Re: [Google/gcc-4_9][PATCH][target/x86_64] PR 63538

[Sriraman Tallam <tmsriram@google.com>]

[-- Attachment #1: Type: text/plain, Size: 1696 bytes --]

On Mon, Oct 20, 2014 at 10:59 AM, Xinliang David Li <davidxl@google.com> wrote:
> Perhaps explicitly allowing STRING_CST to go through the large data
> check, instead of removing the var-decl check? Do you see other
> opcodes that need to be handled too?

I do not see any other opcodes explicitly but the code in
ix86_in_large_data_p seemingly handles all opcodes other than
FUNCTION_DECL through:

if (TREE_CODE (exp) == VAR_DECL && DECL_SECTION_NAME (exp))
    {
        <blah>
    }
else {
    <blah>
  }


However, I have modified the patch to explicitly check for STRING_CST
and I cannot think of any other case where the constant goes into
rodata but is not accessed via a VAR_DECL. Also note that TREE_STATIC
(decl) is true for STRING_CST.

Thanks
Sri




>
> David
>
> On Mon, Oct 20, 2014 at 10:46 AM, Sriraman Tallam <tmsriram@google.com> wrote:
>> On Mon, Oct 20, 2014 at 10:42 AM, Xinliang David Li <davidxl@google.com> wrote:
>>> Why removing the tree_code check?
>>
>> The actual problem happens because STRING_CSTs (end up in .lrodata)
>> are not set a far address as they dont match the VAR_DECL check here.
>> Futher,  "ix86_in_large_data_p" call has the TREE_CODE check to do the
>> right thing so this seems unnecessary & buggy here.
>>
>> Thanks
>> Sri
>>
>>>
>>> David
>>>
>>> On Mon, Oct 20, 2014 at 10:35 AM, Sriraman Tallam <tmsriram@google.com> wrote:
>>>> Hi,
>>>>
>>>>    This patch is under review for trunk GCC :
>>>> https://gcc.gnu.org/ml/gcc-patches/2014-10/msg01638.html.
>>>>
>>>>     In the mean time, is this ok for google/gcc-4_9 branch?  Without
>>>> this, -mcmodel=medium is unusable if .lrodata goes beyond the 2G
>>>> boundary.
>>>>
>>>> Thanks
>>>> Sri

[-- Attachment #2: pr63538.txt --]
[-- Type: text/plain, Size: 1196 bytes --]

Index: config/i386/i386.c
===================================================================
--- config/i386/i386.c	(revision 216287)
+++ config/i386/i386.c	(working copy)
@@ -41331,7 +41331,7 @@ ix86_encode_section_info (tree decl, rtx rtl, int
 {
   default_encode_section_info (decl, rtl, first);
 
-  if (TREE_CODE (decl) == VAR_DECL
+  if ((TREE_CODE (decl) == VAR_DECL || TREE_CODE (decl) == STRING_CST)
       && (TREE_STATIC (decl) || DECL_EXTERNAL (decl))
       && ix86_in_large_data_p (decl))
     SYMBOL_REF_FLAGS (XEXP (rtl, 0)) |= SYMBOL_FLAG_FAR_ADDR;
Index: testsuite/gcc.dg/pr63538.c
===================================================================
--- testsuite/gcc.dg/pr63538.c	(revision 0)
+++ testsuite/gcc.dg/pr63538.c	(revision 0)
@@ -0,0 +1,14 @@
+/* PR63538 is about not using 64-bit addresses for .lrodata accesses when it
+   involves STRING_CSTs.  */
+/* { dg-do compile { target x86_64-*-* } } */
+/* { dg-options "-O2 -mcmodel=medium -mlarge-data-threshold=0" { target x86_64-*-* } } */
+
+#include <stdio.h>
+
+const char *str = "Hello World";
+
+int main() {
+ printf("str = %p %s\n",str, str);
+ return 0;
+}
+/* { dg-final { scan-assembler-not "movl" } } */