[Bug target/114801] [14/15 Regression] arm: ICE in find_cached_value, at rtx-vector-builder.cc:100 with MVE intrinsics

["jakub at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114801

--- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Guess the primary question is why there is the gen_lowpart call at all.
Is it that normally the mode of x is already right due to the prototypes of the
builtins, with the exception that gcc likes to promote QImode/HImode arguments
of calls to SImode, so is the intent in that case to just narrow down SImode
back to HImode (seems VALID_MVE_PRED_MODE is only true for HImode from scalar
MODE_INT modes)?

If so, best would be to limit the call to just that case.
So (completely untested):
--- gcc/config/arm/arm-mve-builtins.cc.jj       2024-03-19 09:51:05.203631194
+0100
+++ gcc/config/arm/arm-mve-builtins.cc  2024-04-26 15:49:55.380344060 +0200
@@ -2100,7 +2100,12 @@ function_expander::add_input_operand (in
       mode = GET_MODE (x);
     }
   else if (VALID_MVE_PRED_MODE (mode))
-    x = gen_lowpart (mode, x);
+    {
+      if (mode == HImode && GET_MODE (x) != HImode)
+       /* GCC promotes QI/HImode arguments to int, undo that
+          here.  */
+       x = lowpart_subreg (mode, x, SImode);
+    }

   m_ops.safe_grow (m_ops.length () + 1, true);
   create_input_operand (&m_ops.last (), x, mode);

I'd hope if the argument is a vector mode x already has that mode...