Already approved in bugzilla and bootstrapped on x86_64-linux-gnu.
Pushed.
expand_simple_binop() is allowed to allocate a new pseudo-register and
return it, instead of forcing the result into the provided
pseudo-register. This can cause a problem when we expand the unrolled
loop for __builtin_strcmp: the compiler always generates code for all n
iterations of the loop, so "result" will be an alias of the
pseudo-register allocated and used in the last iteration; but at runtime
the loop can break early, causing this pseudo-register uninitialized.
Emit a move instruction in the iteration to force the difference into
one register which has been allocated before the loop, to avoid this
issue.
gcc/ChangeLog:
PR other/109086
* builtins.cc (inline_string_cmp): Force the character
difference into "result" pseudo-register, instead of reassign
the pseudo-register.
---
gcc/builtins.cc | 12 ++++++++++--
1 file changed, 10 insertions(+), 2 deletions(-)
diff --git a/gcc/builtins.cc b/gcc/builtins.cc
index 305c65c29be..90246e214d6 100644
--- a/gcc/builtins.cc
+++ b/gcc/builtins.cc
@@ -7142,8 +7142,16 @@ inline_string_cmp (rtx target, tree var_str, const char
*const_str,
op0 = convert_modes (mode, unit_mode, op0, 1);
op1 = convert_modes (mode, unit_mode, op1, 1);
- result = expand_simple_binop (mode, MINUS, op0, op1,
- result, 1, OPTAB_WIDEN);
+ rtx diff = expand_simple_binop (mode, MINUS, op0, op1,
+ result, 1, OPTAB_WIDEN);
+
+ /* Force the difference into result register. We cannot reassign
+ result here ("result = diff") or we may end up returning
+ uninitialized result when expand_simple_binop allocates a new
+ pseudo-register for returning. */
+ if (diff != result)
+ emit_move_insn (result, diff);
+
if (i < length - 1)
emit_cmp_and_jump_insns (result, CONST0_RTX (mode), NE, NULL_RTX,
mode, true, ne_label);
--
2.40.0
