adding support for TLS global variables

runtime/linker/vx_link32.ld

@@ -89,12 +89,22 @@ SECTIONS
   .gcc_except_table   : ONLY_IF_RW { *(.gcc_except_table .gcc_except_table.*) }
   .exception_ranges   : ONLY_IF_RW { *(.exception_ranges*) }
   /* Thread Local Storage sections  */
-  .tdata    :
+  .tdata	  :
    {
      PROVIDE_HIDDEN (__tdata_start = .);
      *(.tdata .tdata.* .gnu.linkonce.td.*)
+     PROVIDE_HIDDEN (__tdata_end = .);
    }
-  .tbss     : { *(.tbss .tbss.* .gnu.linkonce.tb.*) *(.tcommon) }
+   PROVIDE (__tdata_size = SIZEOF (.tdata));
+  .tbss		  : 
+  {
+    PROVIDE_HIDDEN (__tbss_start = .);
+    PROVIDE_HIDDEN (__tbss_offset = ABSOLUTE (__tbss_start - __tdata_start)); 
+    *(.tbss .tbss.* .gnu.linkonce.tb.*) *(.tcommon)
+    PROVIDE_HIDDEN (__tbss_end = .);
+  }
+  PROVIDE (__tbss_size = SIZEOF (.tbss));
+  PROVIDE (__tcb_aligned_size = ALIGN(__tbss_end - __tdata_start, 64));
   .preinit_array    :
   {
     PROVIDE_HIDDEN (__preinit_array_start = .);

runtime/linker/vx_link64.ld

@@ -89,12 +89,22 @@ SECTIONS
   .gcc_except_table   : ONLY_IF_RW { *(.gcc_except_table .gcc_except_table.*) }
   .exception_ranges   : ONLY_IF_RW { *(.exception_ranges*) }
   /* Thread Local Storage sections  */
-  .tdata    :
+  .tdata	  :
    {
      PROVIDE_HIDDEN (__tdata_start = .);
      *(.tdata .tdata.* .gnu.linkonce.td.*)
+     PROVIDE_HIDDEN (__tdata_end = .);
    }
-  .tbss     : { *(.tbss .tbss.* .gnu.linkonce.tb.*) *(.tcommon) }
+   PROVIDE (__tdata_size = SIZEOF (.tdata));
+  .tbss		  : 
+  {
+    PROVIDE_HIDDEN (__tbss_start = .);
+    PROVIDE_HIDDEN (__tbss_offset = ABSOLUTE (__tbss_start - __tdata_start)); 
+    *(.tbss .tbss.* .gnu.linkonce.tb.*) *(.tcommon)
+    PROVIDE_HIDDEN (__tbss_end = .);
+  }
+  PROVIDE (__tbss_size = SIZEOF (.tbss));
+  PROVIDE (__tcb_aligned_size = ALIGN(__tbss_end - __tdata_start, 64));
   .preinit_array    :
   {
     PROVIDE_HIDDEN (__preinit_array_start = .);

runtime/src/vx_start.S

@@ -4,32 +4,44 @@
 .global _start
 .type   _start, @function
 _start:
-  
-  # execute stack initialization on all warps
-  la a1, vx_set_sp
-  csrr a0, CSR_NW  # get num warps
-  .insn s 0x6b, 1, a1, 0(a0)  # wspawn a0, a1
-  jal vx_set_sp
 
+  # initialize per-thread registers
+  csrr a0, CSR_NW  # get num warps
+  la a1, init_regs
+  .insn s 0x6b, 1, a1, 0(a0)  # wspawn a0, a1
+  jal init_regs
+  # return back to single thread execution
+  li a0, 1
+  .insn s 0x6b, 0, x0, 0(a0)  # tmc a0
+
+  # initialize TLS for all warps
+  csrr a0, CSR_NW  # get num warps
+  la a1, __init_tls
+  .insn s 0x6b, 1, a1, 0(a0)  # wspawn a0, a1
+  call __init_tls
   # return back to single thread execution
   li a0, 1
   .insn s 0x6b, 0, x0, 0(a0)  # tmc a0
   
-  # Clear the bss segment
+  # clear BSS segment
   la      a0, _edata
   la      a2, _end
   sub     a2, a2, a0
   li      a1, 0
   call    memset
 
+  # Initialize trap vector
+  # a t0, trap_entry
+  # csrw mtvec, t0
+
   # Register global termination functions
-  la      a0, __libc_fini_array   
+  la      a0, __libc_fini_array
 
   # to be called upon exit
-  call    atexit                  
+  call    atexit
 
   # Run global initialization functions
-  call    __libc_init_array       
+  call    __libc_init_array
 
   # call main program routine
   call    main
@@ -49,9 +61,9 @@ _exit:
   .insn s 0x6b, 0, x0, 0(a0)  # tmc a0
 
 .section .text
-.type vx_set_sp, @function
-.global vx_set_sp
-vx_set_sp:
+.type init_regs, @function
+.global init_regs
+init_regs:
   # activate all threads
   li a0, -1
   .insn s 0x6b, 0, x0, 0(a0)  # tmc a0
@@ -66,12 +78,21 @@ vx_set_sp:
   # set stack pointer
   li sp, SMEM_BASE_ADDR # load stack base address
   #if SM_ENABLE
-  csrr a2, CSR_LTID    # get local thread id
+  csrr a0, CSR_LTID    # get local thread id
   #else
-  csrr a2, CSR_GTID    # get global thread id
+  csrr a0, CSR_GTID    # get global thread id
   #endif
-  slli a1, a2, STACK_LOG2_SIZE
-  sub  sp, sp, a1      # sub thread block
+  sll  a1, a0, STACK_LOG2_SIZE
+  sub  sp, sp, a1
+
+  # set thread pointer register
+  # use address space after BSS region
+  # ensure cacheline alignment
+  la      a1, __tcb_aligned_size
+  mul     a0, a0, a1
+  la      tp, _end + 63
+  add     tp, tp, a0
+  and     tp, tp, -64
 
   # disable active warps except warp0
   csrr a3, CSR_LWID    # get local wid

runtime/src/vx_syscalls.c

@@ -3,6 +3,7 @@
 #include <unistd.h>
 #include <vx_intrinsics.h>
 #include <vx_print.h>
+#include <string.h>
  
 int _close(int file) { return -1; }
  
@@ -22,18 +23,36 @@ caddr_t _sbrk(int incr) {
 }
  
 int _write(int file, char *ptr, int len) {
-    int i; 
-    for (i = 0; i < len; ++i) {
-        vx_putchar(*ptr++);
-    }
-    return len;
- }
+  int i; 
+  for (i = 0; i < len; ++i) {
+    vx_putchar(*ptr++);
+  }
+  return len;
+}
 
- int _kill(int pid, int sig) { return -1; }
+int _kill(int pid, int sig) { return -1; }
 
- int _getpid() {
-     return vx_warp_gid();
- }
+int _getpid() {
+  return vx_warp_gid();
+}
+
+void __init_tls(void) {  
+  extern char __tdata_start[];
+  extern char __tbss_offset[];
+  extern char __tdata_size[];
+  extern char __tbss_size[];
+
+  // activate all threads
+  vx_tmc(-1);
+
+  // TLS memory initialization
+  register char *__thread_self __asm__ ("tp");
+  memcpy(__thread_self, __tdata_start, (size_t)__tdata_size);
+  memset(__thread_self + (size_t)__tbss_offset, 0, (size_t)__tbss_size);
+
+  // back to single thread execution
+  vx_tmc(0 == vx_warp_id());
+}
 
  #ifdef HAVE_INITFINI_ARRAY
 
@@ -48,9 +67,7 @@ extern void _init (void);
 #endif
 
 /* Iterate over all the init routines.  */
-void
-__libc_init_array (void)
-{
+void __libc_init_array (void) {
   size_t count;
   size_t i;
 
@@ -77,9 +94,7 @@ extern void _fini (void);
 #endif
 
 /* Run all the cleanup routines.  */
-void
-__libc_fini_array (void)
-{
+void __libc_fini_array (void) {
   size_t count;
   size_t i;