Hi!
If we discover some bad inline-asm during reg-stack processing and we
error on those, we replace that inline-asm with a (use (const_int 0))
and therefore the various assumptions of reg-stack pass may not hold.
Seems we already have a couple of spots which are more permissive if
any_malformed_asm is true, this patch just adds another one.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2018-03-05 Jakub Jelinek <ja...@redhat.com>
PR inline-asm/84683
* reg-stack.c (move_for_stack_reg): If any_malformed_asm, avoid
assertion failure.
* g++.dg/ext/pr84683.C: New test.
--- gcc/reg-stack.c.jj 2018-01-03 10:19:55.000000000 +0100
+++ gcc/reg-stack.c 2018-03-05 17:41:15.558415480 +0100
@@ -1170,7 +1170,8 @@ move_for_stack_reg (rtx_insn *insn, stac
&& XINT (SET_SRC (XVECEXP (pat, 0, 1)), 1) == UNSPEC_TAN)
emit_swap_insn (insn, regstack, dest);
else
- gcc_assert (get_hard_regnum (regstack, dest) < FIRST_STACK_REG);
+ gcc_assert (get_hard_regnum (regstack, dest) < FIRST_STACK_REG
+ || any_malformed_asm);
gcc_assert (regstack->top < REG_STACK_SIZE);
--- gcc/testsuite/g++.dg/ext/pr84683.C.jj 2018-03-05 17:45:32.901475529
+0100
+++ gcc/testsuite/g++.dg/ext/pr84683.C 2018-03-05 17:44:52.527467872 +0100
@@ -0,0 +1,13 @@
+// PR inline-asm/84683
+// { dg-do compile { target i?86-*-* x86_64-*-* } }
+// { dg-options "-O2" }
+
+void
+foo (float b, double c)
+{
+ for (int e = 0; e < 2; e++)
+ {
+ asm volatile ("" : "+f" (c)); // { dg-error "must specify a single
register" }
+ asm ("" : "+rm" (c = b));
+ }
+}
Jakub
