7
-	mov	r11,	r6
-	mov	r10,	r5
-	mov	r9,	r4
-	mov	r8,	r3
+	pop	{r4-r11,lr}
 
-	pop	{r4-r7}			/* r4-r7 */
-
-	/*
-	 * ... and now get the fuck outta here, hoping I never need to touch
-	 * this again, even though deep inside I know I will return here at some
-	 * point because something broke (or I feel the need to micro-optimize).
-	 */
 	bx	lr
 
-.size irq_pend_sv, .-irq_pend_sv
+func_end irq_pend_sv
 
 /*
  * This file is part of Ardix.
diff --git a/arch/at91sam3x8e/irq_svc.S b/arch/at91sam3x8e/irq_svc.S
index cc318e7..53821ea 100644
--- a/arch/at91sam3x8e/irq_svc.S
+++ b/arch/at91sam3x8e/irq_svc.S
@@ -1,18 +1,13 @@
 /* See the end of this file for copyright, license, and warranty information. */
 
-.syntax unified
-.thumb
+.include "asm.S"
 
 .text
 
 .extern arch_enter
 
-.thumb_func
-.global irq_svc
-.type irq_svc, %function
-
 /* void irq_svc(void); */
-irq_svc:
+func_begin irq_svc
 	/*
 	 * Syscalls on Cortex-M use the following parameter calling convention:
 	 *
@@ -30,30 +25,16 @@ irq_svc:
 	 * like in irq_pend_sv, we save everything on the stack to make early
 	 * process switching possible in case the syscall is blocking.
 	 */
-	push	{r4-r7}			/* r4-r7 */
-
-	mov	r3,	r8
-	mov	r4,	r9
-	mov	r5,	r10
-	mov	r6,	r11
-	mov	r7,	lr
-	push	{r3-r7}			/* r8-r11, lr */
+	push	{r3-r11,lr}			/* r4-r7 */
 
 	mov	r0,	sp
 	bl	arch_enter		/* arch_enter(sp); */
 
-	pop	{r3-r7}			/* r8-r11, lr */
-	mov	lr,	r7
-	mov	r11,	r6
-	mov	r10,	r5
-	mov	r9,	r4
-	mov	r8,	r3
-
-	pop	{r4-r7}			/* r4-r7 */
+	pop	{r3-r11,lr}
 
 	bx	lr
 
-.size irq_svc, .-irq_svc
+func_end irq_svc
 
 /*
  * This file is part of Ardix.
diff --git a/arch/at91sam3x8e/syscall.S b/arch/at91sam3x8e/syscall.S
index 5942f82..b22b597 100644
--- a/arch/at91sam3x8e/syscall.S
+++ b/arch/at91sam3x8e/syscall.S
@@ -1,45 +1,31 @@
 /* See the end of this file for copyright, license, and warranty information. */
 
-.syntax unified
-.thumb
+.include "asm.S"
 
 /* this is only invoked from user space, obviously */
 .section .text.shared
 
-.thumb_func
-.global syscall
-.type syscall, %function
 /* int syscall(int number, int arg1, int arg2, int arg3, int arg4, int arg5, int arg6); */
-syscall:
-	push	{r4-r5,r7}
+func_begin syscall
 
 	/*
-	 * I HAVE NO IDEA WHY THE  F U C K  GCC THINKS I CAN'T MOV TWO LOW REGS
-	 * IN ONE INSTRUCTION WHEN THAT IS  L I T E R A L L Y  IN THE ASM OUTPUT
-	 * OF GCC WHEN COMPILING SOME C CODE.  BUT AT THIS POINT, I WON'T EVEN
-	 * BOTHER WITH WHATEVER SHITTY COMPILE OPTION OR PREPROCESSOR DIRECTIVE
-	 * I HAVE TO TURN ON IN ORDER FOR THIS PIECE OF SHIT THAT DARES CALLING
-	 * ITSELF A COMPILER TO EAT MY SHIT CODE AND SHIT OUT THE SHITTY BINARY
-	 * I WANT.  FUCK YOU, SYSCALLS WILL JUST TAKE 8 INSTRUCTIONS MORE.
+	 *   arg  | syscall |  AAPCS
+	 * -------|---------|---------
+	 * number | r7      | r0
+	 * arg1   | r0      | r1
+	 * arg2   | r1      | r2
+	 * arg3   | r2      | r3
+	 * arg4   | r3      | sp + 0
+	 * arg5   | r4      | sp + 4
+	 * arg6   | r5      | sp + 8
 	 */
 
-	mov	r4,	r12
-	push	{r4}
+	push	{r4-r5,r7}
 
-	mov	r12,	r0		/* syscall number */
-	mov	r7,	r12
-
-	mov	r12,	r1		/* arg1 */
-	mov	r0,	r12
-
-	mov	r12,	r2		/* arg2 */
-	mov	r1,	r12
-
-	mov	r12,	r3		/* arg3 */
-	mov	r2,	r12
-
-	pop	{r4}
-	mov	r12,	r4
+	mov	r7,	r0		/* number */
+	mov	r0,	r1		/* arg1 */
+	mov	r1,	r2		/* arg2 */
+	mov	r2,	r3		/* arg3 */
 
 	/* stack params begin at 12 bytes offset because we already pushed three registers */
 	ldr	r3,	[sp, #12]	/* arg4 */
@@ -53,7 +39,7 @@ syscall:
 	/* r0 (return value) is set by the kernel */
 	bx	lr
 
-.size syscall, .-syscall
+func_end syscall
 
 /*
  * This file is part of Ardix.