Patchwork [PR43513,1/3] Replace vla with array - Implementation.

login
register
mail settings
Submitter Tom de Vries
Date July 28, 2011, 3:09 p.m.
Message ID <4E317B96.6050908@codesourcery.com>
Download mbox | patch
Permalink /patch/107375/
State New
Headers show

Comments

Tom de Vries - July 28, 2011, 3:09 p.m.
On 07/28/2011 12:22 PM, Richard Guenther wrote:
> On Wed, 27 Jul 2011, Tom de Vries wrote:
> 
>> On 07/27/2011 05:27 PM, Richard Guenther wrote:
>>> On Wed, 27 Jul 2011, Tom de Vries wrote:
>>>
>>>> On 07/27/2011 02:12 PM, Richard Guenther wrote:
>>>>> On Wed, 27 Jul 2011, Tom de Vries wrote:
>>>>>
>>>>>> On 07/27/2011 01:50 PM, Tom de Vries wrote:
>>>>>>> Hi Richard,
>>>>>>>
>>>>>>> I have a patch set for bug 43513 - The stack pointer is adjusted twice.
>>>>>>>
>>>>>>> 01_pr43513.3.patch
>>>>>>> 02_pr43513.3.test.patch
>>>>>>> 03_pr43513.3.mudflap.patch
>>>>>>>
>>>>>>> The patch set has been bootstrapped and reg-tested on x86_64.
>>>>>>>
>>>>>>> I will sent out the patches individually.
>>>>>>>
>>>>>>
>>>>>> The patch replaces a vla __builtin_alloca that has a constant argument with an
>>>>>> array declaration.
>>>>>>
>>>>>> OK for trunk?
>>>>>
>>>>> I don't think it is safe to try to get at the VLA type the way you do.
>>>>
>>>> I don't understand in what way it's not safe. Do you mean I don't manage to find
>>>> the type always, or that I find the wrong type, or something else?
>>>
>>> I think you might get the wrong type,
>>
>> Ok, I'll review that code one more time.
>>
>>> you also do not transform code
>>> like
>>>
>>>   int *p = alloca(4);
>>>   *p = 3;
>>>
>>> as there is no array type involved here.
>>>
>>
>> I was trying to stay away from non-vla allocas.  A source declared alloca has
>> function livetime, so we could have a single alloca in a loop, called 10 times,
>> with all 10 instances live at the same time. This patch does not detect such
>> cases, and thus stays away from non-vla allocas. A vla decl does not have such
>> problems, the lifetime ends when it goes out of scope.
> 
> Yes indeed - that probably would require more detailed analysis.
> 
>>>>> In fact I would simply do sth like
>>>>>
>>>>>   elem_type = build_nonstandard_integer_type (BITS_PER_UNIT, 1);
>>>>>   n_elem = size * 8 / BITS_PER_UNIT;
>>>>>   array_type = build_array_type_nelts (elem_type, n_elem);
>>>>>   var = create_tmp_var (array_type, NULL);
>>>>>   return fold_convert (TREE_TYPE (lhs), build_fold_addr_expr (var));
>>>>>
>>>>
>>>> I tried this code on the example, and it works, but the newly declared type has
>>>> an 8-bit alignment, while the vla base type has a 32 bit alignment.  This make
>>>> the memory access in the example potentially unaligned, which prohibits an
>>>> ivopts optimization, so the resulting text size is 68 instead of the 64 achieved
>>>> with my current patch.
>>>
>>> Ok, so then set DECL_ALIGN of the variable to something reasonable
>>> like MIN (size * 8, GET_MODE_PRECISION (word_mode)).  Basically the
>>> alignment that the targets alloca function would guarantee.
>>>
>>
>> I tried that, but that doesn't help. It's the alignment of the type that
>> matters, not of the decl.
> 
> It shouldn't.  All accesses are performed with the original types and
> alignment comes from that (plus the underlying decl).
> 

I managed to get it all working by using build_aligned_type rather that DECL_ALIGN.

>> So should we try to find the base type of the vla, and use that, or use the
>> nonstandard char type?
> 
> I don't think we can reliably find the base type of the vla - well,
> in practice we may because we control how we lower VLAs during
> gimplification, but nothing in the IL constraints say that the
> resulting pointer type should be special.
> 
> Using a char[] decl shouldn't be a problem IMHO.
> 
>>>>> And obviously you lose the optimization we arrange with inserting
>>>>> __builtin_stack_save/restore pairs that way - stack space will no
>>>>> longer be shared for subsequent VLAs.  Which means that you'd
>>>>> better limit the size you allow this promotion.
>>>>>
>>>>
>>>> Right, I could introduce a parameter for this.
>>>
>>> I would think you could use PARAM_LARGE_STACK_FRAME for now and say,
>>> allow a size of PARAM_LARGE_STACK_FRAME / 10?
>>>
>>
>> That unfortunately is too small for the example from bug report. The default
>> value of the param is 250, so that would be a threshold of 25, and the alloca
>> size of the example is 40.  Perhaps we can try a threshold of
>> PARAM_LARGE_STACK_FRAME - estimated_stack_size or some such?
> 
> Hm.  estimated_stack_size is not O(1), so no.  I think we need to
> find a sensible way of allowing stack sharing.  Eventually Michas
> patch for introducing points-of-death would help here, if we'd
> go for folding this during stack-save/restore optimization.
> 

I changed the heuristics to this:

+  /* Heuristic: don't fold large vlas.  */
+  threshold = (unsigned HOST_WIDE_INT)PARAM_VALUE (PARAM_LARGE_STACK_FRAME);
+  /* In case a vla is declared at function scope, it has the same lifetime as a
+     declared array, so we allow a larger size.  */
+  block = gimple_block (stmt);
+  if (!(cfun->after_inlining
+        && TREE_CODE (BLOCK_SUPERCONTEXT (block)) == FUNCTION_DECL))
+    threshold /= 10;
+  if (size > threshold)
+    return NULL_TREE;

The heuristics distinguishes between before and after inlining.

After inlining, vla's declared at function scope have the same lifetimes as
declared arrays, and don't share their space. There should be no negative
effects from folding an alloca in this case, but for safety we set a threshold
of PARAM_LARGE_STACK_FRAME.

Before inlining, such a vla might be inlined and share its space with another
vla, so we stick with the normal threshold before inlining.

However, using this heuristic we still don't generate optimal code.

During the first pass_ccp, the folding is not done, because the size (40) is
larger than the threshold 25. The threshold is 25, because inlining is not yet done.

During pass_fold_builtins, the folding is done because it's after inlining, but
it's later than pass_iv_optimize, so that still doesn't yield the optimal size
of 64.

The folding is not done during any of the other invocations or pass_ccp, because
the argument has already become constant in the earlier invocation.

Using this change, I manage to trigger folding during the second invocation of
pass_ccp, before iv_optimize so we generate optimal code.


But, to me it feels like a hack. Do you have any ideas how to do this better?

Attaching untested patch for reference (will test overnight).

Thanks,
- Tom

2011-07-28  Tom de Vries  <tom@codesourcery.com>

	PR middle-end/43513
	* gimple-fold.c (params.h): Include.
	(fold_builtin_alloca): New function.
	(gimple_fold_builtin): Use fold_builtin_alloca.
	* tree-ssa-ccp.c (ccp_fold_stmt): Force folding of vla-related alloca.
	* Makefile.in (gimple-fold.o): Add $(PARAMS_H) to rule.

Patch

Index: gcc/gimple-fold.c
===================================================================
--- gcc/gimple-fold.c (revision 173734)
+++ gcc/gimple-fold.c (working copy)
@@ -31,6 +31,7 @@  along with GCC; see the file COPYING3.  
 #include "tree-ssa-propagate.h"
 #include "target.h"
 #include "gimple-fold.h"
+#include "params.h"
 
 /* Return true when DECL can be referenced from current unit.
    We can get declarations that are not possible to reference for
@@ -1170,6 +1171,54 @@  get_maxval_strlen (tree arg, tree *lengt
     }
 }
 
+/* Detects a vla-related alloca with a constant argument.  Declares fixed-size
+   array and return the address, if found, otherwise returns NULL_TREE.  */
+
+static tree
+fold_builtin_alloca (gimple stmt)
+{
+  unsigned HOST_WIDE_INT size, threshold, n_elem;
+  tree lhs, arg, block, var, elem_type, array_type;
+  unsigned int align;
+
+  /* Get lhs.  */
+  lhs = gimple_call_lhs (stmt);
+  if (lhs == NULL_TREE)
+    return NULL_TREE;
+
+  /* Only handle vla-related allocas for the moment.  We do not yet detect when
+     a source-level alloca can be safely folded.  */
+  if (!gimple_call_alloca_for_var_p (stmt))
+    return NULL_TREE;
+
+  /* Detect constant argument.  */
+  arg = gimple_call_arg (stmt, 0);
+  if (TREE_CODE (arg) != INTEGER_CST || !host_integerp (arg, 1))
+    return NULL_TREE;
+  size = TREE_INT_CST_LOW (arg);
+
+  /* Heuristic: don't fold large vlas.  */
+  threshold = (unsigned HOST_WIDE_INT)PARAM_VALUE (PARAM_LARGE_STACK_FRAME);
+  /* In case a vla is declared at function scope, it has the same lifetime as a
+     declared array, so we allow a larger size.  */
+  block = gimple_block (stmt);
+  if (!(cfun->after_inlining
+        && TREE_CODE (BLOCK_SUPERCONTEXT (block)) == FUNCTION_DECL))
+    threshold /= 10;
+  if (size > threshold)
+    return NULL_TREE;
+
+  /* Declare array.  */
+  elem_type = build_nonstandard_integer_type (BITS_PER_UNIT, 1);
+  n_elem = size * 8 / BITS_PER_UNIT;
+  align = MIN (size * 8, GET_MODE_PRECISION (word_mode));
+  array_type = build_aligned_type (build_array_type_nelts (elem_type, n_elem),
+                                   align);
+  var = create_tmp_var (array_type, NULL);
+
+  /* Fold alloca to the address of the array.  */
+  return fold_convert (TREE_TYPE (lhs), build_fold_addr_expr (var));
+}
 
 /* Fold builtin call in statement STMT.  Returns a simplified tree.
    We may return a non-constant expression, including another call
@@ -1246,6 +1295,8 @@  gimple_fold_builtin (gimple stmt)
       arg_idx = 1;
       type = 2;
       break;
+    case BUILT_IN_ALLOCA:
+      return fold_builtin_alloca (stmt);
     default:
       return NULL_TREE;
     }
Index: gcc/tree-ssa-ccp.c
===================================================================
--- gcc/tree-ssa-ccp.c (revision 173734)
+++ gcc/tree-ssa-ccp.c (working copy)
@@ -1727,6 +1727,13 @@  ccp_fold_stmt (gimple_stmt_iterator *gsi
 	if (gimple_call_internal_p (stmt))
 	  return false;
 
+        /* The heuristic of fold_builtin_alloca differs before and after
+           inlining, so we don't require the arg to be changed into a constant
+           for folding, but just to be constant.  */
+        if (gimple_call_alloca_for_var_p (stmt)
+            && get_constant_value (gimple_call_arg (stmt, 0)))
+          return true;
+
 	/* Propagate into the call arguments.  Compared to replace_uses_in
 	   this can use the argument slot types for type verification
 	   instead of the current argument type.  We also can safely
Index: gcc/Makefile.in
===================================================================
--- gcc/Makefile.in (revision 173734)
+++ gcc/Makefile.in (working copy)
@@ -2672,7 +2672,7 @@  gimple-iterator.o : gimple-iterator.c $(
 gimple-fold.o : gimple-fold.c $(TREE_FLOW_H) $(CONFIG_H) \
    $(SYSTEM_H) $(RTL_H) $(TREE_H) $(TM_P_H) $(EXPR_H) $(GGC_H) output.h \
    $(DIAGNOSTIC_H) $(FUNCTION_H) $(TIMEVAR_H) $(TM_H) coretypes.h \
-   $(TREE_DUMP_H) $(BASIC_BLOCK_H) $(TREE_PASS_H) langhooks.h \
+   $(TREE_DUMP_H) $(BASIC_BLOCK_H) $(TREE_PASS_H) langhooks.h $(PARAMS_H) \
    tree-ssa-propagate.h value-prof.h $(FLAGS_H) $(TARGET_H) gimple-fold.h
 gimple-low.o : gimple-low.c $(CONFIG_H) $(SYSTEM_H) $(TREE_H) \
    $(DIAGNOSTIC_H) $(GIMPLE_H) $(TREE_INLINE_H) langhooks.h \