Segher suggested that the change to favor (plus (ashift (...))) as a split point be conditionalized on the shift count being a constant -- seems quite reasonable to me.
Bootstrapped and regression tested on x86_64-linux-gnu as well as verified the 300+ test files generate the same code before/after on the PA and hppa.exp passes.
Installed on the trunk. Jeff
diff --git a/gcc/ChangeLog b/gcc/ChangeLog index f02ba5c..bf90c86 100644 --- a/gcc/ChangeLog +++ b/gcc/ChangeLog @@ -1,5 +1,8 @@ 2015-05-26 Jeff Law <l...@redhat.com> + * combine.c (find_split_point): Verify that the shift count is a + constant when choosing (plus (ashift ...)) as a split point. + * tree-ssa-threadupdate.c: Replace 8 space sequences with tabs. No functional changes. diff --git a/gcc/combine.c b/gcc/combine.c index 73d141e..55a2e35 100644 --- a/gcc/combine.c +++ b/gcc/combine.c @@ -5157,7 +5157,8 @@ find_split_point (rtx *loc, rtx_insn *insn, bool set_src) worthless to try this split. */ if (!set_src && (GET_CODE (XEXP (x, 0)) == MULT - || GET_CODE (XEXP (x, 0)) == ASHIFT)) + || (GET_CODE (XEXP (x, 0)) == ASHIFT + && GET_CODE (XEXP (XEXP (x, 0), 1)) == CONST_INT))) return loc; default: