In cpu_exec() we have a longstanding workaround for compilers which
do not correctly implement the part of the sigsetjmp()/siglongjmp()
spec which requires that local variables which are not changed
between the setjmp and the longjmp retain their value.

I recently ran across the upstream clang bug report for this; add a
link to it to the comment describing the workaround, and generally
expand the comment, so that we have a reasonable chance in future of
understanding why it's there and determining when we can remove it,
assuming clang eventually fixes the bug.

Remove the /* buggy compiler */ comments on the #else and #endif:
they don't add anything to understanding and are somewhat misleading
since they're sandwiching the code path for *non*-buggy compilers.

Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
---
 accel/tcg/cpu-exec.c | 25 +++++++++++++++++++------
 1 file changed, 19 insertions(+), 6 deletions(-)

diff --git a/accel/tcg/cpu-exec.c b/accel/tcg/cpu-exec.c
index 6d017e46dd8..0abdae6f9f3 100644
--- a/accel/tcg/cpu-exec.c
+++ b/accel/tcg/cpu-exec.c
@@ -772,17 +772,30 @@ int cpu_exec(CPUState *cpu)
     /* prepare setjmp context for exception handling */
     if (sigsetjmp(cpu->jmp_env, 0) != 0) {
 #if defined(__clang__)
-        /* Some compilers wrongly smash all local variables after
-         * siglongjmp. There were bug reports for gcc 4.5.0 and clang.
+        /*
+         * Some compilers wrongly smash all local variables after
+         * siglongjmp (the spec requires that only non-volatile locals
+         * which are changed between the sigsetjmp and siglongjmp are
+         * permitted to be trashed). There were bug reports for gcc
+         * 4.5.0 and clang.  The bug is fixed in all versions of gcc
+         * that we support, but is still unfixed in clang:
+         *   https://bugs.llvm.org/show_bug.cgi?id=21183
+         *
          * Reload essential local variables here for those compilers.
-         * Newer versions of gcc would complain about this code (-Wclobbered). */
+         * Newer versions of gcc would complain about this code (-Wclobbered),
+         * so we only perform the workaround for clang.
+         */
         cpu = current_cpu;
         cc = CPU_GET_CLASS(cpu);
-#else /* buggy compiler */
-        /* Assert that the compiler does not smash local variables. */
+#else
+        /*
+         * Non-buggy compilers preserve these locals; assert that
+         * they have the correct value.
+         */
         g_assert(cpu == current_cpu);
         g_assert(cc == CPU_GET_CLASS(cpu));
-#endif /* buggy compiler */
+#endif
+
 #ifndef CONFIG_SOFTMMU
         tcg_debug_assert(!have_mmap_lock());
 #endif
-- 
2.20.1

Peter Maydell <peter.maydell@linaro.org> writes:

> In cpu_exec() we have a longstanding workaround for compilers which
> do not correctly implement the part of the sigsetjmp()/siglongjmp()
> spec which requires that local variables which are not changed
> between the setjmp and the longjmp retain their value.
>
> I recently ran across the upstream clang bug report for this; add a
> link to it to the comment describing the workaround, and generally
> expand the comment, so that we have a reasonable chance in future of
> understanding why it's there and determining when we can remove it,
> assuming clang eventually fixes the bug.
>
> Remove the /* buggy compiler */ comments on the #else and #endif:
> they don't add anything to understanding and are somewhat misleading
> since they're sandwiching the code path for *non*-buggy compilers.
>
> Signed-off-by: Peter Maydell <peter.maydell@linaro.org>

Reviewed-by: Alex Bennée <alex.bennee@linaro.org>

-- 
Alex Bennée

On Fri, 29 Jan 2021 at 15:19, Alex Bennée <alex.bennee@linaro.org> wrote:
>
>
> Peter Maydell <peter.maydell@linaro.org> writes:
>
> > In cpu_exec() we have a longstanding workaround for compilers which
> > do not correctly implement the part of the sigsetjmp()/siglongjmp()
> > spec which requires that local variables which are not changed
> > between the setjmp and the longjmp retain their value.
> >
> > I recently ran across the upstream clang bug report for this; add a
> > link to it to the comment describing the workaround, and generally
> > expand the comment, so that we have a reasonable chance in future of
> > understanding why it's there and determining when we can remove it,
> > assuming clang eventually fixes the bug.
> >
> > Remove the /* buggy compiler */ comments on the #else and #endif:
> > they don't add anything to understanding and are somewhat misleading
> > since they're sandwiching the code path for *non*-buggy compilers.
> >
> > Signed-off-by: Peter Maydell <peter.maydell@linaro.org>
>
> Reviewed-by: Alex Bennée <alex.bennee@linaro.org>

Thanks. I'll put this in via target-arm.next unless somebody
has a preference for it going in via some other tree ?

-- PMM