Drop excess size used for run time allocated stack variables.

* get_dynamic_stack_size is passed a SIZE of a data block (which is
  allocated elsewhere), the SIZE_ALIGN of the SIZE (i.e. the alignment
  of the underlying memory units (e.g. 32 bytes split into 4 times 8
  bytes = 64 bit alignment) and the REQUIRED_ALIGN of the data portion
  of the allocated memory.

* Assuming the function is called with SIZE = 2, SIZE_ALIGN = 8 and
  REQUIRED_ALIGN = 64 it first adds 7 bytes to SIZE -> 9.  This is
  what is needed to have two bytes 8-byte-aligned at some memory
  location without any known alignment.

* Finally round_push is called to round up SIZE to a multiple of the
  stack slot size.

The key to understanding this is that the function assumes that
STACK_DYNMAIC_OFFSET is completely unknown at the time its called
and therefore it does not make assumptions about the alignment of
STACKPOINTER + STACK_DYNMAIC_OFFSET.  The latest patch simply
hard-codes that SP + SDO is supposed to be aligned to at least
stack slot size (and does that in a very complicated way).  Since
there is no guarantee that this is the case on all targets, the
patch is broken.  It may miscalculate a SIZE that is too small in
some cases.

However, on many targets there is some guarantee about the
alignment of SP + SDO even if the actual value of SDO is unknown.
On s390x it's always 8-byte-aligned (stack slot size).  So the
right fix should be to add knowledge about the target's guaranteed
alignment of SP + SDO to the function.  I'm right now testing a
much simpler patch that uses
REGNO_POINTER_ALIGN(VIRTUAL_STACK_DYNAMIC_REGNUM) as the
alignment.

gcc/ChangeLog:

2016-08-23  Dominik Vogt  <vogt@linux.vnet.ibm.com>

	* explow.c (get_dynamic_stack_size): Take known alignment of stack
	pointer + STACK_DYNAMIC_OFFSET into account when calculating the
	size needed.  Correct a typo in a comment.

From-SVN: r239688

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 8b1dc4ba0512470e5e533281ec518fb3379e8659..23cf5bc33cf3ec322aface2bc4763968f14e8118 100644 (file)
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,10 @@
+2016-08-23  Dominik Vogt  <vogt@linux.vnet.ibm.com>
+
+       * explow.c (get_dynamic_stack_size): Take known alignment of stack
+       pointer + STACK_DYNAMIC_OFFSET into account when calculating the size
+       needed.
+       Correct a typo in a comment.
+
 2016-08-23  Dominik Vogt  <vogt@linux.vnet.ibm.com>
 
        * config/s390/s390.md ("*andc_split"): New splitter for and with

diff --git a/gcc/explow.c b/gcc/explow.c
index a345690b810873601d5f9aff04c233d9a86acaf6..f97a214f655e483545f642c61cb74cfedc962a4b 100644 (file)
--- a/gcc/explow.c
+++ b/gcc/explow.c
@@ -1224,9 +1224,15 @@ get_dynamic_stack_size (rtx *psize, unsigned size_align,
      example), so we must preventively align the value.  We leave space
      in SIZE for the hole that might result from the alignment operation.  */
 
-  extra = (required_align - BITS_PER_UNIT) / BITS_PER_UNIT;
-  size = plus_constant (Pmode, size, extra);
-  size = force_operand (size, NULL_RTX);
+  unsigned known_align = REGNO_POINTER_ALIGN (VIRTUAL_STACK_DYNAMIC_REGNUM);
+  if (known_align == 0)
+    known_align = BITS_PER_UNIT;
+  if (required_align > known_align)
+    {
+      extra = (required_align - known_align) / BITS_PER_UNIT;
+      size = plus_constant (Pmode, size, extra);
+      size = force_operand (size, NULL_RTX);
+    }
 
   if (flag_stack_usage_info && pstack_usage_size)
     *pstack_usage_size += extra;
@@ -1235,7 +1241,7 @@ get_dynamic_stack_size (rtx *psize, unsigned size_align,
     size_align = BITS_PER_UNIT;
 
   /* Round the size to a multiple of the required stack alignment.
-     Since the stack if presumed to be rounded before this allocation,
+     Since the stack is presumed to be rounded before this allocation,
      this will maintain the required alignment.
 
      If the stack grows downward, we could save an insn by subtracting