[Lldb-commits] [PATCH] D77045: Add invalidate list to primary regs in arm64 register infos

Pavel Labath via Phabricator via lldb-commits lldb-commits at lists.llvm.org
Mon Apr 6 07:02:31 PDT 2020

labath added a comment.

In D77045#1956879 <https://reviews.llvm.org/D77045#1956879>, @omjavaid wrote:

> Adding a testcase would be tricky as these register overlap in memory and we store them with overlapping offsets with their children we should not need to invalidate the children when we write the parent but for some strange unexplainable reason QEMU was behaving strangely and not updating the first half in certain random cases. I just thought invalidation of children will force a read after write for that case.

Thanks for the explanation, but I'm afraid I still don't get what is going on here. Can you walk me through the individual steps here? Something like:

1. user does "register write x0 xxxxxx"
2. lldb translates that to the appropriate `p` packet
3. ???
4. user does "register read w0"
5. bad value comes out because...

Comment at: lldb/source/Plugins/Process/Utility/RegisterInfos_arm64.h:593-594
+#define STRINGIZE2(x) #x
omjavaid wrote:
> labath wrote:
> > What's up with the indirection?
> We want to pass nullptr or register name as is then do stringize on regname only. According to C macro expansion rules if we want our macros expanded first and then do the # then we ll need double indirection.
I've figured it's something like that, though it's still not clear to me why that is needed in this particular case.

However, now I have a different question. As the only thing you're using the alt_name argument for is to turn it into a string, why not just pass a string in the first place:
DEFINE_GPR64(x0, nullptr, ...)
DEFINE_GPR64(fp, "x29", ...)

Comment at: lldb/source/Plugins/Process/Utility/RegisterInfos_arm64.h:597
 static lldb_private::RegisterInfo g_register_infos_arm64_le[] = {
     // DEFINE_GPR64(name, GENERIC KIND)
+    DEFINE_GPR64(x0, nullptr, LLDB_REGNUM_GENERIC_ARG1, g_x0_invalidates),
This comment is out of date now.



More information about the lldb-commits mailing list