On Tue, Jun 9, 2015 at 1:57 PM, Jakub Jelinek <ja...@redhat.com> wrote: > Hi! > > As mentioned in the PR, when trying to split: > (insn 7 15 13 2 (set (reg:DI 0 ax [92]) > (mem:DI (plus:SI (plus:SI (mult:SI (reg/v:SI 1 dx [orig:89 b ] [89]) > (const_int 8 [0x8])) > (unspec:SI [ > (const_int 0 [0]) > ] UNSPEC_TP)) > (reg:SI 0 ax [91])) [1 a S8 A64])) rh1212265.i:2 85 > {*movdi_internal} > (nil)) > which has collisions == 2 (both ax and dx used on lhs and both > ax and dx used in the memory address), we generate invalid insn > - lea with %gs: or %fs: in it. This patch fixes it by using > normal lea instead (so remove the unspec UNSPEC_TP from the address > for lea) and duplicating the unspec UNSPEC_TP to all the memory loads. > > Bootstrapped/regtested on x86_64-linux and i686-linux, ok for > trunk/5/4.9/4.8? > > 2015-06-09 Jakub Jelinek <ja...@redhat.com> > > PR target/66470 > * config/i386/i386.c (ix86_split_long_move): For collisions > involving direct tls segment refs, move the UNSPEC_TP out of > the address for lea, to each of the memory loads. > > * gcc.dg/tls/pr66470.c: New test. > > --- gcc/config/i386/i386.c.jj 2015-06-08 15:41:19.000000000 +0200 > +++ gcc/config/i386/i386.c 2015-06-09 11:50:29.960859723 +0200 > @@ -22866,7 +22866,7 @@ ix86_split_long_move (rtx operands[]) > Do an lea to the last part and use only one colliding move. */ > else if (collisions > 1) > { > - rtx base; > + rtx base, addr, tls_base = NULL_RTX; > > collisions = 1; > > @@ -22877,10 +22877,45 @@ ix86_split_long_move (rtx operands[]) > if (GET_MODE (base) != Pmode) > base = gen_rtx_REG (Pmode, REGNO (base)); > > - emit_insn (gen_rtx_SET (base, XEXP (part[1][0], 0))); > + addr = XEXP (part[1][0], 0); > + if (TARGET_TLS_DIRECT_SEG_REFS) > + { > + struct ix86_address parts; > + int ok = ix86_decompose_address (addr, &parts); > + gcc_assert (ok); > + if (parts.seg == DEFAULT_TLS_SEG_REG) > + { > + /* It is not valid to use %gs: or %fs: in > + lea though, so we need to remove it from the > + address used for lea and add it to each individual > + memory loads instead. */ > + addr = copy_rtx (addr); > + rtx *x = &addr; > + while (GET_CODE (*x) == PLUS)
Why not use RTX iterators here? IMO, it would be much more readable. Uros. > + { > + for (i = 0; i < 2; i++) > + if (GET_CODE (XEXP (*x, i)) == UNSPEC > + && XINT (XEXP (*x, i), 1) == UNSPEC_TP) > + { > + tls_base = XEXP (*x, i); > + *x = XEXP (*x, 1 - i); > + break; > + } > + if (tls_base) > + break; > + x = &XEXP (*x, 0); > + } > + gcc_assert (tls_base); > + } > + } > + emit_insn (gen_rtx_SET (base, addr)); > + if (tls_base) > + base = gen_rtx_PLUS (GET_MODE (base), base, tls_base); > part[1][0] = replace_equiv_address (part[1][0], base); > for (i = 1; i < nparts; i++) > { > + if (tls_base) > + base = copy_rtx (base); > tmp = plus_constant (Pmode, base, UNITS_PER_WORD * i); > part[1][i] = replace_equiv_address (part[1][i], tmp); > } > --- gcc/testsuite/gcc.dg/tls/pr66470.c.jj 2015-06-09 11:59:05.543954781 > +0200 > +++ gcc/testsuite/gcc.dg/tls/pr66470.c 2015-06-09 11:58:43.000000000 +0200 > @@ -0,0 +1,29 @@ > +/* PR target/66470 */ > +/* { dg-do compile } */ > +/* { dg-options "-O2" } */ > +/* { dg-require-effective-target tls } */ > + > +extern __thread unsigned long long a[10]; > +extern __thread struct S { int a, b; } b[10]; > + > +unsigned long long > +foo (long x) > +{ > + return a[x]; > +} > + > +struct S > +bar (long x) > +{ > + return b[x]; > +} > + > +#ifdef __SIZEOF_INT128__ > +extern __thread unsigned __int128 c[10]; > + > +unsigned __int128 > +baz (long x) > +{ > + return c[x]; > +} > +#endif > > Jakub