Combine has this little check:
/* Replacing the operand of a SUBREG or a ZERO_EXTEND with a
CONST_INT is not valid, because after the replacement, the
original mode would be gone. Unfortunately, we can't tell
when do_SUBST is called to replace the operand thereof, so we
perform this test on oldval instead, checking whether an
invalid replacement took place before we got here. */
gcc_assert (!(GET_CODE (oldval) == SUBREG
&& CONST_INT_P (SUBREG_REG (oldval))));
gcc_assert (!(GET_CODE (oldval) == ZERO_EXTEND
&& CONST_INT_P (XEXP (oldval, 0))));
Which indirectly implies that we can never have these in the IL prior to
combine. The ft32 port has recently started regressing a few tests
because it's triggering those assertions. This isn't a huge surprise as
the port allows (zero_extend (const_int)) for a subset of const_int values.
This patch tightens the predicate & constraints which fixes the recent
regression. In fact, it fixes a few dozen long standing failures in the
testsuite on ft32-elf and since the compiler is no longer faulting as
much, there's more execution tests being run as well.
Committed to the trunk.
JEff
commit a8ae23920f9c954ed66bd92ddbdbb1b8486b950f
Author: Stefan Kanthak <stefan.kant...@nexgo.de>
Date: Sun Nov 29 18:05:46 2020 -0700
fixed _bswapsi2 function
libgcc
* libgcc2.c (bswapsi2): Make constants unsigned.
diff --git a/libgcc/libgcc2.c b/libgcc/libgcc2.c
index 1921d80f612..61de3b42e1d 100644
--- a/libgcc/libgcc2.c
+++ b/libgcc/libgcc2.c
@@ -468,10 +468,10 @@ __ashrdi3 (DWtype u, shift_count_type b)
SItype
__bswapsi2 (SItype u)
{
- return ((((u) & 0xff000000) >> 24)
- | (((u) & 0x00ff0000) >> 8)
- | (((u) & 0x0000ff00) << 8)
- | (((u) & 0x000000ff) << 24));
+ return ((((u) & 0xff000000u) >> 24)
+ | (((u) & 0x00ff0000u) >> 8)
+ | (((u) & 0x0000ff00u) << 8)
+ | (((u) & 0x000000ffu) << 24));
}
#endif
#ifdef L_bswapdi2
