On 7/18/21 4:12 AM, Bruno Haible wrote:
The purpose of the test is to verify that the compiler does not eliminate
a call to explicit_bzero, even if data flow analysis reveals that the stack
area is "dead" at the end of the function.
With this patch, it was turned into a weaker assertion: namely, that the
compiler does not eliminate a call to explicit_bzero if it cannot make
inferences about the pointer argument.
Strictly speaking, neither assertion is weaker than the other. However,
I take your point that the patch changes the meaning of the test in an
undesirable way. I installed the attached to implement your suggestion.
This new diagnostic points out a problem with the test, though. If GCC
can determine that memcmp reads uninitialized storage, GCC can optimize
away the memcmp and act as if memcmp returns 0 (or any other constant).
So the test as it stands is problematic given recent advances in
practical compilers.
As an aside, I don't understand the discussion of asynchronous signal
invocations in that test's commentary. There is no asynchronous
signaling in that code.
(These points are of course low-priority, as explicit_bzero is
documented to be best-effort as opposed to being a guarantee that the
information is erased.)
>From f9803478355d038aa060d71bdd9eddf2bd43325f Mon Sep 17 00:00:00 2001
From: Paul Eggert <egg...@cs.ucla.edu>
Date: Sun, 18 Jul 2021 14:08:56 -0500
Subject: [PATCH] explicit_bzero-tests: pacify GCC better
Problem reported by Bruno Haible in:
https://lists.gnu.org/r/bug-gnulib/2021-07/msg00039.html
* tests/test-explicit_bzero.c: Ignore -Wmaybe-uninitialized.
(stackbuf): Remove this static pointer, reverting recent change.
(do_secret_stuff, test_stack): Revert these related changes too.
---
ChangeLog | 9 +++++++++
tests/test-explicit_bzero.c | 16 +++++++++++-----
2 files changed, 20 insertions(+), 5 deletions(-)
diff --git a/ChangeLog b/ChangeLog
index c50808437..d175c39af 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,3 +1,12 @@
+2021-07-18 Paul Eggert <egg...@cs.ucla.edu>
+
+ explicit_bzero-tests: pacify GCC better
+ Problem reported by Bruno Haible in:
+ https://lists.gnu.org/r/bug-gnulib/2021-07/msg00039.html
+ * tests/test-explicit_bzero.c: Ignore -Wmaybe-uninitialized.
+ (stackbuf): Remove this static pointer, reverting recent change.
+ (do_secret_stuff, test_stack): Revert these related changes too.
+
2021-07-17 Paul Eggert <egg...@cs.ucla.edu>
memrchr-tests: pacify GCC
diff --git a/tests/test-explicit_bzero.c b/tests/test-explicit_bzero.c
index c42aba93f..14f0ead2b 100644
--- a/tests/test-explicit_bzero.c
+++ b/tests/test-explicit_bzero.c
@@ -32,6 +32,12 @@ SIGNATURE_CHECK (explicit_bzero, void, (void *, size_t));
#include "vma-iter.h"
#include "macros.h"
+/* Suppress GCC warning that do_secret_stuff (2) reads uninitialized
+ local storage. */
+#if 4 < __GNUC__ + (3 <= __GNUC_MINOR__)
+# pragma GCC diagnostic ignored "-Wmaybe-uninitialized"
+#endif
+
#define SECRET "xyzzy1729"
#define SECRET_SIZE 9
@@ -126,12 +132,14 @@ test_heap (void)
/* There are two passes:
1. Put a secret in memory and invoke explicit_bzero on it.
2. Verify that the memory has been erased.
- Access the memory via a volatile pointer, so the compiler
- does not assume the pointer's value and optimize away accesses. */
-static char *volatile stackbuf;
+ Implement them in the same function, so that they access the same memory
+ range on the stack. That way, the test verifies that the compiler
+ does not eliminate a call to explicit_bzero, even if data flow analysis
+ reveals that the stack area is dead at the end of the function. */
static int _GL_ATTRIBUTE_NOINLINE
do_secret_stuff (volatile int pass)
{
+ char stackbuf[SECRET_SIZE];
if (pass == 1)
{
memcpy (stackbuf, SECRET, SECRET_SIZE);
@@ -147,8 +155,6 @@ do_secret_stuff (volatile int pass)
static void
test_stack (void)
{
- char stack_buffer[SECRET_SIZE];
- stackbuf = stack_buffer;
int count = 0;
int repeat;
--
2.25.1
