...which should have been after the early exit for non-vectorised
statements. (Actually it could be after the !vec_stmt test in current
sources, but one of the later SVE patches needs it to be before, and
that's also the standard position in other vectorizable_* routines.)
Tested on x86_64-linux-gnu and applied as obvious.
It might be difficult to construct a reliable C or Fortran test for this,
since it only triggers if we miss a sinking opportunity. A gimple FE
test might work though. I'll try that when I have more time, but I
didn't want to hold the patch up until then.
Thanks,
Richard
2017-09-16 Richard Sandiford <richard.sandif...@linaro.org>
gcc/
PR tree-optimization/82228
* tree-vect-loop.c (vectorizable_live_operation): Move initialization
of ncopies.
Index: gcc/tree-vect-loop.c
===================================================================
--- gcc/tree-vect-loop.c 2017-09-15 13:03:19.144141117 +0100
+++ gcc/tree-vect-loop.c 2017-09-16 21:30:38.351984079 +0100
@@ -7029,11 +7029,6 @@ vectorizable_live_operation (gimple *stm
gcc_assert (STMT_VINFO_LIVE_P (stmt_info));
- if (slp_node)
- ncopies = 1;
- else
- ncopies = vect_get_num_copies (loop_vinfo, vectype);
-
if (STMT_VINFO_DEF_TYPE (stmt_info) == vect_reduction_def)
return false;
@@ -7054,6 +7049,11 @@ vectorizable_live_operation (gimple *stm
return true;
}
+ if (slp_node)
+ ncopies = 1;
+ else
+ ncopies = vect_get_num_copies (loop_vinfo, vectype);
+
if (!vec_stmt)
/* No transformation required. */
return true;
