[Bug target/36189] Invalid movabsq $mycom_@TPOFF-8, %rbp for TLS

["jakub at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org>]

------- Comment #1 from jakub at gcc dot gnu dot org  2008-08-20 15:42 -------
program foo
  implicit none
  common /mycom/ arr(1000000000_8:1000000002_8)
!$omp threadprivate (/mycom/)
  integer*8 i
  real*8 arr
  do i=1000000000_8,1000000002_8
    write(*,*) i
    arr(i)=0.0d0
  enddo
end program

is another testcase, now with CONST_INT offset that doesn't even fit into
SImode.
I've tried to reject TPOFF/DTPOFF which would be rejected by
x86_64_immediate_operand also in legitimate_constant_p (they really aren't
legitimate, can't be put into .data section nor instructions):
--- i386.c.jj6  2007-06-26 13:38:46.000000000 +0200
+++ i386.c      2008-08-20 16:30:19.000000000 +0200
@@ -6130,6 +6130,8 @@ darwin_local_data_pic (rtx disp)
 bool
 legitimate_constant_p (rtx x)
 {
+  rtx off = NULL_RTX;
+
   switch (GET_CODE (x))
     {
     case CONST:
@@ -6139,6 +6141,7 @@ legitimate_constant_p (rtx x)
        {
          if (GET_CODE (XEXP (x, 1)) != CONST_INT)
            return false;
+         off = XEXP (x, 1);
          x = XEXP (x, 0);
        }

@@ -6154,10 +6157,24 @@ legitimate_constant_p (rtx x)
          case UNSPEC_TPOFF:
          case UNSPEC_NTPOFF:
            x = XVECEXP (x, 0, 0);
+           /* GNU ld only handles 32-bit @TPOFF relocation on input.  */
+           if (TARGET_64BIT
+               && off
+               && (INTVAL (off) < 0
+                   || trunc_int_for_mode (INTVAL (off),
+                                          SImode) != INTVAL (off)))
+             return false;
            return (GET_CODE (x) == SYMBOL_REF
                    && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_EXEC);
          case UNSPEC_DTPOFF:
            x = XVECEXP (x, 0, 0);
+           /* GNU ld only handles 32-bit @DTPOFF relocation on input.  */
+           if (TARGET_64BIT
+               && off
+               && (INTVAL (off) < 0
+                   || trunc_int_for_mode (INTVAL (off),
+                                          SImode) != INTVAL (off)))
+             return false;
            return (GET_CODE (x) == SYMBOL_REF
                    && SYMBOL_REF_TLS_MODEL (x) == TLS_MODEL_LOCAL_DYNAMIC);
          default:

but still an instruction with @TPOFF-8 is created and then ICEs during
recognition, eventhough both legitimate_constant_p and legitimate_address_p
reject it as invalid memory address.  Any ideas?

In 4.3/trunk this is IMHO a latent issue, the memory reference expansion just
reorders the mycom_+(i*8)-8 operations differently.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36189