xtensa: implement call0 ABI support in assembly

Replace hardcoded register and opcode names with ABI-agnostic macros.
Add register save/restore code where necessary. Conditionalize windowed
only or call0 only code. Add stack initialization matching _switch_to
epilogue to copy_thread.
Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>

arch/xtensa/boot/boot-redboot/bootstrap.S

@@ -3,6 +3,7 @@
 #include <asm/regs.h>
 #include <asm/asmmacro.h>
 #include <asm/cacheasm.h>
+#include <asm/processor.h>
 	/*
 	 * RB-Data: RedBoot data/bss
 	 * P:	    Boot-Parameters
@@ -36,7 +37,7 @@
 	.globl __start
 	/* this must be the first byte of the loader! */
 __start:
-	entry	sp, 32		# we do not intend to return
+	abi_entry(32)		# we do not intend to return
 	_call0	_start
 __start_a0:
 	.align 4
@@ -62,10 +63,12 @@ _start:
 	wsr	a4, windowstart
 	rsync
 
-	movi	a4, 0x00040000
+	movi	a4, KERNEL_PS_WOE_MASK
 	wsr	a4, ps
 	rsync
 
+KABI_C0	mov	abi_saved0, abi_arg0
+
 	/* copy the loader to its address
 	 * Note: The loader itself is a very small piece, so we assume we
 	 *       don't partially overlap. We also assume (even more important)
@@ -168,52 +171,52 @@ _reloc:
 
 	movi	a3, __image_load
 	sub	a4, a3, a4
-	add	a8, a0, a4
+	add	abi_arg2, a0, a4
 
 	# a1  Stack
 	# a8(a4)  Load address of the image
 
-	movi	a6, _image_start
-	movi	a10, _image_end
-	movi	a7, 0x1000000
-	sub	a11, a10, a6
-	movi	a9, complen
-	s32i	a11, a9, 0
+	movi	abi_arg0, _image_start
+	movi	abi_arg4, _image_end
+	movi	abi_arg1, 0x1000000
+	sub	abi_tmp0, abi_arg4, abi_arg0
+	movi	abi_arg3, complen
+	s32i	abi_tmp0, abi_arg3, 0
 
 	movi	a0, 0
 
-	# a6 destination
-	# a7 maximum size of destination
-	# a8 source
-	# a9 ptr to length
+	# abi_arg0 destination
+	# abi_arg1 maximum size of destination
+	# abi_arg2 source
+	# abi_arg3 ptr to length
 
 	.extern gunzip
-	movi	a4, gunzip
-	beqz	a4, 1f
+	movi	abi_tmp0, gunzip
+	beqz	abi_tmp0, 1f
 
-	callx4	a4
+	abi_callx	abi_tmp0
 
 	j	2f
 
 
-	# a6 destination start
-	# a7 maximum size of destination
-	# a8 source start
-	# a9 ptr to length
-	# a10 destination end
+	# abi_arg0 destination start
+	# abi_arg1 maximum size of destination
+	# abi_arg2 source start
+	# abi_arg3 ptr to length
+	# abi_arg4 destination end
 
 1:
-        l32i    a9, a8, 0
-        l32i    a11, a8, 4
-        s32i    a9, a6, 0
-        s32i    a11, a6, 4
-        l32i    a9, a8, 8
-        l32i    a11, a8, 12
-        s32i    a9, a6, 8
-        s32i    a11, a6, 12
-        addi    a6, a6, 16
-        addi    a8, a8, 16
-        blt     a6, a10, 1b
+        l32i    abi_tmp0, abi_arg2, 0
+        l32i    abi_tmp1, abi_arg2, 4
+        s32i    abi_tmp0, abi_arg0, 0
+        s32i    abi_tmp1, abi_arg0, 4
+        l32i    abi_tmp0, abi_arg2, 8
+        l32i    abi_tmp1, abi_arg2, 12
+        s32i    abi_tmp0, abi_arg0, 8
+        s32i    abi_tmp1, abi_arg0, 12
+        addi    abi_arg0, abi_arg0, 16
+        addi    abi_arg2, abi_arg2, 16
+        blt     abi_arg0, abi_arg4, 1b
 
 
 	/* jump to the kernel */
@@ -230,6 +233,7 @@ _reloc:
 
 	# a2  Boot parameter list
 
+KABI_C0	mov	abi_arg0, abi_saved0
 	movi	a0, _image_start
 	jx	a0
 

arch/xtensa/kernel/head.S

@@ -15,6 +15,7 @@
  * Kevin Chea
  */
 
+#include <asm/asmmacro.h>
 #include <asm/processor.h>
 #include <asm/page.h>
 #include <asm/cacheasm.h>
@@ -193,9 +194,10 @@ ENTRY(_startup)
 	movi	a1, start_info
 	l32i	a1, a1, 0
 
-	movi	a2, PS_WOE_MASK | LOCKLEVEL
-					# WOE=1, INTLEVEL=LOCKLEVEL, UM=0
-	wsr	a2, ps			# (enable reg-windows; progmode stack)
+	/* Disable interrupts. */
+	/* Enable window exceptions if kernel is built with windowed ABI. */
+	movi	a2, KERNEL_PS_WOE_MASK | LOCKLEVEL
+	wsr	a2, ps
 	rsync
 
 #ifdef CONFIG_SMP
@@ -267,13 +269,13 @@ ENTRY(_startup)
 	l32i	a1, a1, 0
 #endif
 
-	movi	a6, 0
-	xsr	a6, excsave1
+	movi	abi_arg0, 0
+	xsr	abi_arg0, excsave1
 
 	/* init_arch kick-starts the linux kernel */
 
-	call4	init_arch
-	call4	start_kernel
+	abi_call	init_arch
+	abi_call	start_kernel
 
 should_never_return:
 	j	should_never_return
@@ -297,10 +299,10 @@ should_never_return:
 	s32i	a3, a2, 0
 	memw
 
-	movi	a6, 0
-	wsr	a6, excsave1
+	movi	abi_arg0, 0
+	wsr	abi_arg0, excsave1
 
-	call4	secondary_start_kernel
+	abi_call	secondary_start_kernel
 	j	should_never_return
 
 #endif  /* CONFIG_SMP */

arch/xtensa/kernel/mcount.S

@@ -17,11 +17,16 @@
 /*
  * Entry condition:
  *
- *   a2:	a0 of the caller
+ *   a2:	a0 of the caller in windowed ABI
+ *   a10:	a0 of the caller in call0 ABI
+ *
+ * In call0 ABI the function _mcount is called with the special ABI:
+ * its argument is in a10 and all the usual argument registers (a2 - a7)
+ * must be preserved in addition to callee-saved a12 - a15.
  */
 
 ENTRY(_mcount)
-
+#if defined(__XTENSA_WINDOWED_ABI__)
 	abi_entry_default
 
 	movi	a4, ftrace_trace_function
@@ -42,7 +47,36 @@ ENTRY(_mcount)
 	callx4	a4
 
 	abi_ret_default
+#elif defined(__XTENSA_CALL0_ABI__)
+	abi_entry_default
+
+	movi	a9, ftrace_trace_function
+	l32i	a9, a9, 0
+	movi	a11, ftrace_stub
+	bne	a9, a11, 1f
+	abi_ret_default
 
+1:	abi_entry(28)
+	s32i	a0, sp, 0
+	s32i	a2, sp, 4
+	s32i	a3, sp, 8
+	s32i	a4, sp, 12
+	s32i	a5, sp, 16
+	s32i	a6, sp, 20
+	s32i	a7, sp, 24
+	addi	a2, a10, -MCOUNT_INSN_SIZE
+	callx0	a9
+	l32i	a0, sp, 0
+	l32i	a2, sp, 4
+	l32i	a3, sp, 8
+	l32i	a4, sp, 12
+	l32i	a5, sp, 16
+	l32i	a6, sp, 20
+	l32i	a7, sp, 24
+	abi_ret(28)
+#else
+#error Unsupported Xtensa ABI
+#endif
 ENDPROC(_mcount)
 
 ENTRY(ftrace_stub)

arch/xtensa/kernel/process.c

@@ -211,11 +211,18 @@ int copy_thread(unsigned long clone_flags, unsigned long usp_thread_fn,
 	struct thread_info *ti;
 #endif
 
+#if defined(__XTENSA_WINDOWED_ABI__)
 	/* Create a call4 dummy-frame: a0 = 0, a1 = childregs. */
 	SPILL_SLOT(childregs, 1) = (unsigned long)childregs;
 	SPILL_SLOT(childregs, 0) = 0;
 
 	p->thread.sp = (unsigned long)childregs;
+#elif defined(__XTENSA_CALL0_ABI__)
+	/* Reserve 16 bytes for the _switch_to stack frame. */
+	p->thread.sp = (unsigned long)childregs - 16;
+#else
+#error Unsupported Xtensa ABI
+#endif
 
 	if (!(p->flags & (PF_KTHREAD | PF_IO_WORKER))) {
 		struct pt_regs *regs = current_pt_regs();
@@ -272,11 +279,25 @@ int copy_thread(unsigned long clone_flags, unsigned long usp_thread_fn,
 		p->thread.ra = MAKE_RA_FOR_CALL(
 				(unsigned long)ret_from_kernel_thread, 1);
 
-		/* pass parameters to ret_from_kernel_thread:
-		 * a2 = thread_fn, a3 = thread_fn arg
+		/* pass parameters to ret_from_kernel_thread: */
+#if defined(__XTENSA_WINDOWED_ABI__)
+		/*
+		 * a2 = thread_fn, a3 = thread_fn arg.
+		 * Window underflow will load registers from the
+		 * spill slots on the stack on return from _switch_to.
 		 */
-		SPILL_SLOT(childregs, 3) = thread_fn_arg;
 		SPILL_SLOT(childregs, 2) = usp_thread_fn;
+		SPILL_SLOT(childregs, 3) = thread_fn_arg;
+#elif defined(__XTENSA_CALL0_ABI__)
+		/*
+		 * a12 = thread_fn, a13 = thread_fn arg.
+		 * _switch_to epilogue will load registers from the stack.
+		 */
+		((unsigned long *)p->thread.sp)[0] = usp_thread_fn;
+		((unsigned long *)p->thread.sp)[1] = thread_fn_arg;
+#else
+#error Unsupported Xtensa ABI
+#endif
 
 		/* Childregs are only used when we're going to userspace
 		 * in which case start_thread will set them up.