tree-stdarg.c:check_all_va_list_escapes looks at va_arg accesses
to determine cfun->va_list_gpr_size. The value set is the offset of
the highest access in the va_list area, plus the size of that access.
For big-endian rs6000 this always results in a multiple of the gpr
register size, but that isn't true for little-endian. We currently
round down the size, resulting in gcc.c-torture/execute/20071213-1.c
failure on powerpc64le at -O3.
Bootstrapped and regression tested powerpc64-linux. OK to apply?
* config/rs6000/rs6000.c (setup_incoming_varargs): Round up
va_list_gpr_size.
Index: gcc/config/rs6000/rs6000.c
===================================================================
--- gcc/config/rs6000/rs6000.c (revision 199781)
+++ gcc/config/rs6000/rs6000.c (working copy)
@@ -9371,19 +9371,20 @@
&& cfun->va_list_gpr_size)
{
int nregs = GP_ARG_NUM_REG - first_reg_offset;
+ int n_gpr;
if (va_list_gpr_counter_field)
{
/* V4 va_list_gpr_size counts number of registers needed. */
- if (nregs > cfun->va_list_gpr_size)
- nregs = cfun->va_list_gpr_size;
+ n_gpr = cfun->va_list_gpr_size;
}
else
{
/* char * va_list instead counts number of bytes needed. */
- if (nregs > cfun->va_list_gpr_size / reg_size)
- nregs = cfun->va_list_gpr_size / reg_size;
+ n_gpr = (cfun->va_list_gpr_size + reg_size - 1) / reg_size;
}
+ if (nregs > n_gpr)
+ nregs = n_gpr;
mem = gen_rtx_MEM (BLKmode,
plus_constant (Pmode, save_area,
--
Alan Modra
Australia Development Lab, IBM
