Vlad, Jeff and Segher,
I think I have determined what is happening with the aarch64 test case that
is failing after my r264897 commit. It appears my patch is just exposing
an issue in lra-constraints.c:process_alt_operands() when processing an insn
with early clobber operands. Jeff & Segher, I have reverted all of my
previous patches we were making to lra-lives.c for the analysis below and
am just using straight trunk revision 264897 (ie, my last patch).
The aarch64 test case is below. Note that the first 2 input/output operands
are early clobber and are used with asm defined hard regs (x0 and x1).
Also notice that oldval1 and oldval2 are pseudos with the same values as
those in/out operands input values.
long
__cmpxchg_double (unsigned long arg)
{
unsigned long old1 = 1;
unsigned long old2 = arg;
unsigned long new1 = 2;
unsigned long new2 = 3;
void *ptr = 0;
unsigned long oldval1 = old1;
unsigned long oldval2 = old2;
register unsigned long x0 asm ("x0") = old1;
register unsigned long x1 asm ("x1") = old2;
register unsigned long x2 asm ("x2") = new1;
register unsigned long x3 asm ("x3") = new2;
register unsigned long x4 asm ("x4") = (unsigned long) ptr;
asm volatile ("casp %[old1], %[old2], %[new1], %[new2], %[v]\n"
"eor %[old1], %[old1], %[oldval1]\n"
"eor %[old2], %[old2], %[oldval2]\n"
"orr %[old1], %[old1], %[old2]\n"
: [old1] "+&r" (x0), [old2] "+&r" (x1), [v] "+Q" (* (unsigned
long *) ptr)
: [new1] "r" (x2), [new2] "r" (x3), [ptr] "r" (x4), [oldval1]
"r" (oldval1), [oldval2] "r" (oldval2)
: "x16", "x17", "x30");
return x0;
}
After IRA, we have the following register assignment and rtl. The only
difference between before my patch and after my patch is that r92 used
to be assigned to a hardreg that is not x0 - x4, which hides the lurking
problem.
r92 -> x1, r93 -> x5, r94 -> x6
(insn 2 3 8 2 (set (reg/v:DI 92 [arg]) (reg:DI 0 x0)) (REG_DEAD (reg:DI 0 x0)))
(insn 8 2 7 2 (set (reg/v:DI 2 x2 [*x2]) (const_int 2)))
(insn 7 8 6 2 (set (reg/v:DI 1 x1 [*x1]) (reg/v:DI 92 [arg])))
(insn 6 7 9 2 (set (reg/v:DI 0 x0 [*x0]) (const_int 1)))
(insn 9 6 12 2 (set (reg/v:DI 3 x3 [*x3]) (const_int 3)))
(insn 12 9 10 2 (set (reg:DI 94 [*x0]) (reg/v:DI 0 x0 [*x0])))
(insn 10 12 11 2 (set (reg/v:DI 4 x4 [*x4]) (const_int 0)))
(insn 11 10 14 2 (set (reg/f:DI 93) (const_int 0)))
(insn 14 11 21 2 (parallel [
(set (reg/v:DI 0 x0 [*x0])
(asm_operands/v:DI ("casp %0, %1, %3, %4, %2
eor %0, %0, %6
eor %1, %1, %7
orr %0, %0, %1
") ("=&r") 0 [(reg/v:DI 2 x2 [*x2])
(reg/v:DI 3 x3 [*x3])
(reg/v:DI 4 x4 [*x4])
(reg:DI 94 [*x0])
(reg/v:DI 92 [arg])
(reg/v:DI 0 x0 [*x0])
(reg/v:DI 1 x1 [*x1])
(mem:DI (reg/f:DI 93))]
[(asm_input:DI ("r"))
(asm_input:DI ("r"))
(asm_input:DI ("r"))
(asm_input:DI ("r"))
(asm_input:DI ("r"))
(asm_input:DI ("0"))
(asm_input:DI ("1"))
(asm_input:DI ("Q"))])
(set (reg/v:DI 1 x1 [*x1])
<same mnemonics as above>
") ("=&r") 1 [<same operands as above>]
(set (mem:DI (reg/f:DI 93) [1 MEM[(long unsigned intD.11 *)0B]+0 S8
A128])
<same mnemonics as above>
") ("=Q") 2 [<same operands as above>]
(clobber (reg:DI 30 x30))
(clobber (reg:DI 17 x17))
(clobber (reg:DI 16 x16))
]) "slub-min.c":17 -1
(expr_list:REG_DEAD (reg:DI 94 [*x0])
(expr_list:REG_DEAD (reg/f:DI 93)
(expr_list:REG_DEAD (reg/v:DI 92 [arg])
(expr_list:REG_DEAD (reg/v:DI 4 x4 [*x4])
(expr_list:REG_DEAD (reg/v:DI 3 x3 [*x3])
(expr_list:REG_DEAD (reg/v:DI 2 x2 [*x2])
(expr_list:REG_UNUSED (reg:DI 30 x30)
(expr_list:REG_UNUSED (reg:DI 17 x17)
(expr_list:REG_UNUSED (reg:DI 16 x16)
(expr_list:REG_UNUSED (reg/v:DI 1 x1
[*x1])
(nil))))))))))))
(insn 21 14 23 2 (use (reg/i:DI 0 x0)))
In lra-constraints.c:process_alt_operands(), we notice that pseudo 92 is
assigned to x1 and that an early clobber operand is also assigned to x1, or
rather, that it uses x1 explicitly. This is enough to trigger reload(s),
but the problem is we end up trying to reload the early clobber operand
which has been forced into x1 via register asm assignment, instead of
pseudo 92 which conflicts with it.
The problematic code in question is:
/* If earlyclobber operand conflicts with another
non-matching operand which is actually the same register
as the earlyclobber operand, it is better to reload the
another operand as an operand matching the earlyclobber
operand can be also the same. */
if (first_conflict_j == last_conflict_j
&& operand_reg[last_conflict_j] != NULL_RTX
&& ! curr_alt_match_win[last_conflict_j]
&& REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
{
curr_alt_win[last_conflict_j] = false;
curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
= last_conflict_j;
losers++;
/* Early clobber was already reflected in REJECT. */
lra_assert (reject > 0);
if (lra_dump_file != NULL)
fprintf
(lra_dump_file,
" %d Conflict early clobber reload: reject--\n",
i);
reject--;
overall += LRA_LOSER_COST_FACTOR - 1;
}
else
{
/* We need to reload early clobbered register and the
matched registers. */
....
In our test case, first_conflict_j == last_conflict_j and they point
to the operand pseudo 92. The first three tests in the "if" condition
are all true and then we check:
REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j])
Now REGNO (operand_reg[i]) is our 2nd early clobber operand and is regno x1.
However, REGNO (operand_reg[last_conflict_j]) returns regno 92, not the
hard register pseudo 92 was assigned to (ie, x1). This causes us to execute
the "else" clause of this code which tries to reload the early clobber
operand. We end up replacing the hard coded x1 in the 2nd early clobber
operand and then assign it to some reg like x7 and then the assembler
barfs because it isn't x1 which it has to be.
Now the REGNO (...) == REGNO (...) test predates the code that added
first_conflict_j and last_conflict_j, so I think we can just eliminate that
code altogether, since we've already proved there is a conflict in register
usage between the early clobber operand and this non-matching operand.
If I eliminate that check, we correctly reload input pseudo 92 and everything
is good. Secondly, I don't think it's ever legal to reload an operand
that has been hard coded by the user to a hard register via a register asm
definition like in the test case above.
With that in mind, what do people think of the patch below? This fixes the
AARCH64 test case. However, it ICE's on the following test case:
long foo (long arg)
{
register long var asm("x0");
asm("bla %0 %1" : "+&r"(var) : "r"(arg));
return var;
}
...but that is due to a combine bug where combine replaces the use of
"arg"'s pseudo in the inline asm with the incoming argument reg x0 which
should be illegal. Ie, it's taking a valid inline asm and creating an
invalid one since the earlyclobber op and non-matching op have both
been forced to use the same hard reg. Segher has a combine patch to
stop that which he is going to submit (commit?) soon. With my patch,
we are also now able to catch the following user error, when before we
could not:
bergner@pike:~/gcc/BUGS/PR87507/AARCH64$ cat ice4.i
long foo (void)
{
register long arg asm("x1");
register long var asm("x1");
asm ("bla %0 %1" : "=&r"(arg) : "r"(var));
return arg;
}
bergner@pike:~/gcc/BUGS/PR87507/AARCH64$ .../xgcc -B.../gcc -O2
-march=armv8.1-a -S ice4.i
ice4.i: In function âfooâ:
ice4.i:7:1: error: unable to generate reloads for:
7 | }
| ^
(insn 5 6 10 2 (set (reg/v:DI 1 x1 [ arg ])
(asm_operands:DI ("bla %0 %1") ("=&r") 0 [
(reg/v:DI 1 x1 [ var ])
]
[
(asm_input:DI ("r") ice4.i:5)
]
[] ice4.i:5)) "ice4.i":5 -1
(nil))
during RTL pass: reload
ice4.i:7:1: internal compiler error: in process_alt_operands, at
lra-constraints.c:2911
...
Thoughts? I'll note that this does not fix the S390 bugs, since those seem
to be due to problems with early clobber operands and "matching" constraint
operands. I'm still working on that and hope to have something soon.
I am currently bootstrapping this on s390x just to make sure I don't break
anything there, before debugging the existing s390x issue(s). I can't test
this on aarch64 or arm, other than knowing my aarch64 cross doesn't ICE on
the test case above.
Peter
* lra-constraints.c (process_alt_operands): Abort on illegal hard
register usage. Prefer reloading non hard register operands.
Index: gcc/lra-constraints.c
===================================================================
--- gcc/lra-constraints.c (revision 264897)
+++ gcc/lra-constraints.c (working copy)
@@ -2904,18 +2904,29 @@ process_alt_operands (int only_alternati
if (first_conflict_j < 0)
first_conflict_j = j;
last_conflict_j = j;
+ /* Both the earlyclobber operand and conflicting operand
+ cannot both be hard registers. */
+ if (REGNO (operand_reg[i]) < FIRST_PSEUDO_REGISTER
+ && operand_reg[j] != NULL_RTX
+ && REGNO (operand_reg[j]) < FIRST_PSEUDO_REGISTER)
+ fatal_insn ("unable to generate reloads for:", curr_insn);
}
if (last_conflict_j < 0)
continue;
- /* If earlyclobber operand conflicts with another
- non-matching operand which is actually the same register
- as the earlyclobber operand, it is better to reload the
- another operand as an operand matching the earlyclobber
- operand can be also the same. */
- if (first_conflict_j == last_conflict_j
- && operand_reg[last_conflict_j] != NULL_RTX
- && ! curr_alt_match_win[last_conflict_j]
- && REGNO (operand_reg[i]) == REGNO (operand_reg[last_conflict_j]))
+
+ /* If an earlyclobber operand conflicts with another non-matching
+ operand (ie, they have been assigned the same hard register),
+ then it is better to reload the other operand, as there may
+ exist yet another operand with a matching constraint associated
+ with the earlyclobber operand. However, if one of the operands
+ is an explicit use of a hard register, then we must reload the
+ other non-hard register operand. */
+ if (REGNO (operand_reg[i]) < FIRST_PSEUDO_REGISTER
+ || (operand_reg[last_conflict_j] != NULL_RTX
+ && REGNO (operand_reg[last_conflict_j])
+ >= FIRST_PSEUDO_REGISTER
+ && first_conflict_j == last_conflict_j
+ && ! curr_alt_match_win[last_conflict_j]))
{
curr_alt_win[last_conflict_j] = false;
curr_alt_dont_inherit_ops[curr_alt_dont_inherit_ops_num++]
