This reverts part of the earlier patch for PR52478 which turns out
to be not necessary. It also revealed that at least switch
expansion builds new arithmetic in signed char - going the -ftrapv
expansion way (where we obviously don't have __addvqi3 as C
promotes everything to int). We probably should be using
unsigned types here.
I chose to do the simple reversion instead of not using
trapv optabs for modes < INT_TYPE_SIZE.
Bootstrapped and tested on x86_64-unknown-linux-gnu, applied.
Richard.
2014-07-29 Richard Biener <rguent...@suse.de>
PR middle-end/52478
* expr.c (expand_expr_real_2): Revert last change.
Index: gcc/expr.c
===================================================================
--- gcc/expr.c (revision 213119)
+++ gcc/expr.c (working copy)
@@ -9212,9 +9212,7 @@ expand_expr_real_2 (sepops ops, rtx targ
if (modifier == EXPAND_STACK_PARM)
target = 0;
temp = expand_binop (mode, this_optab, op0, op1, target,
- unsignedp,
- trapv_binoptab_p (this_optab)
- ? OPTAB_LIB : OPTAB_LIB_WIDEN);
+ unsignedp, OPTAB_LIB_WIDEN);
gcc_assert (temp);
/* Bitwise operations do not need bitfield reduction as we expect their
operands being properly truncated. */
