Hi!
This patch adds verification that vec_select in *.md files
doesn't have any out of bounds indices in the selector.
Bootstrapped/regtested on x86_64-linux and i686-linux, tested
by building cc1 of aarch64, arm, powerpc64le and s390x cross,
and tested by hacking up sse.md to have an out of bounds access there,
ok for trunk?
2017-10-11 Jakub Jelinek <ja...@redhat.com>
* genrecog.c (validate_pattern): For VEC_SELECT verify that
CONST_INT selectors are 0 to GET_MODE_NUNITS (imode) - 1.
--- gcc/genrecog.c.jj 2017-09-01 09:25:40.000000000 +0200
+++ gcc/genrecog.c 2017-10-11 17:53:20.107198630 +0200
@@ -751,6 +751,21 @@ validate_pattern (rtx pattern, md_rtx_in
error_at (info->loc,
"vec_select parallel with %d elements, expected %d",
XVECLEN (XEXP (pattern, 1), 0), expected);
+ else if (VECTOR_MODE_P (imode))
+ {
+ unsigned int nelems = GET_MODE_NUNITS (imode);
+ int i;
+ for (i = 0; i < expected; ++i)
+ if (CONST_INT_P (XVECEXP (XEXP (pattern, 1), 0, i))
+ && (UINTVAL (XVECEXP (XEXP (pattern, 1), 0, i))
+ >= nelems))
+ error_at (info->loc,
+ "out of bounds selector %u in vec_select, "
+ "expected at most %u",
+ (unsigned)
+ UINTVAL (XVECEXP (XEXP (pattern, 1), 0, i)),
+ nelems - 1);
+ }
}
if (imode != VOIDmode && !VECTOR_MODE_P (imode))
error_at (info->loc, "%smode of first vec_select operand is not a "
Jakub
