On Thu, 15 Mar 2012, Richard Guenther wrote:
>
> This removes the use of TREE_LISTs for VECTOR_CSTs and instead employs
> a similar way of storing elements as TREE_VECs. I copied the
> macro interface bits of the CONSTRUCTOR accesses and did a 1:1 transform
> at most places to not let refactoring errors creep in (well, where
> possible). I'm not sure if it's worth omitting the explicit length
> field in favor of using that of the type - at least we are getting
> consistency (no implicit zero elements) here for free.
>
> Bootstrap and regtest on x86_64-unknown-linux-gnu running. I cannot
> test sparc apart from compiling a cc1, which works.
>
> Ok for the c-common and sparc bits? Any other comments?
Bootstrapped ok on x86_64-unknown-linux-gnu. Incremental patch to
fix issues revealed by testing below.
Richard.
Index: gcc/tree.c
===================================================================
--- gcc/tree.c.orig 2012-03-15 14:02:31.000000000 +0100
+++ gcc/tree.c 2012-03-15 13:51:27.000000000 +0100
@@ -1328,6 +1328,7 @@ build_vector_stat (tree type, tree *vals
v = ggc_alloc_zone_cleared_tree_node_stat (&tree_zone, length PASS_MEM_STAT);
TREE_SET_CODE (v, VECTOR_CST);
+ TREE_CONSTANT (v) = 1;
TREE_TYPE (v) = type;
/* Iterate through elements and check for overflow. */
Index: gcc/varasm.c
===================================================================
--- gcc/varasm.c.orig 2012-03-15 14:02:31.000000000 +0100
+++ gcc/varasm.c 2012-03-15 13:59:37.000000000 +0100
@@ -4596,8 +4596,8 @@ output_constant (tree exp, unsigned HOST
elt_size = GET_MODE_SIZE (inner);
- thissize = 0;
output_constant (VECTOR_CST_ELT (exp, 0), elt_size, align);
+ thissize = elt_size;
for (i = 1; i < VECTOR_CST_NELTS (exp); ++i)
{
output_constant (VECTOR_CST_ELT (exp, i), elt_size, nalign);
