Index: contrib/jemalloc/include/jemalloc/jemalloc_FreeBSD.h
===================================================================
--- contrib/jemalloc/include/jemalloc/jemalloc_FreeBSD.h
+++ contrib/jemalloc/include/jemalloc/jemalloc_FreeBSD.h
@@ -46,6 +46,7 @@
 #else
 #  define LG_SIZEOF_PTR		2
 #endif
+#  define JEMALLOC_TLS_MODEL	__attribute__((tls_model("initial-exec")))
 #endif
 #ifdef __powerpc64__
 #  define LG_SIZEOF_PTR		3
Index: lib/libc/mips/gen/_set_tp.c
===================================================================
--- lib/libc/mips/gen/_set_tp.c
+++ lib/libc/mips/gen/_set_tp.c
@@ -30,11 +30,12 @@
 #include <stdint.h>
 
 #include <machine/sysarch.h>
+#include <machine/tls.h>
 #include "libc_private.h"
 
 void
 _set_tp(void *tp)
 {
 
-	sysarch(MIPS_SET_TLS, tp);
+	sysarch(MIPS_SET_TLS, (char *)tp + TLS_TP_OFFSET + TLS_TCB_SIZE);
 }
Index: lib/libthr/arch/mips/include/pthread_md.h
===================================================================
--- lib/libthr/arch/mips/include/pthread_md.h
+++ lib/libthr/arch/mips/include/pthread_md.h
@@ -40,6 +40,7 @@
 
 #define	CPU_SPINWAIT
 #define	DTV_OFFSET		offsetof(struct tcb, tcb_dtv)
+#define	TP_OFFSET		(TLS_TP_OFFSET + TLS_TCB_SIZE)
 
 /*
  * Variant I tcb. The structure layout is fixed, don't blindly
@@ -55,7 +56,7 @@
 _tcb_set(struct tcb *tcb)
 {
 
-	sysarch(MIPS_SET_TLS, tcb);
+	sysarch(MIPS_SET_TLS, (char *)tcb + TP_OFFSET);
 }
 
 /*
@@ -64,10 +65,10 @@
 static __inline struct tcb *
 _tcb_get(void)
 {
-	struct tcb *tcb;
+	char *tcb;
 
 	sysarch(MIPS_GET_TLS, &tcb);
-	return tcb;
+	return (struct tcb *)(tcb - TP_OFFSET);
 }
 
 extern struct pthread *_thr_initial;
Index: libexec/rtld-elf/mips/reloc.c
===================================================================
--- libexec/rtld-elf/mips/reloc.c
+++ libexec/rtld-elf/mips/reloc.c
@@ -621,7 +621,7 @@
 allocate_initial_tls(Obj_Entry *objs)
 {
 	char *tls;
-	
+
 	/*
 	 * Fix the size of the static TLS block by using the maximum
 	 * offset allocated so far and adding a bit for dynamic modules to
@@ -629,7 +629,8 @@
 	 */
 	tls_static_space = tls_last_offset + tls_last_size + RTLD_STATIC_TLS_EXTRA;
 
-	tls = (char *) allocate_tls(objs, NULL, TLS_TCB_SIZE, 8);
+	tls = (char *)allocate_tls(objs, NULL, TLS_TCB_SIZE, 8)
+	    + TLS_TP_OFFSET + TLS_TCB_SIZE;
 
 	sysarch(MIPS_SET_TLS, tls);
 }
@@ -637,12 +638,13 @@
 void *
 __tls_get_addr(tls_index* ti)
 {
-	Elf_Addr** tls;
+	Elf_Addr** tp;
 	char *p;
 
-	sysarch(MIPS_GET_TLS, &tls);
+	sysarch(MIPS_GET_TLS, &tp);
 
-	p = tls_get_addr_common(tls, ti->ti_module, ti->ti_offset + TLS_DTP_OFFSET);
+	p = tls_get_addr_common((Elf_Addr**)((Elf_Addr)tp - TLS_TP_OFFSET
+	    - TLS_TCB_SIZE), ti->ti_module, ti->ti_offset);
 
-	return (p);
+	return (p + TLS_DTP_OFFSET);
 }
Index: sys/mips/mips/trap.c
===================================================================
--- sys/mips/mips/trap.c
+++ sys/mips/mips/trap.c
@@ -899,12 +899,6 @@
 					if (inst.RType.rd == 29) {
 						frame_regs = &(trapframe->zero);
 						frame_regs[inst.RType.rt] = (register_t)(intptr_t)td->td_md.md_tls;
-#if defined(__mips_n64) && defined(COMPAT_FREEBSD32)
-						if (SV_PROC_FLAG(td->td_proc, SV_ILP32))
-							frame_regs[inst.RType.rt] += TLS_TP_OFFSET + TLS_TCB_SIZE32;
-						else
-#endif
-						frame_regs[inst.RType.rt] += TLS_TP_OFFSET + TLS_TCB_SIZE;
 						trapframe->pc += sizeof(int);
 						goto out;
 					}
Index: sys/mips/mips/vm_machdep.c
===================================================================
--- sys/mips/mips/vm_machdep.c
+++ sys/mips/mips/vm_machdep.c
@@ -492,7 +492,7 @@
 cpu_set_user_tls(struct thread *td, void *tls_base)
 {
 
-	td->td_md.md_tls = (char*)tls_base;
+	td->td_md.md_tls = (char*)tls_base + TLS_TP_OFFSET + TLS_TCP_SIZE;
 
 	return (0);
 }