rs6000_hard_regno_mode_ok_uncached result depends on target flags,
and nowadays it is possible that TARGET_OPTION_OVERRIDE be called not
just at the start of compilation but per-function by the action of
function attributes or "#pragma GCC target". At the start of
compilation rs6000_hard_regno_mode_ok_p[][] is a clean slate, but
not so later on.
Bootstrapped etc. powerpc64le-linux. If gcc wasn't in stage 4 I'd
apply this as obvious, but I doubt it is a regression, and the patch
is just a tiny bit more than a doc fix. So, OK now or wait for
stage 1?
* config/rs6000/rs6000.c (rs6000_init_hard_regno_mode_ok): Always
assign rs6000_hard_regno_mode_ok_p[m][r]. Formatting.
diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
index 06ce2892bf9..f132c3a27c8 100644
--- a/gcc/config/rs6000/rs6000.c
+++ b/gcc/config/rs6000/rs6000.c
@@ -3463,13 +3463,13 @@ rs6000_init_hard_regno_mode_ok (bool global_init_p)
for (r = 0; HARD_REGISTER_NUM_P (r); ++r)
for (m = 0; m < NUM_MACHINE_MODES; ++m)
rs6000_hard_regno_nregs[m][r]
- = rs6000_hard_regno_nregs_internal (r, (machine_mode)m);
+ = rs6000_hard_regno_nregs_internal (r, (machine_mode) m);
/* Precalculate TARGET_HARD_REGNO_MODE_OK. */
for (r = 0; HARD_REGISTER_NUM_P (r); ++r)
for (m = 0; m < NUM_MACHINE_MODES; ++m)
- if (rs6000_hard_regno_mode_ok_uncached (r, (machine_mode)m))
- rs6000_hard_regno_mode_ok_p[m][r] = true;
+ rs6000_hard_regno_mode_ok_p[m][r]
+ = rs6000_hard_regno_mode_ok_uncached (r, (machine_mode) m);
/* Precalculate CLASS_MAX_NREGS sizes. */
for (c = 0; c < LIM_REG_CLASSES; ++c)
--
Alan Modra
Australia Development Lab, IBM
