https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92433
--- Comment #4 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
Reduced testcase that results in the warning also on x86_64-linux with -O2
-Wall:
struct S { void *p; struct S *q; };
void bar (int, ...);
void
foo (struct S *x, int n, int y)
{
void *arg_type[3];
for (int i = 0; i < n; i++)
{
arg_type[i] = x->p;
x = x->q;
}
if (y)
{
void *t = arg_type[2]; arg_type[2] = arg_type[1]; arg_type[1] = t;
}
switch (n)
{
case 0: bar (0); break;
case 1: bar (1, arg_type[0]); break;
case 2: bar (2, arg_type[0], arg_type[1]); break;
case 3: bar (3, arg_type[0], arg_type[1], arg_type[2]); break;
}
}
Using if (n == 3 && y) makes the warning go away.
The compiler indeed doesn't know if y implies n == 3, although the caller
guarantees it.
So, we could do:
--- gcc/config/rs6000/rs6000-c.c.jj 2019-08-27 12:26:30.115019661 +0200
+++ gcc/config/rs6000/rs6000-c.c 2019-11-11 10:12:00.954282097 +0100
@@ -6076,13 +6076,13 @@ altivec_build_resolved_builtin (tree *ar
condition (LT vs. EQ, which is recognizable by bit 1 of the first
argument) is reversed. Patch the arguments here before building
the resolved CALL_EXPR. */
- if (desc->code == ALTIVEC_BUILTIN_VEC_VCMPGE_P
+ if (n == 3
+ && desc->code == ALTIVEC_BUILTIN_VEC_VCMPGE_P
&& desc->overloaded_code != ALTIVEC_BUILTIN_VCMPGEFP_P
&& desc->overloaded_code != VSX_BUILTIN_XVCMPGEDP_P)
{
- tree t;
- t = args[2], args[2] = args[1], args[1] = t;
- t = arg_type[2], arg_type[2] = arg_type[1], arg_type[1] = t;
+ std::swap (args[1], args[2]);
+ std::swap (arg_type[1], arg_type[2]);
args[0] = fold_build2 (BIT_XOR_EXPR, TREE_TYPE (args[0]), args[0],
build_int_cst (NULL_TREE, 2));
to make it clear to the compiler that VCMPGE_P has 3 arguments and that it is
actually safe to use all of them, or of course we could clear arg_type first.
I don't find the may be used uninitialized warning wrong (at least not
completely) when the compiler sees a possible code path where it is
uninitialized (desc->code == ALTIVEC_BULTIN_VEC_VCMPGE_P with n < 3).
