[PATCH, i386]: Fix PR59153, ICE: in memory_address_length, at config/i386/i386.c
Uros Bizjak
ubizjak@gmail.com
Mon Nov 18 07:38:00 GMT 2013
Hello!
ix86_decompose_address is called from many places in i386.c, also to
calculate various attributes of the insn (length, etc). The testcase
failed since addsi_1 pattern was declared as TYPE_LEA and its pattern
(involving subregs of SFmode) was passed to length attribute
calculation as memory operand. The failure was in
ix86_address_subreg_operand that rejected non-integer subregs.
ix86_decompose_address should fail only if the *structure* of address
is totaly wrong, and should not bother too much about its operands.
Operand checks should be done in ix86_legitimate_address_p.
The patch moves a couple of check to ix86_legitimate_address_p. The
prevention of non-integer registers in address was already there, so
the check in ix86_address_subreg_operand was not needed anyway. The
check for invalid x86_64 constant address was also already present in
ix86_legitimate_address_p, so corresponding x32 check was also moved
there.
2013-11-17 Uros Bizjak <ubizjak@gmail.com>
PR target/59153
* config/i386/i386.c (ix86_address_subreg_operand): Do not
reject non-integer subregs.
(ix86_decompose_address): Do not reject invalid CONST_INT RTXes.
Move check for invalid x32 constant addresses ...
(ix86_legitimate_address_p): ... here.
testsuite/ChangeLog:
2013-11-17 Uros Bizjak <ubizjak@gmail.com>
PR target/59153
* gcc.target/i386/pr59153.c: New test.
Patch was bootstrapped and regression tested on x86_64-pc-linux-gnu
{,-m32} and committed to mainline. The patch will be backported to
other release branches after a couple of days without problems in
mainline.
Uros.
-------------- next part --------------
Index: config/i386/i386.c
===================================================================
--- config/i386/i386.c (revision 204921)
+++ config/i386/i386.c (working copy)
@@ -11785,9 +11785,6 @@ ix86_address_subreg_operand (rtx op)
mode = GET_MODE (op);
- if (GET_MODE_CLASS (mode) != MODE_INT)
- return false;
-
/* Don't allow SUBREGs that span more than a word. It can lead to spill
failures when the register is one word out of a two word structure. */
if (GET_MODE_SIZE (mode) > UNITS_PER_WORD)
@@ -11962,19 +11959,6 @@ ix86_decompose_address (rtx addr, struct ix86_addr
scale = 1 << scale;
retval = -1;
}
- else if (CONST_INT_P (addr))
- {
- if (!x86_64_immediate_operand (addr, VOIDmode))
- return 0;
-
- /* Constant addresses are sign extended to 64bit, we have to
- prevent addresses from 0x80000000 to 0xffffffff in x32 mode. */
- if (TARGET_X32
- && val_signbit_known_set_p (SImode, INTVAL (addr)))
- return 0;
-
- disp = addr;
- }
else
disp = addr; /* displacement */
@@ -12706,6 +12690,12 @@ ix86_legitimate_address_p (enum machine_mode mode
&& !x86_64_immediate_operand (disp, VOIDmode))
/* Displacement is out of range. */
return false;
+ /* In x32 mode, constant addresses are sign extended to 64bit, so
+ we have to prevent addresses from 0x80000000 to 0xffffffff. */
+ else if (TARGET_X32 && !(index || base)
+ && CONST_INT_P (disp)
+ && val_signbit_known_set_p (SImode, INTVAL (disp)))
+ return false;
}
/* Everything looks valid. */
Index: testsuite/gcc.target/i386/pr59153.c
===================================================================
--- testsuite/gcc.target/i386/pr59153.c (revision 0)
+++ testsuite/gcc.target/i386/pr59153.c (working copy)
@@ -0,0 +1,13 @@
+/* { dg-do compile } */
+/* { dg-options "-O -flive-range-shrinkage -mdispatch-scheduler -march=bdver1" } */
+
+int foo (float f)
+{
+ union
+ {
+ float f;
+ int i;
+ } z = { .f = f };
+
+ return z.i - 1;
+}
More information about the Gcc-patches
mailing list