The issue with wrong-code bug PR71976 is that combine.c:get_last_value was
called for a hard register in a wider mode (DImode) than the stored value had
(QImode). The patch introduces a test of the mode precision and only returns
the recorded value if the according mode is not smaller than the mode for which
get_last_value is called.
Segher already tested the patch on 32 different (sub-)architectures of Linux
without any regression, cf.
https://gcc.gnu.org/ml/gcc/2016-07/msg00215.html
Ok for trunk and to backport?
Johann
gcc/
PR rtl-optimization/71976
* combine.c (get_last_value): Return 0 if the argument for which
the function is called has a wider mode than the recorded value.
Index: combine.c
===================================================================
--- combine.c (revision 238849)
+++ combine.c (working copy)
@@ -13210,6 +13210,12 @@ get_last_value (const_rtx x)
&& DF_INSN_LUID (rsp->last_set) >= subst_low_luid)
return 0;
+ /* If fewer bits were set than what we are asked for now, we cannot use
+ the value. */
+ if (GET_MODE_PRECISION (rsp->last_set_mode)
+ < GET_MODE_PRECISION (GET_MODE (x)))
+ return 0;
+
/* If the value has all its registers valid, return it. */
if (get_last_value_validate (&value, rsp->last_set, rsp->last_set_label, 0))
return value;
