pmu_init() register its event checking the pm_event::supported()
handler. For INST_RETIRED, the event is only registered and the
bit enabled in the PMU Common Event Identification register when
icount is enabled as ICOUNT_PRECISE.
PMU events are TCG-only, hardware accelerators handle them
directly. Unfortunately we register the events in non-TCG builds,
leading to linking error such:
ld: Undefined symbols:
_icount_to_ns, referenced from:
_instructions_ns_per in target_arm_helper.c.o
clang: error: linker command failed with exit code 1 (use -v to see invocation)
As a kludge, give a hint to the compiler by asserting the
pm_event::get_count() and pm_event::ns_per_count() handler will
only be called under this icount mode.
Signed-off-by: Philippe Mathieu-Daudé <philmd@linaro.org>
---
As discussed in
https://lore.kernel.org/qemu-devel/CAFEAcA-HVf8vWLzmdStEo2NrSKQdZV612rBjiaj-gLW4vXyvpA@mail.gmail.com/
better would be to restrict the PMU events to TCG, but this is
out of the scope of this series.
---
target/arm/helper.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/target/arm/helper.c b/target/arm/helper.c
index adb0960bba..333fd5f4bf 100644
--- a/target/arm/helper.c
+++ b/target/arm/helper.c
@@ -940,11 +940,13 @@ static bool instructions_supported(CPUARMState *env)
static uint64_t instructions_get_count(CPUARMState *env)
{
+ assert(icount_enabled() == ICOUNT_PRECISE);
return (uint64_t)icount_get_raw();
}
static int64_t instructions_ns_per(uint64_t icount)
{
+ assert(icount_enabled() == ICOUNT_PRECISE);
return icount_to_ns((int64_t)icount);
}
#endif
--
2.41.0