Hi!
As mentioned in the PR, aarch64 used to allow VOIDmode CONST_INTs
as %aN operands, but r255230 started ICEing on it and r257907 turned
that ICE into error (output_operand_lossage).
The following patch restores the previous behavior, by allowing such
CONST_INTs through. They will fail aarch64_classify_address a few lines
later and so aarch64_print_address_internal will return false and either
cause output_operand_lossage there, or if it is aarch64_print_address,
let the generic code handle the constant.
Bootstrapped/regtested on aarch64-linux on GCCFarm, ok for trunk?
2018-11-29 Jakub Jelinek <ja...@redhat.com>
PR target/87598
* config/aarch64/aarch64.c (aarch64_print_address_internal): Don't
call output_operand_lossage on VOIDmode CONST_INTs. After
output_operand_lossage do return false.
* gcc.target/aarch64/asm-5.c: New test.
--- gcc/config/aarch64/aarch64.c.jj 2018-11-26 22:21:24.891607602 +0100
+++ gcc/config/aarch64/aarch64.c 2018-11-27 14:16:48.586358824 +0100
@@ -7635,8 +7635,14 @@ aarch64_print_address_internal (FILE *f,
unsigned int size;
/* Check all addresses are Pmode - including ILP32. */
- if (GET_MODE (x) != Pmode)
- output_operand_lossage ("invalid address mode");
+ if (GET_MODE (x) != Pmode
+ && (GET_MODE (x) != VOIDmode
+ || !CONST_INT_P (x)
+ || trunc_int_for_mode (INTVAL (x), Pmode) != INTVAL (x)))
+ {
+ output_operand_lossage ("invalid address mode");
+ return false;
+ }
if (aarch64_classify_address (&addr, x, mode, true, type))
switch (addr.type)
--- gcc/testsuite/gcc.target/aarch64/asm-5.c.jj 2018-11-27 14:24:18.774957407
+0100
+++ gcc/testsuite/gcc.target/aarch64/asm-5.c 2018-11-27 14:19:01.290176187
+0100
@@ -0,0 +1,8 @@
+/* PR target/87598 */
+/* { dg-do compile } */
+
+void
+foo (void)
+{
+ __asm__ ("# %a0" : : "i" (0));
+}
Jakub
