[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [Tinycc-devel] Largish mob update from me

From: Michael Matz
Subject: Re: [Tinycc-devel] Largish mob update from me
Date: Tue, 20 Dec 2016 05:59:37 +0100 (CET)
User-agent: Alpine 2.20 (LSU 67 2015-01-07)


On Sun, 18 Dec 2016, grischka wrote:

Ok, I guess the best we can do to make sure the assert of above
will never trigger is to tell the compiler to optimize it out,
like so:

   ST_FUNC void g(int c)
       int ind1;
       if (nocode_wanted)
       if (nocode_wanted)
            tcc_error("internal: code generated but nocode_wanted");

Yes, that's of course the nicest thing as it's the most aggressive way to ensure no code is emitted. But I had problems with that that I've not analyzed further back then.

Why I think this is necessary is because I got some regression wrt.
your improved "expr_land/lor()", and I wondered whether maybe the
code was wrong in itself or in its trust on the underlying nocode
machinery.  So to find out I needed a machinery that works ;)

Okay, so I was now forced to indeed analyze the problems (miscompiled the linux kernel) ;) Result:

There aren't any if's with (nocode_wanted) in tccgen.c anywhere
anymore, exept the 3 mentioned in the patch, and it I'll be happy
if that can stay so.

Unfortunately I had to add one other place where it checks nocode_wanted and that's in the ugly dealing with VT_CMP/VT_JMP values in vtop. See testcases in the above, and my comment included here for citing/discussing:

+       Don't do this when nocode_wanted.  vtop might come from
+       !nocode_wanted regions (see 88_codeopt.c) and transforming
+       it to a register without actually generating code is wrong
+       as their value might still be used for real.  All values
+       we push under nocode_wanted will eventually be popped
+       again, so that the VT_CMP/VT_JMP value will be in vtop
+       when code is unsuppressed again.  */
+    if (vtop >= vstack && !nocode_wanted) {
         v = vtop->r & VT_VALMASK;
         if (v == VT_CMP || (v & ~1) == VT_JMP)

I hope I haven't broken anything that you've fixed with your three nocode_wanted patches but lacking testcases I can't be sure (hint hint) :)


reply via email to

[Prev in Thread] Current Thread [Next in Thread]