| Message ID | 40551004.PRfSaV3QlZ@polaris |
|---|---|
| State | New |
| Headers | show |
Index: optabs.c =================================================================== --- optabs.c (revision 250226) +++ optabs.c (working copy) @@ -3844,9 +3844,9 @@ prepare_cmp_insn (rtx x, rtx y, enum rtx if (cfun->can_throw_non_call_exceptions) { if (may_trap_p (x)) - x = force_reg (mode, x); + x = copy_to_reg (x); if (may_trap_p (y)) - y = force_reg (mode, y); + y = copy_to_reg (y); } if (GET_MODE_CLASS (mode) == MODE_CC)
This is a regression present on the mainline and 7 branch: the compiler aborts on the attached testcase for x86 at -O2 -march=i686: +===========================GNAT BUG DETECTED==============================+ | 8.0.0 20170716 (experimental) [trunk revision 250237] (x86_64-suse-linux) GCC error:| | in emit_move_insn, at expr.c:3704 | | Error detected around opt65.adb:30:4 because it is trying to generate a move from HImode to SImode. The problem comes from the relatively new code in prepare_cmp_insn: /* Don't allow operands to the compare to trap, as that can put the compare and branch in different basic blocks. */ if (cfun->can_throw_non_call_exceptions) { if (may_trap_p (x)) x = force_reg (mode, x); if (may_trap_p (y)) y = force_reg (mode, y); } which assumes that "mode" is the mode of "x" and "y". While that's true on entry to the function, that can be wrong at this point because just above: x = result; y = const0_rtx; mode = TYPE_MODE (integer_type_node); methods = OPTAB_LIB_WIDEN; unsignedp = false; can change it to something totally different. That's why the attached fix replaces the call to force_reg with a call to copy_to_reg, which should be OK since the mode-less RTXes, i.e. constants, cannot satisfy may_trap_p. Bootstrapped/regtested on x86_64-suse-linux, applied on mainline and 7 branch. 2017-07-16 Eric Botcazou <ebotcazou@adacore.com> PR rtl-optimization/81424 * optabs.c (prepare_cmp_insn): Use copy_to_reg instead of force_reg to remove potential trapping from operands if -fnon-call-exceptions. 2017-07-16 Eric Botcazou <ebotcazou@adacore.com> * gnat.dg/opt65.adb: New test.