[Bug target/64226] New: Secondary reload incorrect TOC address

[Bug target/64226] New: Secondary reload incorrect TOC address

[dje at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64226

            Bug ID: 64226
           Summary: Secondary reload incorrect TOC address
           Product: gcc
           Version: 5.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: dje at gcc dot gnu.org

Many of the GCC testsuite VMX tests fail on AIX because secondary reload is
generating an incorrect TOC reference. A MEM is being stripped from a
TOC reference causing GCC to use a pattern for a raw address,
generating a bogus reference.

226r.ira:

(insn 347 346 348 8 (set (reg/f:SI 489)
        (mem/u/c:SI (unspec:SI [
                    (symbol_ref/u:SI ("*LC..32") [flags 0x2])
                    (reg:SI 2 2)
                ] UNSPEC_TOCREL) [0  S4 A8])) ld.c:83 411 {*movsi_internal1}
     (expr_list:REG_EQUIV (symbol_ref/u:SI ("*LC..31") [flags 0x2])
        (nil)))
(insn 348 347 349 8 (set (reg:V8HI 488)
        (mem/u/c:V8HI (reg/f:SI 489) [0  S16 A128])) ld.c:83 972
{*altivec_movv8
hi}
     (expr_list:REG_DEAD (reg/f:SI 489)
        (expr_list:REG_EQUIV (const_vector:V8HI [
                    (const_int 0 [0])
                    (const_int 1 [0x1])
                    (const_int 2 [0x2])
                    (const_int 3 [0x3])
                    (const_int 4 [0x4])
                    (const_int 5 [0x5])
                    (const_int 6 [0x6])
                    (const_int 7 [0x7])
                ])
            (nil))))
(insn 349 348 350 8 (parallel [
            (set (reg:CC 74 6)
                (unspec:CC [
                        (eq:CC (reg:V8HI 456)
                            (reg:V8HI 488))
                    ] UNSPEC_PREDICATE))
            (set (reg:V8HI 490)
                (eq:V8HI (reg:V8HI 456)
                    (reg:V8HI 488)))
        ]) ld.c:83 1206 {*altivec_vcmpequh_p}
     (expr_list:REG_DEAD (reg:V8HI 456)
        (expr_list:REG_UNUSED (reg:V8HI 490)
            (nil))))


227r.reload:

Reloads for insn # 349
Reload 0: reload_in (SI) = (plus:SI (reg/f:SI 1 1)
                                                    (const_int 96 [0x60]))
        BASE_REGS, RELOAD_FOR_INPUT_ADDRESS (opnum = 1), can't combine
        reload_in_reg: (plus:SI (reg/f:SI 1 1)
                                                    (const_int 96 [0x60]))
        reload_reg_rtx: (reg:SI 7 7)
Reload 1: reload_in (SI) = (symbol_ref/u:SI ("*LC..31") [flags 0x2])
        BASE_REGS, RELOAD_FOR_INPUT_ADDRESS (opnum = 2)
        reload_in_reg: (symbol_ref/u:SI ("*LC..31") [flags 0x2])
        reload_reg_rtx: (reg/f:SI 9 9 [489])
Reload 2: reload_in (V8HI) = (mem/c:V8HI (plus:SI (reg/f:SI 1 1)
                                                        (const_int 96 [0x60]))
[
0 %sfp+96 S16 A128])
        ALTIVEC_REGS, RELOAD_FOR_INPUT (opnum = 1), can't combine
        reload_in_reg: (reg:V8HI 456)
        reload_reg_rtx: (reg:V8HI 78 1)Reload 3: BASE_REGS,
RELOAD_FOR_INPUT_ADDRESS (opnum = 2), can't combine, second
ary_reload_p
        reload_reg_rtx: (reg:SI 7 7)
Reload 4: reload_in (V8HI) = (mem/u/c:V8HI (symbol_ref/u:SI ("*LC..31") [flags
0
x2]) [0  S16 A128])
        ALTIVEC_REGS, RELOAD_FOR_INPUT (opnum = 2), can't combine
        reload_in_reg: (reg:V8HI 488)
        reload_reg_rtx: (reg:V8HI 90 13)
        secondary_in_reload = 3
        secondary_in_icode = reload_v8hi_si_load

(insn 498 497 499 8 (set (reg:SI 7 7)
        (unspec:SI [
                (symbol_ref/u:SI ("*LC..31") [flags 0x2])
                (reg:SI 2 2)
            ] UNSPEC_TOCREL)) ld.c:83 532 {*tocrefsi}
     (nil))
(insn 499 498 349 8 (set (reg:V8HI 90 13)
        (mem/u/c:V8HI (reg:SI 7 7) [0  S16 A128])) ld.c:83 972
{*altivec_movv8hi
}
     (nil))
(insn 349 499 350 8 (parallel [
            (set (reg:CC 74 6)
                (unspec:CC [
                        (eq:CC (reg:V8HI 78 1)
                            (reg:V8HI 90 13))
                    ] UNSPEC_PREDICATE))
            (set (reg:V8HI 77 0 [490])
                (eq:V8HI (reg:V8HI 78 1)
                    (reg:V8HI 90 13)))
        ]) ld.c:83 1206 {*altivec_vcmpequh_p}
     (nil))

Pseudo 489 was loaded with (mem:SI (unspec:SI [LC..32]
TOCREL)) but r7 is loaded directly with (unspec:SI [LC..31] TOCREL).

I assume that find_replacement is finding symbol_ref LC..31 and using
that.  There is no valid way to reference LC..31 directly on AIX.
It's clearer with -fno-section-anchors.  LC..31 is in read-only data:

        .csect _ld.rw_[RO],4
        .align 4
...
LC..31:
        .short  0
        .short  1
        .short  2
        .short  3
        .short  4
        .short  5
        .short  6
        .short  7

LC..32 would have been a TOC reference to LC..31

        .toc
LC..32:
        .tc LC..31[TC],LC..31

but that is deleted and replaced with a direct reference to the RO
data.  The final instruction stream is

        addi 7,1,96
        lvx 1,0,7
        la 7,LC..31(2)     <-----
        lvx 13,0,7
        vcmpequh. 0,1,13
        lwz 4,LC..34(2)
        mfcr 3
        rlwinm 3,3,25,1
        bl .check

which tries to load the RO data directly and use LC..31 as a TOC displacement.

[Bug target/64226] [5.0 Regression] Secondary reload incorrect TOC address

[dje at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64226

David Edelsohn <dje at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2014-12-08
                 CC|                            |meissner at gcc dot gnu.org,
                   |                            |uweigand at gcc dot gnu.org
   Target Milestone|---                         |5.0
            Summary|Secondary reload incorrect  |[5.0 Regression] Secondary
                   |TOC address                 |reload incorrect TOC
                   |                            |address
     Ever confirmed|0                           |1

--- Comment #1 from David Edelsohn <dje at gcc dot gnu.org> ---
Confirmed.

[Bug target/64226] [5.0 Regression] Secondary reload incorrect TOC address

[dje at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64226

--- Comment #2 from David Edelsohn <dje at gcc dot gnu.org> ---
Uli mentioned in private email:

"I think the piece of code quoted above from rs6000_secondary_reload_inner
is wrong; it should not call create_TOC_reference unconditionally.

"Other places that use TOC-relative addressing first verify whether the
symbol actually supports that, using use_toc_relative_ref.  So this
check should probably be done here too.  If a symbol cannot be accessed
via TOC-relative addressing, we should call rs6000_emit_move, which
should do the right thing (forcing the address into the TOC)."

And the following patch to remove the special call to create_TOC_reference
seems to fix the problem.

Index: rs6000.c
===================================================================
--- rs6000.c    (revision 218484)
+++ rs6000.c    (working copy)
@@ -17379,12 +17379,7 @@
     case SYMBOL_REF:
     case CONST:
     case LABEL_REF:
-      if (TARGET_TOC)
-       emit_insn (gen_rtx_SET (VOIDmode, scratch,
-                               create_TOC_reference (addr, scratch)));
-      else
-       rs6000_emit_move (scratch, addr, Pmode);
-
+      rs6000_emit_move (scratch, addr, Pmode);
       new_addr = scratch;
       break;

I am testing bootstrap.

[Bug target/64226] [5.0 Regression] Secondary reload incorrect TOC address

[dje at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64226

--- Comment #3 from David Edelsohn <dje at gcc dot gnu.org> ---
Author: dje
Date: Mon Dec  8 23:47:39 2014
New Revision: 218497

URL: https://gcc.gnu.org/viewcvs?rev=218497&root=gcc&view=rev
Log:
        PR target/64226
        * config/rs6000/rs6000.c (rs6000_secondary_reload_inner)
        [SYMBOL_REF]: Do not explicitly call create_TOC_reference for
        TARGET_TOC. Always use rs6000_emit_move.

Modified:
    trunk/gcc/ChangeLog
    trunk/gcc/config/rs6000/rs6000.c

[Bug target/64226] [5.0 Regression] Secondary reload incorrect TOC address

[dje at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64226

David Edelsohn <dje at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|NEW                         |RESOLVED
         Resolution|---                         |FIXED

--- Comment #4 from David Edelsohn <dje at gcc dot gnu.org> ---
Fixed.