On 2014/6/11 下午 06:32, James Greenhalgh wrote:
>
> Hi,
>
> A recent change somewhere exposed a latent bug between LRA and the definition
> of the movsi_compare0 pattern.
>
> This pattern ties the source and destination register of a set together
> a (match_dup) and register constraints:
>
> [(set (reg:CC CC_REGNUM)
> (compare:CC (match_operand:SI 1 "s_register_operand" "0,r")
> (const_int 0)))
> (set (match_operand:SI 0 "s_register_operand" "=r,r")
> (match_dup 1))]
>
> This confuses LRA which expects the source and destination register of
> a set to be different.
>
> reduced.c: In function '_IO_vfscanf_internal':
> reduced.c:104:1: internal compiler error: in lra_create_copy, at lra.c:1512
> }
> ^
> 0x8c3f9a lra_create_copy(int, int, int)
> /work/gcc-dev/src/gcc/gcc/lra.c:1512
> 0x8e4ab0 process_bb_lives
> /work/gcc-dev/src/gcc/gcc/lra-lives.c:568
> 0x8e4ab0 lra_create_live_ranges(bool)
> /work/gcc-dev/src/gcc/gcc/lra-lives.c:1019
> 0x8c5a39 lra(_IO_FILE*)
> /work/gcc-dev/src/gcc/gcc/lra.c:2356
> 0x873a96 do_reload
> /work/gcc-dev/src/gcc/gcc/ira.c:5415
> 0x873a96 execute
> /work/gcc-dev/src/gcc/gcc/ira.c:5576
> Please submit a full bug report,
>
> We can fix the pattern by moving away from match_dup and using register
> tying with constraints consistently.
>
> I'm not entirely convinced that this is legitimate (my vague recollection is
> that register tying should only be used to tie inputs to outputs).
>
> This has passed testing on a bunch of ARM targets, and fixes the build
> issues I've been seeing.
Looking at this too, as an LRA exercise. I don't really think the
pattern is wrong, rather LRA should just avoid creating the copy in this
case; it's a result of operand constraining, after all.
Attached is the small LRA patch, pending testing. Vladimir should weight
in on this.
Thanks,
Chung-Lin
* ira-lives.c (process_bb_lives): Skip creating copy during
insn sca when src/dest has constrained to same regno.
Index: lra-lives.c
===================================================================
--- lra-lives.c (revision 211398)
+++ lra-lives.c (working copy)
@@ -558,7 +558,11 @@ process_bb_lives (basic_block bb, int &curr_point)
/* It might be 'inheritance pseudo <- reload pseudo'. */
|| (src_regno >= lra_constraint_new_regno_start
&& ((int) REGNO (SET_DEST (set))
- >= lra_constraint_new_regno_start))))
+ >= lra_constraint_new_regno_start)
+ /* Remember to skip special cases where src/dest regnos are
+ the same, e.g. insn SET pattern has matching constraints
+ like =r,0. */
+ && src_regno != (int) REGNO (SET_DEST (set)))))
{
int hard_regno = -1, regno = -1;
