From: Michiel Derhaeg <[email protected]>

<voptab>v2si3 advertised V2SImode optabs (mul/div/smin/smax) whenever the
scalar op existed, but V2SImode is only a supported vector mode under
TARGET_PLUS_QMACW.

Since r15-5493-g485ab50c204ddb ("Add helpers to test whether an optab
can be implemented") the middle-end keeps the operation vectorized on
such cores.  The operands degrade to DImode and it is expanded as a
double-word scalar op: mul and div silently compute on the packed pair,
min and max ICE in expand_expr_real_2 for lack of a DImode expansion.

Require TARGET_PLUS_QMACW so the optab matches
arc_vector_mode_supported_p.

No tests added because this is already covered by existing tests.

Every core without TARGET_PLUS_QMACW, since smin/smax are advertised
unconditionally, e.g. -mcpu=em, ICEs in expand_expr_real_2 on:

        g++.dg/ext/vector21.C    -std=gnu++{98,20,29}
        g++.dg/opt/vectcond-1.C  -std=gnu++{98,20,29}
        g++.dg/pr94920-1.C       -std=c++{98,20,29}

Cores that also have a multiplier or divider but no QMACW (arcem, em4_*,
hs3*, archs) additionally miscompile the following:

        gcc.c-torture/execute/pr23135.c
        gcc.c-torture/execute/pr53645.c
        gcc.c-torture/execute/scal-to-vec2.c
        gcc.c-torture/execute/simd-1.c
        gcc.dg/pr88598-4.c

        PR target/126243

gcc/ChangeLog:
        * config/arc/simdext.md (<voptab>v2si3): Require TARGET_PLUS_QMACW.

Signed-off-by: Michiel Derhaeg <[email protected]>
---
v2: Added test failure info & mention PR

 gcc/config/arc/simdext.md | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/gcc/config/arc/simdext.md b/gcc/config/arc/simdext.md
index 003f2c4e6e6..509ca55d3f2 100644
--- a/gcc/config/arc/simdext.md
+++ b/gcc/config/arc/simdext.md
@@ -2010,7 +2010,7 @@
   [(set (match_operand:V2SI 0 "register_operand" "=r")
        (EMUVEC:V2SI (match_operand:V2SI 1 "register_operand" "r")
                     (match_operand:V2SI 2 "nonmemory_operand" "ri")))]
-  ""
+  "TARGET_PLUS_QMACW"
   "#"
   "reload_completed"
   [(const_int 0)]
-- 
2.54.0

Reply via email to