--- zzzz-none-000/linux-3.10.107/arch/tile/kernel/head_64.S	2017-06-27 09:49:32.000000000 +0000
+++ scorpion-7490-727/linux-3.10.107/arch/tile/kernel/head_64.S	2021-02-04 17:41:59.000000000 +0000
@@ -25,6 +25,15 @@
 #include <arch/chip.h>
 #include <arch/spr_def.h>
 
+/* Extract two 32-bit bit values that were read into one register. */
+#ifdef __BIG_ENDIAN__
+#define GET_FIRST_INT(rd, rs) shrsi rd, rs, 32
+#define GET_SECOND_INT(rd, rs) addxi rd, rs, 0
+#else
+#define GET_FIRST_INT(rd, rs) addxi rd, rs, 0
+#define GET_SECOND_INT(rd, rs) shrsi rd, rs, 32
+#endif
+
 /*
  * This module contains the entry code for kernel images. It performs the
  * minimal setup needed to call the generic C routines.
@@ -46,11 +55,11 @@
 	  movei r2, TILE_CHIP_REV
 	  movei r3, KERNEL_PL
 	}
-	jal hv_init
+	jal _hv_init
 	/* Get a reasonable default ASID in r0 */
 	{
 	  move r0, zero
-	  jal hv_inquire_asid
+	  jal _hv_inquire_asid
 	}
 
 	/*
@@ -61,7 +70,7 @@
 	 * other CPUs should see a properly-constructed page table.
 	 */
 	{
-	  v4int_l r2, zero, r0    /* ASID for hv_install_context */
+	  GET_FIRST_INT(r2, r0)    /* ASID for hv_install_context */
 	  moveli r4, hw1_last(swapper_pgprot - PAGE_OFFSET)
 	}
 	{
@@ -77,7 +86,7 @@
 	{
 	  /* After initializing swapper_pgprot, HV_PTE_GLOBAL is set. */
 	  bfextu r7, r1, HV_PTE_INDEX_GLOBAL, HV_PTE_INDEX_GLOBAL
-	  inv r4
+	  finv r4
 	}
 	bnez r7, .Lno_write
 	{
@@ -121,29 +130,24 @@
 	}
 	{
 	  moveli r3, CTX_PAGE_FLAG
-	  j hv_install_context
+	  j _hv_install_context
 	}
 1:
 
 	/* Install the interrupt base. */
-	moveli r0, hw2_last(MEM_SV_START)
-	shl16insli r0, r0, hw1(MEM_SV_START)
-	shl16insli r0, r0, hw0(MEM_SV_START)
+	moveli r0, hw2_last(intrpt_start)
+	shl16insli r0, r0, hw1(intrpt_start)
+	shl16insli r0, r0, hw0(intrpt_start)
 	mtspr SPR_INTERRUPT_VECTOR_BASE_K, r0
 
-	/*
-	 * Get our processor number and save it away in SAVE_K_0.
-	 * Extract stuff from the topology structure: r4 = y, r6 = x,
-	 * r5 = width.  FIXME: consider whether we want to just make these
-	 * 64-bit values (and if so fix smp_topology write below, too).
-	 */
-	jal hv_inquire_topology
+	/* Get our processor number and save it away in SAVE_K_0. */
+	jal _hv_inquire_topology
 	{
-	  v4int_l r5, zero, r1    /* r5 = width */
-	  shrui r4, r0, 32        /* r4 = y */
+	  GET_FIRST_INT(r5, r1)   /* r5 = width */
+	  GET_SECOND_INT(r4, r0)  /* r4 = y */
 	}
 	{
-	  v4int_l r6, zero, r0    /* r6 = x */
+	  GET_FIRST_INT(r6, r0)   /* r6 = x */
 	  mul_lu_lu r4, r4, r5
 	}
 	{
@@ -154,7 +158,7 @@
 	/*
 	 * Load up our per-cpu offset.  When the first (master) tile
 	 * boots, this value is still zero, so we will load boot_pc
-	 * with start_kernel, and boot_sp with init_stack + THREAD_SIZE.
+	 * with start_kernel, and boot_sp with at the top of init_stack.
 	 * The master tile initializes the per-cpu offset array, so that
 	 * when subsequent (secondary) tiles boot, they will instead load
 	 * from their per-cpu versions of boot_sp and boot_pc.
@@ -198,9 +202,9 @@
 	}
 	ld r0, r0
 	ld sp, r1
-	or r4, sp, r4
+	shli r4, r4, CPU_SHIFT
+	bfins r4, sp, 0, CPU_SHIFT-1
 	mtspr SPR_SYSTEM_SAVE_K_0, r4  /* save ksp0 + cpu */
-	addi sp, sp, -STACK_TOP_DELTA
 	{
 	  move lr, zero   /* stop backtraces in the called function */
 	  jr r0