On 11/3/20 11:16 AM, Richard Biener wrote:
> On Tue, 3 Nov 2020, Bernd Edlinger wrote:
> 
>>
>>
>> On 11/3/20 10:34 AM, Richard Biener wrote:
>>> On Mon, 2 Nov 2020, Bernd Edlinger wrote:
>>>
>>>> On 11/2/20 3:07 PM, Richard Biener wrote:
>>>>> On Mon, 2 Nov 2020, Bernd Edlinger wrote:
>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> this makes sure that stack allocated SSA_NAMEs are
>>>>>> at least MODE_ALIGNED.  Also increase the MEM_ALIGN
>>>>>> for the corresponding rtl objects.
>>>>>>
>>>>>>
>>>>>> Tested on x86_64-pc-linux-gnu and arm-none-eabi.
>>>>>>
>>>>>> OK for trunk?
>>>>>
>>>>>
>>>>> @@ -1022,6 +1030,14 @@ expand_one_stack_var_at (tree decl, rtx base,
>>>>> unsigned base_align,
>>>>>      }
>>>>>  
>>>>>    set_rtl (decl, x);
>>>>> +
>>>>> +  if (TREE_CODE (decl) == SSA_NAME
>>>>> +      && GET_MODE (x) != BLKmode
>>>>> +      && MEM_ALIGN (x) < GET_MODE_ALIGNMENT (GET_MODE (x)))
>>>>> +    {
>>>>> +      gcc_checking_assert (GET_MODE_ALIGNMENT (GET_MODE (x)) <=
>>>>> base_align);
>>>>> +      set_mem_align (x, GET_MODE_ALIGNMENT (GET_MODE (x)));
>>>>> +    }
>>>>>  }
>>>>>  
>>>>>
>>>>> I wonder whether we cannot "fix" set_rtl to not call
>>>>> set_mem_attributes in this path, maybe directly call
>>>>> set_mem_align there instead?  That is, the preceeding
>>>>> code for ! SSA_NAME already tries to adjust alignment
>>>>> to honor that of the actual stack slot - IMHO the
>>>>> non-SSA and SSA cases should be merged after this
>>>>> patch, but maybe simply by calling set_mem_align
>>>>> instead of doing the DECL_ALIGN frobbing and do
>>>>> the alignment compute also for SSA_NAMEs?
>>>>>
>>>>> The other pieces look OK but the above is a bit ugly
>>>>> at the moment.
>>>>>
>>>>
>>>> Hmm, how about this?
>>>
>>> That would work for me.  Guess removing the DECL_ALIGN frobbing
>>> in the != SSA_NAME path didn't work out or you didn't try out
>>> of caution?
>>>
>>
>> I didn't try, since it felt simply more correct this way,
>> and get_object_alignment would probably give a different
>> answer since it uses DECL_ALIGN too.
> 
> OK, I see.
> 
> Richard.
> 

Ok, good.

So this is what I will commit shortly.

Thanks
Bernd.
From e281da96736655c42296806162d1b979dd368544 Mon Sep 17 00:00:00 2001
From: Bernd Edlinger <bernd.edlin...@hotmail.de>
Date: Sun, 1 Nov 2020 07:32:20 +0100
Subject: [PATCH] Fix PR97205

This makes sure that stack allocated SSA_NAMEs are
at least MODE_ALIGNED.  Also increase the MEM_ALIGN
for the corresponding rtl objects.

gcc:
2020-11-03  Bernd Edlinger  <bernd.edlin...@hotmail.de>

	PR target/97205
	* cfgexpand.c (align_local_variable): Make SSA_NAMEs
	at least MODE_ALIGNED.
	(expand_one_stack_var_at): Increase MEM_ALIGN for SSA_NAMEs.

testsuite:
2020-11-03  Bernd Edlinger  <bernd.edlin...@hotmail.de>

	PR target/97205
	* gcc.c-torture/compile/pr97205.c: New test.
---
 gcc/cfgexpand.c                               | 43 ++++++++++++++++-----------
 gcc/testsuite/gcc.c-torture/compile/pr97205.c |  7 +++++
 2 files changed, 33 insertions(+), 17 deletions(-)
 create mode 100644 gcc/testsuite/gcc.c-torture/compile/pr97205.c

diff --git a/gcc/cfgexpand.c b/gcc/cfgexpand.c
index f3f17d3..6c41a7e 100644
--- a/gcc/cfgexpand.c
+++ b/gcc/cfgexpand.c
@@ -366,7 +366,15 @@ align_local_variable (tree decl, bool really_expand)
   unsigned int align;
 
   if (TREE_CODE (decl) == SSA_NAME)
-    align = TYPE_ALIGN (TREE_TYPE (decl));
+    {
+      tree type = TREE_TYPE (decl);
+      machine_mode mode = TYPE_MODE (type);
+
+      align = TYPE_ALIGN (type);
+      if (mode != BLKmode
+	  && align < GET_MODE_ALIGNMENT (mode))
+	align = GET_MODE_ALIGNMENT (mode);
+    }
   else
     {
       align = LOCAL_DECL_ALIGNMENT (decl);
@@ -999,20 +1007,21 @@ expand_one_stack_var_at (tree decl, rtx base, unsigned base_align,
   x = plus_constant (Pmode, base, offset);
   x = gen_rtx_MEM (TREE_CODE (decl) == SSA_NAME
 		   ? TYPE_MODE (TREE_TYPE (decl))
-		   : DECL_MODE (SSAVAR (decl)), x);
+		   : DECL_MODE (decl), x);
+
+  /* Set alignment we actually gave this decl if it isn't an SSA name.
+     If it is we generate stack slots only accidentally so it isn't as
+     important, we'll simply set the alignment directly on the MEM.  */
+
+  if (base == virtual_stack_vars_rtx)
+    offset -= frame_phase;
+  align = known_alignment (offset);
+  align *= BITS_PER_UNIT;
+  if (align == 0 || align > base_align)
+    align = base_align;
 
   if (TREE_CODE (decl) != SSA_NAME)
     {
-      /* Set alignment we actually gave this decl if it isn't an SSA name.
-         If it is we generate stack slots only accidentally so it isn't as
-	 important, we'll simply use the alignment that is already set.  */
-      if (base == virtual_stack_vars_rtx)
-	offset -= frame_phase;
-      align = known_alignment (offset);
-      align *= BITS_PER_UNIT;
-      if (align == 0 || align > base_align)
-	align = base_align;
-
       /* One would think that we could assert that we're not decreasing
 	 alignment here, but (at least) the i386 port does exactly this
 	 via the MINIMUM_ALIGNMENT hook.  */
@@ -1022,6 +1031,8 @@ expand_one_stack_var_at (tree decl, rtx base, unsigned base_align,
     }
 
   set_rtl (decl, x);
+
+  set_mem_align (x, align);
 }
 
 class stack_vars_data
@@ -1327,13 +1338,11 @@ expand_one_stack_var_1 (tree var)
     {
       tree type = TREE_TYPE (var);
       size = tree_to_poly_uint64 (TYPE_SIZE_UNIT (type));
-      byte_align = TYPE_ALIGN_UNIT (type);
     }
   else
-    {
-      size = tree_to_poly_uint64 (DECL_SIZE_UNIT (var));
-      byte_align = align_local_variable (var, true);
-    }
+    size = tree_to_poly_uint64 (DECL_SIZE_UNIT (var));
+
+  byte_align = align_local_variable (var, true);
 
   /* We handle highly aligned variables in expand_stack_vars.  */
   gcc_assert (byte_align * BITS_PER_UNIT <= MAX_SUPPORTED_STACK_ALIGNMENT);
diff --git a/gcc/testsuite/gcc.c-torture/compile/pr97205.c b/gcc/testsuite/gcc.c-torture/compile/pr97205.c
new file mode 100644
index 0000000..6600011
--- /dev/null
+++ b/gcc/testsuite/gcc.c-torture/compile/pr97205.c
@@ -0,0 +1,7 @@
+int a;
+typedef __attribute__((aligned(2))) int x;
+int f ()
+{
+  x b = a;
+  return b;
+}
-- 
1.9.1

Reply via email to