Fix GH-11188: Error when building TSRM in ARM64

Although the issue mentioned FreeBSD, this is a broader problem:
the current ARM64 code to load the TLS offset assumes a setup with
the non-default TLS model.
We have to call into the TLS descriptor similar as how the C compiler does it.
Although it looks like a lot of inline assembly, the compiler actually
optimizes this into only a few instructions on my emulator.

Author: nielsdos

TSRM/TSRM.c

@@ -784,6 +784,16 @@ TSRM_API size_t tsrm_get_ls_cache_tcb_offset(void)
 	asm("adrp %0, #__tsrm_ls_cache@TLVPPAGE\n\t"
 	    "ldr %0, [%0, #__tsrm_ls_cache@TLVPPAGEOFF]"
 	     : "=r" (ret));
+# elif defined(TSRM_TLS_MODEL_DEFAULT)
+	asm("stp x29, x30, [sp, #-16]!\n\t"
+		"adrp x0, #:tlsdesc:_tsrm_ls_cache\n\t"
+		"ldr x8, [x0, #:tlsdesc_lo12:_tsrm_ls_cache]\n\t"
+		"add x0, x0, #:tlsdesc_lo12:_tsrm_ls_cache\n\t"
+		".tlsdesccall _tsrm_ls_cache\n\t"
+		"blr x8\n\t"
+		"mov %0, x0\n\t"
+		"ldp x29, x30, [sp], #16"
+		 : "=r" (ret) : : "x0", "x8");
 # else
 	asm("mov %0, xzr\n\t"
 	    "add %0, %0, #:tprel_hi12:_tsrm_ls_cache, lsl #12\n\t"

TSRM/TSRM.h

@@ -149,6 +149,7 @@ TSRM_API const char *tsrm_api_name(void);
 
 #if !__has_attribute(tls_model) || defined(__FreeBSD__) || defined(__MUSL__) || defined(__HAIKU__)
 # define TSRM_TLS_MODEL_ATTR
+# define TSRM_TLS_MODEL_DEFAULT
 #elif __PIC__
 # define TSRM_TLS_MODEL_ATTR __attribute__((tls_model("initial-exec")))
 #else