Index: libexec/rtld-elf/aarch64/rtld_machdep.h
===================================================================
--- libexec/rtld-elf/aarch64/rtld_machdep.h
+++ libexec/rtld-elf/aarch64/rtld_machdep.h
@@ -76,7 +76,6 @@
 	round(16, align)
 #define	calculate_tls_offset(prev_offset, prev_size, size, align, offset) \
 	round(prev_offset + prev_size, align)
-#define	calculate_tls_end(off, size) 	((off) + (size))
 #define calculate_tls_post_size(align) \
 	round(TLS_TCB_SIZE, align) - TLS_TCB_SIZE
 
Index: libexec/rtld-elf/amd64/reloc.c
===================================================================
--- libexec/rtld-elf/amd64/reloc.c
+++ libexec/rtld-elf/amd64/reloc.c
@@ -563,9 +563,3 @@
 {
 	return calculate_tls_offset(0, 0, size, align, offset);
 }
-
-size_t
-calculate_tls_end(size_t off, size_t size __unused)
-{
-	return (off);
-}
Index: libexec/rtld-elf/amd64/rtld_machdep.h
===================================================================
--- libexec/rtld-elf/amd64/rtld_machdep.h
+++ libexec/rtld-elf/amd64/rtld_machdep.h
@@ -73,10 +73,10 @@
 
 #define md_abi_variant_hook(x)
 
+#define	TLS_BELOW_TP
 #define	TLS_DTV_OFFSET	0
 
 size_t calculate_first_tls_offset(size_t size, size_t align, size_t offset);
 size_t calculate_tls_offset(size_t prev_offset, size_t prev_size, size_t size,
     size_t align, size_t offset);
-size_t calculate_tls_end(size_t off, size_t size);
 #endif
Index: libexec/rtld-elf/arm/rtld_machdep.h
===================================================================
--- libexec/rtld-elf/arm/rtld_machdep.h
+++ libexec/rtld-elf/arm/rtld_machdep.h
@@ -68,7 +68,6 @@
     round(8, align)
 #define calculate_tls_offset(prev_offset, prev_size, size, align, offset) \
     round(prev_offset + prev_size, align)
-#define calculate_tls_end(off, size)    ((off) + (size))
 #define calculate_tls_post_size(align) \
     round(TLS_TCB_SIZE, align) - TLS_TCB_SIZE
 	
Index: libexec/rtld-elf/i386/reloc.c
===================================================================
--- libexec/rtld-elf/i386/reloc.c
+++ libexec/rtld-elf/i386/reloc.c
@@ -554,9 +554,3 @@
 {
 	return calculate_tls_offset(0, 0, size, align, offset);
 }
-
-size_t
-calculate_tls_end(size_t off, size_t size __unused)
-{
-	return (off);
-}
Index: libexec/rtld-elf/i386/rtld_machdep.h
===================================================================
--- libexec/rtld-elf/i386/rtld_machdep.h
+++ libexec/rtld-elf/i386/rtld_machdep.h
@@ -74,10 +74,10 @@
 
 #define md_abi_variant_hook(x)
 
+#define	TLS_BELOW_TP
 #define	TLS_DTV_OFFSET	0
 
 size_t calculate_first_tls_offset(size_t size, size_t align, size_t offset);
 size_t calculate_tls_offset(size_t prev_offset, size_t prev_size, size_t size,
     size_t align, size_t offset);
-size_t calculate_tls_end(size_t off, size_t size);
 #endif
Index: libexec/rtld-elf/mips/rtld_machdep.h
===================================================================
--- libexec/rtld-elf/mips/rtld_machdep.h
+++ libexec/rtld-elf/mips/rtld_machdep.h
@@ -68,7 +68,6 @@
     TLS_TCB_SIZE
 #define calculate_tls_offset(prev_offset, prev_size, size, align, offset) \
     round(prev_offset + prev_size, align)
-#define calculate_tls_end(off, size)    ((off) + (size))
 #define calculate_tls_post_size(align)  0
 
 extern void *__tls_get_addr(tls_index *ti);
Index: libexec/rtld-elf/powerpc/rtld_machdep.h
===================================================================
--- libexec/rtld-elf/powerpc/rtld_machdep.h
+++ libexec/rtld-elf/powerpc/rtld_machdep.h
@@ -83,7 +83,6 @@
     TLS_TCB_SIZE
 #define calculate_tls_offset(prev_offset, prev_size, size, align, offset) \
     round(prev_offset + prev_size, align)
-#define calculate_tls_end(off, size)    ((off) + (size))
 #define calculate_tls_post_size(align)  0
  
 typedef struct {
Index: libexec/rtld-elf/powerpc64/rtld_machdep.h
===================================================================
--- libexec/rtld-elf/powerpc64/rtld_machdep.h
+++ libexec/rtld-elf/powerpc64/rtld_machdep.h
@@ -75,7 +75,6 @@
     TLS_TCB_SIZE
 #define calculate_tls_offset(prev_offset, prev_size, size, align, offset) \
     round(prev_offset + prev_size, align)
-#define calculate_tls_end(off, size)    ((off) + (size))
 #define calculate_tls_post_size(align)  0
  
 typedef struct {
Index: libexec/rtld-elf/riscv/rtld_machdep.h
===================================================================
--- libexec/rtld-elf/riscv/rtld_machdep.h
+++ libexec/rtld-elf/riscv/rtld_machdep.h
@@ -92,7 +92,6 @@
     TLS_TCB_SIZE
 #define calculate_tls_offset(prev_offset, prev_size, size, align, offset) \
     round(prev_offset + prev_size, align)
-#define calculate_tls_end(off, size)    ((off) + (size))
 #define calculate_tls_post_size(align)  0
 
 typedef struct {
Index: libexec/rtld-elf/rtld.c
===================================================================
--- libexec/rtld-elf/rtld.c
+++ libexec/rtld-elf/rtld.c
@@ -4939,8 +4939,7 @@
 	return (tls_get_addr_slow(dtvp, index, offset, false));
 }
 
-#if defined(__aarch64__) || defined(__arm__) || defined(__mips__) || \
-    defined(__powerpc__) || defined(__riscv)
+#ifndef TLS_BELOW_TP
 
 /*
  * Return pointer to allocated TLS block
@@ -5071,9 +5070,7 @@
     free_aligned(get_tls_block_ptr(tcb, tcbsize));
 }
 
-#endif
-
-#if defined(__i386__) || defined(__amd64__)
+#else
 
 /*
  * Allocate Static TLS using the Variant II method.
@@ -5227,6 +5224,11 @@
 	off = calculate_tls_offset(tls_last_offset, tls_last_size,
 	  obj->tlssize, obj->tlsalign, obj->tlspoffset);
 
+    obj->tlsoffset = off;
+#ifndef TLS_BELOW_TP
+    off += obj->tlssize;
+#endif
+
     /*
      * If we have already fixed the size of the static TLS block, we
      * must stay within that size. When allocating the static TLS, we
@@ -5234,13 +5236,13 @@
      * loading modules which use static TLS.
      */
     if (tls_static_space != 0) {
-	if (calculate_tls_end(off, obj->tlssize) > tls_static_space)
+	if (off > tls_static_space)
 	    return false;
     } else if (obj->tlsalign > tls_static_max_align) {
 	    tls_static_max_align = obj->tlsalign;
     }
 
-    tls_last_offset = obj->tlsoffset = off;
+    tls_last_offset = off;
     tls_last_size = obj->tlssize;
     obj->tls_done = true;
 
@@ -5257,8 +5259,11 @@
      * simplistic workaround to allow libGL.so.1 to be loaded and
      * unloaded multiple times.
      */
-    if (calculate_tls_end(obj->tlsoffset, obj->tlssize)
-	== calculate_tls_end(tls_last_offset, tls_last_size)) {
+    size_t off = obj->tlsoffset;
+#ifndef TLS_BELOW_TP
+    off += obj->tlssize;
+#endif
+    if (off == tls_last_offset) {
 	tls_last_offset -= obj->tlssize;
 	tls_last_size = 0;
     }