RFA (libstdc++): PATCH to implement C++17 over-aligned new

[Jason Merrill]

I'm checking in this patch, which should fix the remaining issues:

On Sat, Sep 10, 2016 at 2:14 PM, Andreas Schwab <schwab@linux-m68k.org> wrote:
> FAIL: g++.dg/cpp1z/aligned-new1.C   (test for excess errors)
> Excess errors:
> /daten/aranym/gcc/gcc-20160910/gcc/testsuite/g++.dg/cpp1z/aligned-new1.C:10:20: warning: requested alignment 64 is larger than 16 [-Wattributes]
> FAIL: g++.dg/cpp1z/aligned-new1.C   execution test
> FAIL: g++.dg/cpp1z/aligned-new4.C    (test for warnings, line 11)
> FAIL: g++.dg/cpp1z/aligned-new4.C   (test for excess errors)
> Excess errors:
> /daten/aranym/gcc/gcc-20160910/gcc/testsuite/g++.dg/cpp1z/aligned-new4.C:3:20: warning: requested alignment 64 is larger than 16 [-Wattributes]
> /daten/aranym/gcc/gcc-20160910/gcc/testsuite/g++.dg/cpp1z/aligned-new4.C:4:20: warning: requested alignment 64 is larger than 16 [-Wattributes]
> FAIL: g++.dg/cpp1z/aligned-new5.C  -std=gnu++11 (test for excess errors)
> Excess errors:
> aligned-new5.C:(.text+0xe): undefined reference to `operator new(unsigned int, std::align_val_t)'
>
> Andreas.
>
> --
> Andreas Schwab, schwab@linux-m68k.org
> GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
> "And now for something completely different."
-------------- next part --------------
commit da8e3c2d6ca085aeb815d741e4d858b1216473a1
Author: Jason Merrill <jason@redhat.com>
Date:   Mon Sep 12 15:01:06 2016 -0400

            Fix aligned-new tests on m68k.
    
            * c-common.c (check_cxx_fundamental_alignment_constraints): Fix
            bit/byte confusion, allow large alignment for types.

diff --git a/gcc/c-family/c-common.c b/gcc/c-family/c-common.c
index 16f6548..b4f4d10 100644
--- a/gcc/c-family/c-common.c
+++ b/gcc/c-family/c-common.c
@@ -7836,8 +7836,7 @@ check_user_alignment (const_tree align, bool allow_zero)
   return i;
 }
 
-/* 
-   If in c++-11, check if the c++-11 alignment constraint with respect
+/* If in c++-11, check if the c++-11 alignment constraint with respect
    to fundamental alignment (in [dcl.align]) are satisfied.  If not in
    c++-11 mode, does nothing.
 
@@ -7862,7 +7861,7 @@ check_cxx_fundamental_alignment_constraints (tree node,
 					     int flags)
 {
   bool alignment_too_large_p = false;
-  unsigned requested_alignment = 1U << align_log;
+  unsigned requested_alignment = (1U << align_log) * BITS_PER_UNIT;
   unsigned max_align = 0;
 
   if ((!(flags & ATTR_FLAG_CXX11) && !warn_cxx_compat)
@@ -7906,15 +7905,19 @@ check_cxx_fundamental_alignment_constraints (tree node,
     }
   else if (TYPE_P (node))
     {
-      /* Let's be liberal for types.  */
-      if (requested_alignment > (max_align = BIGGEST_ALIGNMENT))
+      /* Let's be liberal for types.  BIGGEST_ALIGNMENT is the largest
+	 alignment a built-in type can require, MAX_OFILE_ALIGNMENT is the
+	 largest alignment the object file can represent, but a type that is
+	 only allocated dynamically could request even larger alignment.  So
+	 only limit type alignment to what TYPE_ALIGN can represent.  */
+      if (requested_alignment > (max_align = 8U << 28))
 	alignment_too_large_p = true;
     }
 
   if (alignment_too_large_p)
     pedwarn (input_location, OPT_Wattributes,
 	     "requested alignment %d is larger than %d",
-	     requested_alignment, max_align);
+	     requested_alignment / BITS_PER_UNIT, max_align / BITS_PER_UNIT);
 
   return !alignment_too_large_p;
 }