xtensa: rearrange unaligned exception handler

- extract initialization part of the exception handler into a separate
  function.
- use single label for invalid instruction instead of two labels, one
  for load and one for store, at one place.
- use sext instruction for sign extension when available.
- store SAR on the stack instead of in a0.
- replace numeric labels for load and store writeback with .Lload_w and
  .Lstore_w respectively.

Signed-off-by: Max Filippov <[email protected]>

Author: jcmvbkbc

arch/xtensa/kernel/align.S

@@ -157,58 +157,7 @@
 	.literal_position
 ENTRY(fast_unaligned)
 
-	/* Note: We don't expect the address to be aligned on a word
-	 *       boundary. After all, the processor generated that exception
-	 *       and it would be a hardware fault.
-	 */
-
-	/* Save some working register */
-
-	s32i	a4, a2, PT_AREG4
-	s32i	a5, a2, PT_AREG5
-	s32i	a6, a2, PT_AREG6
-	s32i	a7, a2, PT_AREG7
-	s32i	a8, a2, PT_AREG8
-
-	rsr	a0, depc
-	s32i	a0, a2, PT_AREG2
-	s32i	a3, a2, PT_AREG3
-
-	rsr	a3, excsave1
-	movi	a4, fast_unaligned_fixup
-	s32i	a4, a3, EXC_TABLE_FIXUP
-
-	/* Keep value of SAR in a0 */
-
-	rsr	a0, sar
-	rsr	a8, excvaddr		# load unaligned memory address
-
-	/* Now, identify one of the following load/store instructions.
-	 *
-	 * The only possible danger of a double exception on the
-	 * following l32i instructions is kernel code in vmalloc
-	 * memory. The processor was just executing at the EPC_1
-	 * address, and indeed, already fetched the instruction.  That
-	 * guarantees a TLB mapping, which hasn't been replaced by
-	 * this unaligned exception handler that uses only static TLB
-	 * mappings. However, high-level interrupt handlers might
-	 * modify TLB entries, so for the generic case, we register a
-	 * TABLE_FIXUP handler here, too.
-	 */
-
-	/* a3...a6 saved on stack, a2 = SP */
-
-	/* Extract the instruction that caused the unaligned access. */
-
-	rsr	a7, epc1	# load exception address
-	movi	a3, ~3
-	and	a3, a3, a7	# mask lower bits
-
-	l32i	a4, a3, 0	# load 2 words
-	l32i	a5, a3, 4
-
-	__ssa8	a7
-	__src_b	a4, a4, a5	# a4 has the instruction
+	call0	.Lsave_and_load_instruction
 
 	/* Analyze the instruction (load or store?). */
 
@@ -249,29 +198,32 @@ ENTRY(fast_unaligned)
 	addi	a7, a7, 2		# increment PC (assume 16-bit insn)
 
 	extui	a5, a4, INSN_OP0, 4
-	_beqi	a5, OP0_L32I_N, 1f	# l32i.n: jump
+	_beqi	a5, OP0_L32I_N, .Lload_w# l32i.n: jump
 
 	addi	a7, a7, 1
 #else
 	addi	a7, a7, 3
 #endif
 
 	extui	a5, a4, INSN_OP1, 4
-	_beqi	a5, OP1_L32I, 1f	# l32i: jump
+	_beqi	a5, OP1_L32I, .Lload_w	# l32i: jump
 
 	extui	a3, a3, 0, 16		# extract lower 16 bits
-	_beqi	a5, OP1_L16UI, 1f
+	_beqi	a5, OP1_L16UI, .Lload_w
 	addi	a5, a5, -OP1_L16SI
-	_bnez	a5, .Linvalid_instruction_load
+	_bnez	a5, .Linvalid_instruction
 
 	/* sign extend value */
-
+#if XCHAL_HAVE_SEXT
+	sext	a3, a3, 15
+#else
 	slli	a3, a3, 16
 	srai	a3, a3, 16
+#endif
 
 	/* Set target register. */
 
-1:
+.Lload_w:
 	extui	a4, a4, INSN_T, 4	# extract target register
 	movi	a5, .Lload_table
 	addx8	a4, a4, a5
@@ -297,35 +249,35 @@ ENTRY(fast_unaligned)
 	mov	a15, a3		;	_j .Lexit;	.align 8
 
 .Lstore_table:
-	l32i	a3, a2, PT_AREG0;	_j 1f;	.align 8
-	mov	a3, a1;			_j 1f;	.align 8	# fishy??
-	l32i	a3, a2, PT_AREG2;	_j 1f;	.align 8
-	l32i	a3, a2, PT_AREG3;	_j 1f;	.align 8
-	l32i	a3, a2, PT_AREG4;	_j 1f;	.align 8
-	l32i	a3, a2, PT_AREG5;	_j 1f;	.align 8
-	l32i	a3, a2, PT_AREG6;	_j 1f;	.align 8
-	l32i	a3, a2, PT_AREG7;	_j 1f;	.align 8
-	l32i	a3, a2, PT_AREG8;	_j 1f;	.align 8
-	mov	a3, a9		;	_j 1f;	.align 8
-	mov	a3, a10		;	_j 1f;	.align 8
-	mov	a3, a11		;	_j 1f;	.align 8
-	mov	a3, a12		;	_j 1f;	.align 8
-	mov	a3, a13		;	_j 1f;	.align 8
-	mov	a3, a14		;	_j 1f;	.align 8
-	mov	a3, a15		;	_j 1f;	.align 8
+	l32i	a3, a2, PT_AREG0;	_j .Lstore_w;	.align 8
+	mov	a3, a1;			_j .Lstore_w;	.align 8	# fishy??
+	l32i	a3, a2, PT_AREG2;	_j .Lstore_w;	.align 8
+	l32i	a3, a2, PT_AREG3;	_j .Lstore_w;	.align 8
+	l32i	a3, a2, PT_AREG4;	_j .Lstore_w;	.align 8
+	l32i	a3, a2, PT_AREG5;	_j .Lstore_w;	.align 8
+	l32i	a3, a2, PT_AREG6;	_j .Lstore_w;	.align 8
+	l32i	a3, a2, PT_AREG7;	_j .Lstore_w;	.align 8
+	l32i	a3, a2, PT_AREG8;	_j .Lstore_w;	.align 8
+	mov	a3, a9		;	_j .Lstore_w;	.align 8
+	mov	a3, a10		;	_j .Lstore_w;	.align 8
+	mov	a3, a11		;	_j .Lstore_w;	.align 8
+	mov	a3, a12		;	_j .Lstore_w;	.align 8
+	mov	a3, a13		;	_j .Lstore_w;	.align 8
+	mov	a3, a14		;	_j .Lstore_w;	.align 8
+	mov	a3, a15		;	_j .Lstore_w;	.align 8
 
 	/* We cannot handle this exception. */
 
 	.extern _kernel_exception
-.Linvalid_instruction_load:
-.Linvalid_instruction_store:
+.Linvalid_instruction:
 
 	movi	a4, 0
 	rsr	a3, excsave1
 	s32i	a4, a3, EXC_TABLE_FIXUP
 
 	/* Restore a4...a8 and SAR, set SP, and jump to default exception. */
 
+	l32i	a0, a2, PT_SAR
 	l32i	a8, a2, PT_AREG8
 	l32i	a7, a2, PT_AREG7
 	l32i	a6, a2, PT_AREG6
@@ -343,8 +295,8 @@ ENTRY(fast_unaligned)
 2:	movi	a0, _user_exception
 	jx	a0
 
-1: 	# a7: instruction pointer, a4: instruction, a3: value
-
+	# a7: instruction pointer, a4: instruction, a3: value
+.Lstore_w:
 	movi	a6, 0			# mask: ffffffff:00000000
 
 #if XCHAL_HAVE_DENSITY
@@ -361,7 +313,7 @@ ENTRY(fast_unaligned)
 
 	extui	a5, a4, INSN_OP1, 4	# extract OP1
 	_beqi	a5, OP1_S32I, 1f	# jump if 32 bit store
-	_bnei	a5, OP1_S16I, .Linvalid_instruction_store
+	_bnei	a5, OP1_S16I, .Linvalid_instruction
 
 	movi	a5, -1
 	__extl	a3, a3			# get 16-bit value
@@ -434,6 +386,7 @@ ENTRY(fast_unaligned)
 
 	/* Restore working register */
 
+	l32i	a0, a2, PT_SAR
 	l32i	a8, a2, PT_AREG8
 	l32i	a7, a2, PT_AREG7
 	l32i	a6, a2, PT_AREG6
@@ -448,6 +401,59 @@ ENTRY(fast_unaligned)
 	l32i	a2, a2, PT_AREG2
 	rfe
 
+	.align	4
+.Lsave_and_load_instruction:
+
+	/* Save some working register */
+
+	s32i	a3, a2, PT_AREG3
+	s32i	a4, a2, PT_AREG4
+	s32i	a5, a2, PT_AREG5
+	s32i	a6, a2, PT_AREG6
+	s32i	a7, a2, PT_AREG7
+	s32i	a8, a2, PT_AREG8
+
+	rsr	a4, depc
+	s32i	a4, a2, PT_AREG2
+
+	rsr	a5, sar
+	s32i	a5, a2, PT_SAR
+
+	rsr	a3, excsave1
+	movi	a4, fast_unaligned_fixup
+	s32i	a4, a3, EXC_TABLE_FIXUP
+
+	rsr	a8, excvaddr		# load unaligned memory address
+
+	/* Now, identify one of the following load/store instructions.
+	 *
+	 * The only possible danger of a double exception on the
+	 * following l32i instructions is kernel code in vmalloc
+	 * memory. The processor was just executing at the EPC_1
+	 * address, and indeed, already fetched the instruction.  That
+	 * guarantees a TLB mapping, which hasn't been replaced by
+	 * this unaligned exception handler that uses only static TLB
+	 * mappings. However, high-level interrupt handlers might
+	 * modify TLB entries, so for the generic case, we register a
+	 * TABLE_FIXUP handler here, too.
+	 */
+
+	/* a3...a6 saved on stack, a2 = SP */
+
+	/* Extract the instruction that caused the unaligned access. */
+
+	rsr	a7, epc1	# load exception address
+	movi	a3, ~3
+	and	a3, a3, a7	# mask lower bits
+
+	l32i	a4, a3, 0	# load 2 words
+	l32i	a5, a3, 4
+
+	__ssa8	a7
+	__src_b	a4, a4, a5	# a4 has the instruction
+
+	ret
+
 ENDPROC(fast_unaligned)
 
 ENTRY(fast_unaligned_fixup)
@@ -459,10 +465,11 @@ ENTRY(fast_unaligned_fixup)
 	l32i	a7, a2, PT_AREG7
 	l32i	a6, a2, PT_AREG6
 	l32i	a5, a2, PT_AREG5
-	l32i	a4, a2, PT_AREG4
+	l32i	a4, a2, PT_SAR
 	l32i	a0, a2, PT_AREG2
-	xsr	a0, depc			# restore depc and a0
-	wsr	a0, sar
+	wsr	a4, sar
+	wsr	a0, depc			# restore depc and a0
+	l32i	a4, a2, PT_AREG4
 
 	rsr	a0, exccause
 	s32i	a0, a2, PT_DEPC			# mark as a regular exception