Patchwork PR inline-asm/55934

login
register
mail settings
Submitter Vladimir Makarov
Date Jan. 23, 2013, 6:43 p.m.
Message ID <51002F41.8090005@redhat.com>
Download mbox | patch
Permalink /patch/215019/
State New
Headers show

Comments

Vladimir Makarov - Jan. 23, 2013, 6:43 p.m.
On 13-01-22 4:41 AM, Steven Bosscher wrote:
> On Mon, Jan 21, 2013 at 5:22 PM, Vladimir Makarov wrote:
>> I'd prefer the above change than just keeping
>> lra_invalidate_insn_data call.  I think it is more safe solution for other
>> parts of LRA code.
> I agree, but unfortunately the compiler does not...
>
> With that lra.c change, I get extra fails:
>
> +FAIL: gcc.target/i386/20011029-2.c (internal compiler error)
> +FAIL: gcc.target/i386/20011029-2.c (test for excess errors)
> +FAIL: gcc.target/i386/pr21291.c (internal compiler error)
> +FAIL: gcc.target/i386/pr21291.c (test for excess errors)
>
> These are constrain_operands(1) failures in check_rtl. Apparently some
> relevant info is lost in lra_update_insn_recog_data. Not sure how to
> debug this...
>
Steven, sorry for the delay with the answer.  I was busy with other things.

The error occurs because a pseudo in asm insn is not changed into hard 
register as the pseudo info is incorrect after info updating.

You should just use lra_update_insn_regno_info. The patch (and the patch 
is ok to commit) should look like

               /* If this operand accepts a register, and if the
                  register class has at least one allocatable register,
                  then this operand can be reloaded.  */
@@ -2742,10 +2758,6 @@ curr_insn_transform (void)
         swap_operands (commutative);
      }

-  /* The operands don't meet the constraints.  goal_alt describes the
-     alternative that we could reach by reloading the fewest operands.
-     Reload so as to fit it.  */
-
    if (! alt_p && ! sec_mem_p)
      {
        /* No alternative works with reloads??  */
Steven Bosscher - Jan. 24, 2013, 10:30 a.m.
On Wed, Jan 23, 2013 at 7:43 PM, Vladimir Makarov wrote:
> The error occurs because a pseudo in asm insn is not changed into hard
> register as the pseudo info is incorrect after info updating.
>
> You should just use lra_update_insn_regno_info. The patch (and the patch is
> ok to commit) should look like

Right, that solves the problem. Thanks for the help!

I've committed the attached patch.

Ciao!
Steven

Patch

Index: lra-assigns.c
===================================================================
--- lra-assigns.c       (revision 195278)
+++ lra-assigns.c       (working copy)
@@ -1240,6 +1240,23 @@  assign_by_spills (void)
                   asm_p = true;
                   error_for_asm (insn,
                                  "%<asm%> operand has impossible 
constraints");
+                 /* Avoid further trouble with this insn.
+                    For asm goto, instead of fixing up all the edges
+                    just clear the template and clear input operands
+                    (asm goto doesn't have any output operands).  */
+                 if (JUMP_P (insn))
+                   {
+                     rtx asm_op = extract_asm_operands (PATTERN (insn));
+                     ASM_OPERANDS_TEMPLATE (asm_op) = ggc_strdup ("");
+                     ASM_OPERANDS_INPUT_VEC (asm_op) = rtvec_alloc (0);
+                     ASM_OPERANDS_INPUT_CONSTRAINT_VEC (asm_op) = 
rtvec_alloc (0);
+                    lra_update_insn_regno_info (insn);
+                   }
+                 else
+                   {
+                     PATTERN (insn) = gen_rtx_USE (VOIDmode, const0_rtx);
+                     lra_set_insn_deleted (insn);
+                   }
                 }
             }
           lra_assert (asm_p);
@@ -1263,6 +1280,9 @@  assign_by_spills (void)
           bitmap_ior_into (&changed_insns,
&lra_reg_info[sorted_pseudos[i]].insn_bitmap);
         }
+
+      /* FIXME: Look up the changed insns in the cached LRA insn data using
+        an EXECUTE_IF_SET_IN_BITMAP over changed_insns. */
        FOR_EACH_BB (bb)
         FOR_BB_INSNS (bb, insn)
         if (bitmap_bit_p (&changed_insns, INSN_UID (insn)))
Index: lra-constraints.c
===================================================================
--- lra-constraints.c   (revision 195322)
+++ lra-constraints.c   (working copy)
@@ -1847,11 +1847,27 @@  process_alt_operands (int only_alternati
               int const_to_mem = 0;
               bool no_regs_p;

+             /* If this alternative asks for a specific reg class, see 
if there
+                is at least one allocatable register in that class.  */
               no_regs_p
                 = (this_alternative == NO_REGS
                    || (hard_reg_set_subset_p
(reg_class_contents[this_alternative],
                         lra_no_alloc_regs)));
+
+             /* For asms, verify that the class for this alternative is 
possible
+                for the mode that is specified.  */
+             if (!no_regs_p && REG_P (op) && INSN_CODE (curr_insn) < 0)
+               {
+                 int i;
+                 for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
+                   if (HARD_REGNO_MODE_OK (i, mode)
+                       && in_hard_reg_set_p 
(reg_class_contents[this_alternative], mode, i))
+                     break;
+                 if (i == FIRST_PSEUDO_REGISTER)
+                   winreg = false;
+               }
+