https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118362
--- Comment #1 from Stefan Schulze Frielinghaus <stefansf at gcc dot gnu.org>
---
Yikes, the optimization should only apply for constant vectors which are
supported by the hardware. That means vectors up to 16 byte. For
s390_constant_via_vgm_p() and s390_constant_via_vrepi_p() we utilize
s390_constant_via_vgm_vrepi_1() and bail out for any constant vector larger
than 16 byte. For s390_constant_via_vgbm_p() I have been missing this.
Something along the lines
diff --git a/gcc/config/s390/s390.cc b/gcc/config/s390/s390.cc
index 918a2cd6c6d..08acb69de3e 100644
--- a/gcc/config/s390/s390.cc
+++ b/gcc/config/s390/s390.cc
@@ -2818,7 +2818,7 @@ s390_constant_via_vgbm_p (rtx op, unsigned *mask)
unsigned tmp_mask = 0;
int nunit, unit_size;
- if (GET_CODE (op) == CONST_VECTOR)
+ if (GET_CODE (op) == CONST_VECTOR && GET_MODE_SIZE (GET_MODE (op)) <= 16)
{
if (GET_MODE_INNER (GET_MODE (op)) == TImode
|| GET_MODE_INNER (GET_MODE (op)) == TFmode)
should fix it (maybe a power-of-two check should be added, too). I will
revisit this tomorrow. Sorry for the hassle.
