[PATCH] Fix --enable-gather-detailed-mem-stats build.

[PATCH] Fix --enable-gather-detailed-mem-stats build.

[Martin Liška]

The build suffers from the static initialization order fiasco:

==30085== Invalid read of size 4
==30085==    at 0x1D451CD: hash_table_mod1 (hash-table.h:344)
==30085==    by 0x1D451CD: hash_table<hash_map<mem_alloc_description<vec_usage>::mem_location_hash, vec_usage*, simple_hashmap_traits<default_hash_traits<mem_alloc_description<vec_usage>::mem_location_hash>, vec_usage*> >::hash_entry, false, xcallocator>::find_with_hash(mem_location* const&, unsigned int) (hash-table.h:911)
==30085==    by 0x1D411F7: get (hash-map.h:185)
==30085==    by 0x1D411F7: register_descriptor (mem-stats.h:417)
==30085==    by 0x1D411F7: register_descriptor (mem-stats.h:451)
==30085==    by 0x1D411F7: vec_prefix::register_overhead(void*, unsigned long, unsigned long, char const*, int, char const*) (vec.c:132)
==30085==    by 0xA2DB28: reserve<loc_spans::span> (vec.h:294)
==30085==    by 0xA2DB28: vec<loc_spans::span, va_heap, vl_ptr>::reserve(unsigned int, bool, char const*, int, char const*) [clone .isra.0] (vec.h:1778)
==30085==    by 0x9039C7: reserve_exact (vec.h:1798)
==30085==    by 0x9039C7: create (vec.h:1813)
==30085==    by 0x9039C7: loc_spans (module.cc:3281)
==30085==    by 0x9039C7: __static_initialization_and_destruction_0 (module.cc:3340)
==30085==    by 0x9039C7: _GLOBAL__sub_I_map_context_from (gt-cp-module.h:360)
==30085==    by 0x1E00F6C: __libc_csu_init (elf-init.c:89)
==30085==    by 0x4FFF0DD: (below main) (in /lib64/libc-2.32.so)
==30085==  Address 0x28 is not stack'd, malloc'd or (recently) free'd

So vec_mem_desc is not initialized before a static member in module.cc.
That can be fixed by usage of GNU C++ extension init_priority.

Ready to be installed?
Thanks,
Martin

gcc/ChangeLog:

	* vec.c: Use init_priority for vec_mem_desc.
---
  gcc/vec.c | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/gcc/vec.c b/gcc/vec.c
index a28899170ed..1671f26c045 100644
--- a/gcc/vec.c
+++ b/gcc/vec.c
@@ -121,7 +121,8 @@ public:
  };
  
  /* Vector memory description.  */
-static mem_alloc_description <vec_usage> vec_mem_desc;
+static mem_alloc_description <vec_usage> vec_mem_desc
+  __attribute__ ((init_priority (101)));
  
  /* Account the overhead.  */
  
-- 
2.29.2

Re: [PATCH] Fix --enable-gather-detailed-mem-stats build.

[Iain Sandoe]

Martin Liška <mliska@suse.cz> wrote:

> The build suffers from the static initialization order fiasco:
>
> ==30085== Invalid read of size 4
> ==30085==    at 0x1D451CD: hash_table_mod1 (hash-table.h:344)
> ==30085==    by 0x1D451CD:  
> hash_table<hash_map<mem_alloc_description<vec_usage>::mem_location_hash,  
> vec_usage*,  
> simple_hashmap_traits<default_hash_traits<mem_alloc_description<vec_usage>::mem_location_hash>,  
> vec_usage*> >::hash_entry, false,  
> xcallocator>::find_with_hash(mem_location* const&, unsigned int)  
> (hash-table.h:911)
> ==30085==    by 0x1D411F7: get (hash-map.h:185)
> ==30085==    by 0x1D411F7: register_descriptor (mem-stats.h:417)
> ==30085==    by 0x1D411F7: register_descriptor (mem-stats.h:451)
> ==30085==    by 0x1D411F7: vec_prefix::register_overhead(void*, unsigned  
> long, unsigned long, char const*, int, char const*) (vec.c:132)
> ==30085==    by 0xA2DB28: reserve<loc_spans::span> (vec.h:294)
> ==30085==    by 0xA2DB28: vec<loc_spans::span, va_heap,  
> vl_ptr>::reserve(unsigned int, bool, char const*, int, char const*)  
> [clone .isra.0] (vec.h:1778)
> ==30085==    by 0x9039C7: reserve_exact (vec.h:1798)
> ==30085==    by 0x9039C7: create (vec.h:1813)
> ==30085==    by 0x9039C7: loc_spans (module.cc:3281)
> ==30085==    by 0x9039C7: __static_initialization_and_destruction_0  
> (module.cc:3340)
> ==30085==    by 0x9039C7: _GLOBAL__sub_I_map_context_from  
> (gt-cp-module.h:360)
> ==30085==    by 0x1E00F6C: __libc_csu_init (elf-init.c:89)
> ==30085==    by 0x4FFF0DD: (below main) (in /lib64/libc-2.32.so)
> ==30085==  Address 0x28 is not stack'd, malloc'd or (recently) free'd
>
> So vec_mem_desc is not initialized before a static member in module.cc.
> That can be fixed by usage of GNU C++ extension init_priority.
>
> Ready to be installed?
> Thanks,
> Martin
>
> gcc/ChangeLog:
>
> 	* vec.c: Use init_priority for vec_mem_desc.
> ---
> gcc/vec.c | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
>
> diff --git a/gcc/vec.c b/gcc/vec.c
> index a28899170ed..1671f26c045 100644
> --- a/gcc/vec.c
> +++ b/gcc/vec.c
> @@ -121,7 +121,8 @@ public:
> };
>  /* Vector memory description.  */
> -static mem_alloc_description <vec_usage> vec_mem_desc;
> +static mem_alloc_description <vec_usage> vec_mem_desc
> +  __attribute__ ((init_priority (101)));

This will break bootstrap on targets without init_priority (e.g. Darwin)

Iain

Re: [PATCH] Fix --enable-gather-detailed-mem-stats build.

[Martin Liška]

[-- Attachment #1: Type: text/plain, Size: 268 bytes --]

On 12/16/20 10:08 AM, Iain Sandoe wrote:
> This will break bootstrap on targets without init_priority (e.g. Darwin)

All right, based on the chat on IRC, I suggest using C static constructor
with corresponding priority.

Ready to be installed?
Thanks,
Martin

[-- Attachment #2: 0001-Fix-enable-gather-detailed-mem-stats-build.patch --]
[-- Type: text/x-patch, Size: 4239 bytes --]

From 4ecf4f8df8d5b6b75cfe7cc5df90176e0a5014b8 Mon Sep 17 00:00:00 2001
From: Martin Liska <mliska@suse.cz>
Date: Wed, 16 Dec 2020 10:02:10 +0100
Subject: [PATCH] Fix --enable-gather-detailed-mem-stats build.

The build suffers from the static initialization order fiasco:

==30085== Invalid read of size 4
==30085==    at 0x1D451CD: hash_table_mod1 (hash-table.h:344)
==30085==    by 0x1D451CD: hash_table<hash_map<mem_alloc_description<vec_usage>::mem_location_hash, vec_usage*, simple_hashmap_traits<default_hash_traits<mem_alloc_description<vec_usage>::mem_location_hash>, vec_usage*> >::hash_entry, false, xcallocator>::find_with_hash(mem_location* const&, unsigned int) (hash-table.h:911)
==30085==    by 0x1D411F7: get (hash-map.h:185)
==30085==    by 0x1D411F7: register_descriptor (mem-stats.h:417)
==30085==    by 0x1D411F7: register_descriptor (mem-stats.h:451)
==30085==    by 0x1D411F7: vec_prefix::register_overhead(void*, unsigned long, unsigned long, char const*, int, char const*) (vec.c:132)
==30085==    by 0xA2DB28: reserve<loc_spans::span> (vec.h:294)
==30085==    by 0xA2DB28: vec<loc_spans::span, va_heap, vl_ptr>::reserve(unsigned int, bool, char const*, int, char const*) [clone .isra.0] (vec.h:1778)
==30085==    by 0x9039C7: reserve_exact (vec.h:1798)
==30085==    by 0x9039C7: create (vec.h:1813)
==30085==    by 0x9039C7: loc_spans (module.cc:3281)
==30085==    by 0x9039C7: __static_initialization_and_destruction_0 (module.cc:3340)
==30085==    by 0x9039C7: _GLOBAL__sub_I_map_context_from (gt-cp-module.h:360)
==30085==    by 0x1E00F6C: __libc_csu_init (elf-init.c:89)
==30085==    by 0x4FFF0DD: (below main) (in /lib64/libc-2.32.so)
==30085==  Address 0x28 is not stack'd, malloc'd or (recently) free'd

So vec_mem_desc is not initialized before a static member in module.cc.
We can fix it by using constructor attribute.

gcc/ChangeLog:

	* vec.c (init_vec_mem_desc): New function.
	(vec_prefix::register_overhead): Use vec_mem_desc as pointer
	now.
	(vec_prefix::release_overhead): Likewise.
	(dump_vec_loc_statistics): Likewise.
---
 gcc/vec.c | 28 +++++++++++++++++++---------
 1 file changed, 19 insertions(+), 9 deletions(-)

diff --git a/gcc/vec.c b/gcc/vec.c
index a28899170ed..ac91400c2df 100644
--- a/gcc/vec.c
+++ b/gcc/vec.c
@@ -121,7 +121,17 @@ public:
 };
 
 /* Vector memory description.  */
-static mem_alloc_description <vec_usage> vec_mem_desc;
+static mem_alloc_description <vec_usage> *vec_mem_desc;
+
+/* Static constructor for vec_mem_desc that happens before ithe initialization
+   of all static variables.  */
+
+static void
+__attribute__((constructor (101)))
+init_vec_mem_desc (void)
+{
+  vec_mem_desc = new mem_alloc_description<vec_usage> ();
+}
 
 /* Account the overhead.  */
 
@@ -129,10 +139,10 @@ void
 vec_prefix::register_overhead (void *ptr, size_t elements,
 			       size_t element_size MEM_STAT_DECL)
 {
-  vec_mem_desc.register_descriptor (ptr, VEC_ORIGIN, false
-				    FINAL_PASS_MEM_STAT);
+  vec_mem_desc->register_descriptor (ptr, VEC_ORIGIN, false
+				     FINAL_PASS_MEM_STAT);
   vec_usage *usage
-    = vec_mem_desc.register_instance_overhead (elements * element_size, ptr);
+    = vec_mem_desc->register_instance_overhead (elements * element_size, ptr);
   usage->m_element_size = element_size;
   usage->m_items += elements;
   if (usage->m_items_peak < usage->m_items)
@@ -145,11 +155,11 @@ void
 vec_prefix::release_overhead (void *ptr, size_t size, size_t elements,
 			      bool in_dtor MEM_STAT_DECL)
 {
-  if (!vec_mem_desc.contains_descriptor_for_instance (ptr))
-    vec_mem_desc.register_descriptor (ptr, VEC_ORIGIN,
+  if (!vec_mem_desc->contains_descriptor_for_instance (ptr))
+    vec_mem_desc->register_descriptor (ptr, VEC_ORIGIN,
 				      false FINAL_PASS_MEM_STAT);
-  vec_usage *usage = vec_mem_desc.release_instance_overhead (ptr, size,
-							     in_dtor);
+  vec_usage *usage = vec_mem_desc->release_instance_overhead (ptr, size,
+							      in_dtor);
   usage->m_items -= elements;
 }
 
@@ -183,7 +193,7 @@ vec_prefix::calculate_allocation_1 (unsigned alloc, unsigned desired)
 void
 dump_vec_loc_statistics (void)
 {
-  vec_mem_desc.dump (VEC_ORIGIN);
+  vec_mem_desc->dump (VEC_ORIGIN);
 }
 
 #if CHECKING_P
-- 
2.29.2

Re: [PATCH] Fix --enable-gather-detailed-mem-stats build.

[Jakub Jelinek]

On Wed, Dec 16, 2020 at 10:20:09AM +0100, Martin Liška wrote:
> So vec_mem_desc is not initialized before a static member in module.cc.
> We can fix it by using constructor attribute.
> 
> gcc/ChangeLog:
> 
> 	* vec.c (init_vec_mem_desc): New function.
> 	(vec_prefix::register_overhead): Use vec_mem_desc as pointer
> 	now.
> 	(vec_prefix::release_overhead): Likewise.
> 	(dump_vec_loc_statistics): Likewise.
> ---
>  gcc/vec.c | 28 +++++++++++++++++++---------
>  1 file changed, 19 insertions(+), 9 deletions(-)
> 
> diff --git a/gcc/vec.c b/gcc/vec.c
> index a28899170ed..ac91400c2df 100644
> --- a/gcc/vec.c
> +++ b/gcc/vec.c
> @@ -121,7 +121,17 @@ public:
>  };
>  
>  /* Vector memory description.  */
> -static mem_alloc_description <vec_usage> vec_mem_desc;
> +static mem_alloc_description <vec_usage> *vec_mem_desc;
> +
> +/* Static constructor for vec_mem_desc that happens before ithe initialization

s/ithe/the/

> +   of all static variables.  */
> +
> +static void
> +__attribute__((constructor (101)))

I think this needs to be guarded based on which compiler is used to compile
GCC.  Perhaps we could say that we don't support
--enable-gather-detailed-mem-stats when the compiler isn't built by GCC (or
other compiler that supports the constructor attribute) and #error on that.

> +init_vec_mem_desc (void)
> +{
> +  vec_mem_desc = new mem_alloc_description<vec_usage> ();
> +}
>  
>  /* Account the overhead.  */
>  
> @@ -129,10 +139,10 @@ void
>  vec_prefix::register_overhead (void *ptr, size_t elements,
>  			       size_t element_size MEM_STAT_DECL)
>  {
> -  vec_mem_desc.register_descriptor (ptr, VEC_ORIGIN, false
> -				    FINAL_PASS_MEM_STAT);
> +  vec_mem_desc->register_descriptor (ptr, VEC_ORIGIN, false
> +				     FINAL_PASS_MEM_STAT);
>    vec_usage *usage
> -    = vec_mem_desc.register_instance_overhead (elements * element_size, ptr);
> +    = vec_mem_desc->register_instance_overhead (elements * element_size, ptr);
>    usage->m_element_size = element_size;
>    usage->m_items += elements;
>    if (usage->m_items_peak < usage->m_items)
> @@ -145,11 +155,11 @@ void
>  vec_prefix::release_overhead (void *ptr, size_t size, size_t elements,
>  			      bool in_dtor MEM_STAT_DECL)
>  {
> -  if (!vec_mem_desc.contains_descriptor_for_instance (ptr))
> -    vec_mem_desc.register_descriptor (ptr, VEC_ORIGIN,
> +  if (!vec_mem_desc->contains_descriptor_for_instance (ptr))
> +    vec_mem_desc->register_descriptor (ptr, VEC_ORIGIN,
>  				      false FINAL_PASS_MEM_STAT);
> -  vec_usage *usage = vec_mem_desc.release_instance_overhead (ptr, size,
> -							     in_dtor);
> +  vec_usage *usage = vec_mem_desc->release_instance_overhead (ptr, size,
> +							      in_dtor);
>    usage->m_items -= elements;
>  }
>  
> @@ -183,7 +193,7 @@ vec_prefix::calculate_allocation_1 (unsigned alloc, unsigned desired)
>  void
>  dump_vec_loc_statistics (void)
>  {
> -  vec_mem_desc.dump (VEC_ORIGIN);
> +  vec_mem_desc->dump (VEC_ORIGIN);
>  }
>  
>  #if CHECKING_P

	Jakub

Re: [PATCH] Fix --enable-gather-detailed-mem-stats build.

[Martin Liška]

On 12/16/20 10:26 AM, Jakub Jelinek wrote:
> I think this needs to be guarded based on which compiler is used to compile
> GCC.

Do you mean basing that on __GNUC__?

>  Perhaps we could say that we don't support
> --enable-gather-detailed-mem-stats when the compiler isn't built by GCC (or
> other compiler that supports the constructor attribute) and #error on that.

Yes, it's a reasonable requirement I would say.

Martin

Re: [PATCH] Fix --enable-gather-detailed-mem-stats build.

[Rainer Orth]

Hi Jakub,

> On Wed, Dec 16, 2020 at 10:20:09AM +0100, Martin Liška wrote:
>> So vec_mem_desc is not initialized before a static member in module.cc.
>> We can fix it by using constructor attribute.
[...]
>> +   of all static variables.  */
>> +
>> +static void
>> +__attribute__((constructor (101)))
>
> I think this needs to be guarded based on which compiler is used to compile
> GCC.  Perhaps we could say that we don't support
> --enable-gather-detailed-mem-stats when the compiler isn't built by GCC (or
> other compiler that supports the constructor attribute) and #error on that.

not only that: if a target doesn't support constructor priorities (like
Solaris 11.3, unlike 11.4), this makes gcc error out.

	Rainer

-- 
-----------------------------------------------------------------------------
Rainer Orth, Center for Biotechnology, Bielefeld University

Re: [PATCH] Fix --enable-gather-detailed-mem-stats build.

[Martin Liška]

[-- Attachment #1: Type: text/plain, Size: 921 bytes --]

On 12/16/20 10:38 AM, Rainer Orth wrote:
> Hi Jakub,
> 
>> On Wed, Dec 16, 2020 at 10:20:09AM +0100, Martin Liška wrote:
>>> So vec_mem_desc is not initialized before a static member in module.cc.
>>> We can fix it by using constructor attribute.
> [...]
>>> +   of all static variables.  */
>>> +
>>> +static void
>>> +__attribute__((constructor (101)))
>>
>> I think this needs to be guarded based on which compiler is used to compile
>> GCC.  Perhaps we could say that we don't support
>> --enable-gather-detailed-mem-stats when the compiler isn't built by GCC (or
>> other compiler that supports the constructor attribute) and #error on that.
> 
> not only that: if a target doesn't support constructor priorities (like
> Solaris 11.3, unlike 11.4), this makes gcc error out.
> 
> 	Rainer
> 

I see, I'm then suggesting version 3 of the patch that does not depend
on a constructor.

Thoughts?
Thanks,
Martin

Martin

[-- Attachment #2: 0001-Fix-enable-gather-detailed-mem-stats-build.patch --]
[-- Type: text/x-patch, Size: 4023 bytes --]

From 3ac0d258887426b30d3e1885841b1bdf4e53100b Mon Sep 17 00:00:00 2001
From: Martin Liska <mliska@suse.cz>
Date: Wed, 16 Dec 2020 10:02:10 +0100
Subject: [PATCH] Fix --enable-gather-detailed-mem-stats build.

The build suffers from the static initialization order fiasco:

==30085== Invalid read of size 4
==30085==    at 0x1D451CD: hash_table_mod1 (hash-table.h:344)
==30085==    by 0x1D451CD: hash_table<hash_map<mem_alloc_description<vec_usage>::mem_location_hash, vec_usage*, simple_hashmap_traits<default_hash_traits<mem_alloc_description<vec_usage>::mem_location_hash>, vec_usage*> >::hash_entry, false, xcallocator>::find_with_hash(mem_location* const&, unsigned int) (hash-table.h:911)
==30085==    by 0x1D411F7: get (hash-map.h:185)
==30085==    by 0x1D411F7: register_descriptor (mem-stats.h:417)
==30085==    by 0x1D411F7: register_descriptor (mem-stats.h:451)
==30085==    by 0x1D411F7: vec_prefix::register_overhead(void*, unsigned long, unsigned long, char const*, int, char const*) (vec.c:132)
==30085==    by 0xA2DB28: reserve<loc_spans::span> (vec.h:294)
==30085==    by 0xA2DB28: vec<loc_spans::span, va_heap, vl_ptr>::reserve(unsigned int, bool, char const*, int, char const*) [clone .isra.0] (vec.h:1778)
==30085==    by 0x9039C7: reserve_exact (vec.h:1798)
==30085==    by 0x9039C7: create (vec.h:1813)
==30085==    by 0x9039C7: loc_spans (module.cc:3281)
==30085==    by 0x9039C7: __static_initialization_and_destruction_0 (module.cc:3340)
==30085==    by 0x9039C7: _GLOBAL__sub_I_map_context_from (gt-cp-module.h:360)
==30085==    by 0x1E00F6C: __libc_csu_init (elf-init.c:89)
==30085==    by 0x4FFF0DD: (below main) (in /lib64/libc-2.32.so)
==30085==  Address 0x28 is not stack'd, malloc'd or (recently) free'd

So vec_mem_desc is not initialized before a static member in module.cc.

gcc/ChangeLog:

	* vec.c (vec_prefix::register_overhead): Initialize vec_mem_desc
	if it is NULL.
	(vec_prefix::release_overhead): Use pointer type for
	vec_mem_desc.
	(dump_vec_loc_statistics): Likewise.
---
 gcc/vec.c | 21 ++++++++++++---------
 1 file changed, 12 insertions(+), 9 deletions(-)

diff --git a/gcc/vec.c b/gcc/vec.c
index a28899170ed..7df0f54605e 100644
--- a/gcc/vec.c
+++ b/gcc/vec.c
@@ -121,7 +121,7 @@ public:
 };
 
 /* Vector memory description.  */
-static mem_alloc_description <vec_usage> vec_mem_desc;
+static mem_alloc_description <vec_usage> *vec_mem_desc = NULL;
 
 /* Account the overhead.  */
 
@@ -129,10 +129,13 @@ void
 vec_prefix::register_overhead (void *ptr, size_t elements,
 			       size_t element_size MEM_STAT_DECL)
 {
-  vec_mem_desc.register_descriptor (ptr, VEC_ORIGIN, false
-				    FINAL_PASS_MEM_STAT);
+  if (vec_mem_desc == NULL)
+    vec_mem_desc = new mem_alloc_description<vec_usage> ();
+
+  vec_mem_desc->register_descriptor (ptr, VEC_ORIGIN, false
+				     FINAL_PASS_MEM_STAT);
   vec_usage *usage
-    = vec_mem_desc.register_instance_overhead (elements * element_size, ptr);
+    = vec_mem_desc->register_instance_overhead (elements * element_size, ptr);
   usage->m_element_size = element_size;
   usage->m_items += elements;
   if (usage->m_items_peak < usage->m_items)
@@ -145,11 +148,11 @@ void
 vec_prefix::release_overhead (void *ptr, size_t size, size_t elements,
 			      bool in_dtor MEM_STAT_DECL)
 {
-  if (!vec_mem_desc.contains_descriptor_for_instance (ptr))
-    vec_mem_desc.register_descriptor (ptr, VEC_ORIGIN,
+  if (!vec_mem_desc->contains_descriptor_for_instance (ptr))
+    vec_mem_desc->register_descriptor (ptr, VEC_ORIGIN,
 				      false FINAL_PASS_MEM_STAT);
-  vec_usage *usage = vec_mem_desc.release_instance_overhead (ptr, size,
-							     in_dtor);
+  vec_usage *usage = vec_mem_desc->release_instance_overhead (ptr, size,
+							      in_dtor);
   usage->m_items -= elements;
 }
 
@@ -183,7 +186,7 @@ vec_prefix::calculate_allocation_1 (unsigned alloc, unsigned desired)
 void
 dump_vec_loc_statistics (void)
 {
-  vec_mem_desc.dump (VEC_ORIGIN);
+  vec_mem_desc->dump (VEC_ORIGIN);
 }
 
 #if CHECKING_P
-- 
2.29.2

Re: [PATCH] Fix --enable-gather-detailed-mem-stats build.

[Richard Biener]

On Wed, Dec 16, 2020 at 10:52 AM Martin Liška <mliska@suse.cz> wrote:
>
> On 12/16/20 10:38 AM, Rainer Orth wrote:
> > Hi Jakub,
> >
> >> On Wed, Dec 16, 2020 at 10:20:09AM +0100, Martin Liška wrote:
> >>> So vec_mem_desc is not initialized before a static member in module.cc.
> >>> We can fix it by using constructor attribute.
> > [...]
> >>> +   of all static variables.  */
> >>> +
> >>> +static void
> >>> +__attribute__((constructor (101)))
> >>
> >> I think this needs to be guarded based on which compiler is used to compile
> >> GCC.  Perhaps we could say that we don't support
> >> --enable-gather-detailed-mem-stats when the compiler isn't built by GCC (or
> >> other compiler that supports the constructor attribute) and #error on that.
> >
> > not only that: if a target doesn't support constructor priorities (like
> > Solaris 11.3, unlike 11.4), this makes gcc error out.
> >
> >       Rainer
> >
>
> I see, I'm then suggesting version 3 of the patch that does not depend
> on a constructor.
>
> Thoughts?

I guess that works.  Note we're trying to limit the number of dynamic
initializers so the new loc_spans one would be another candidate to
dynamically allocate instead.

But OK.

Thanks,
Richard.

> Thanks,
> Martin
>
> Martin