From 49bb5a7e06049c771a28a2cc507c7789dc54a88b Mon Sep 17 00:00:00 2001 From: Michael Matz Date: Sat, 3 Sep 2016 23:23:13 +0200 Subject: Fix __builtin_constant_p(1000/x) was incorrectly treated as constant because the vpop removed all traces of non-constness. --- tccgen.c | 17 +++++++++++++++++ 1 file changed, 17 insertions(+) (limited to 'tccgen.c') diff --git a/tccgen.c b/tccgen.c index cd6e690..f79d30e 100644 --- a/tccgen.c +++ b/tccgen.c @@ -1787,6 +1787,13 @@ static void gen_opic(int op) gen_opi(op); } else { vtop--; + /* Ensure vtop isn't marked VT_CONST in case something + up our callchain is interested in const-ness of the + expression. Also make it a non-LVAL if it was, + so that further code can't accidentally generate + a deref (happen only for buggy uses of e.g. + gv() under nocode_wanted). */ + vtop->r &= ~(VT_VALMASK | VT_LVAL); } } } @@ -5014,6 +5021,16 @@ static void expr_cond(void) } } else { + /* XXX This doesn't handle nocode_wanted correctly at all. + It unconditionally calls gv/gvtst and friends. That's + the case for many of the expr_ routines. Currently + that should generate only useless code, but depending + on other operand handling this might also generate + pointer derefs for lvalue conversions whose result + is useless, but nevertheless can lead to segfault. + + Somewhen we need to overhaul the whole nocode_wanted + handling. */ if (vtop != vstack) { /* needed to avoid having different registers saved in each branch */ -- cgit v1.3.1