This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [RFC PATCH, i386]: Enable post-reload compare elimination pass


On Sat, May 13, 2017 at 2:47 PM, Markus Trippelsdorf
<markus@trippelsdorf.de> wrote:
>> 2017-05-12  Uros Bizjak  <ubizjak@gmail.com>
>>
>>     * compare-elim.c (try_eliminate_compare): Canonicalize
>>     operation with embedded compare to
>>     [(set (reg:CCM) (compare:CCM (operation) (immediate)))
>>      (set (reg) (operation)].
>>
>>     * config/i386/i386.c (TARGET_FLAGS_REGNUM): New define.
>>
>> Re-bootstrapped and re-tested on x86_64-linux-gnu {,-m32}.
>>
>> Committed to mainline SVN.
>
> This causes gcc.dg/atomic/c11-atomic-exec-2.c to ICE with e.g.
> -march=nehalem:
>
> markus@x4 gcc % gcc -std=c11 -pedantic-errors -march=nehalem -O2 ./gcc/testsuite/gcc.dg/atomic/c11-atomic-exec-2.c
> ./gcc/testsuite/gcc.dg/atomic/c11-atomic-exec-2.c: In function ‘test_minus’:
> ./gcc/testsuite/gcc.dg/atomic/c11-atomic-exec-2.c:120:1: internal compiler error: in ix86_cc_mode, at config/i386/i386.c:22485
>  }
>  ^
> 0xea7b37 ix86_cc_mode(rtx_code, rtx_def*, rtx_def*)
>         /home/markus/gcc/gcc/config/i386/i386.c:22485
> 0x1246e11 maybe_select_cc_mode
>         /home/markus/gcc/gcc/compare-elim.c:500
> 0x1246e11 try_eliminate_compare
>         /home/markus/gcc/gcc/compare-elim.c:665
> 0x1246e11 execute_compare_elim_after_reload
>         /home/markus/gcc/gcc/compare-elim.c:727
> 0x1246e11 execute
>         /home/markus/gcc/gcc/compare-elim.c:770

-mtune=intel is able to copy SFmode value through general registers,
resulting in the following sequence:

(insn 288 1008 1009 29 (parallel [
            (set (reg:DI 4 si [687])
                (lshiftrt:DI (reg:DI 4 si [687])
                    (const_int 32 [0x20])))
            (clobber (reg:CC 17 flags))
        ]) "ttt.c":41 546 {*lshrdi3_1}
     (nil))
(insn 1009 288 289 29 (set (reg:DI 2 cx [419])
        (reg:DI 4 si [687])) "ttt.c":41 81 {*movdi_internal}
     (nil))
(insn 289 1009 291 29 (set (reg:SF 21 xmm0 [425])
        (reg:SF 2 cx [419])) "ttt.c":41 127 {*movsf_internal}
     (nil))
(insn 291 289 292 29 (set (reg:CCFPU 17 flags)
        (compare:CCFPU (reg:SF 21 xmm0 [425])
            (reg:SF 27 xmm6 [688]))) "ttt.c":41 51 {*cmpiusf}
     (expr_list:REG_EQUAL (compare:CCFPU (reg:SF 21 xmm0 [425])
            (const_double:SF 0.0 [0x0.0p+0]))
        (nil)))

Looking at compare-elim.c, equivalent_reg_at_start, the function
traces the value in %xmm0 to %rsi. So, try_eliminate_compare tries to
check if the target supports the compare of

                 (lshiftrt:DI (reg:DI 4 si [687])
                    (const_int 32 [0x20])))

with

(reg:SF 27 xmm6 [688])

which won't fly, since compare operands must have the same modes.

I'm testing the following patch that ensures that the mode of register
copy, found by equivalent_reg_at_start equals original mode.

--cut here--
Index: compare-elim.c
===================================================================
--- compare-elim.c      (revision 247995)
+++ compare-elim.c      (working copy)
@@ -526,6 +526,7 @@ maybe_select_cc_mode (struct comparison *cmp, rtx
 static rtx
 equivalent_reg_at_start (rtx reg, rtx_insn *end, rtx_insn *start)
 {
+  machine_mode orig_mode = GET_MODE (reg);
   rtx_insn *bb_head = BB_HEAD (BLOCK_FOR_INSN (end));

   for (rtx_insn *insn = PREV_INSN (end);
@@ -572,6 +573,9 @@ equivalent_reg_at_start (rtx reg, rtx_insn *end, r
        return NULL_RTX;
     }

+  if (GET_MODE (reg) != orig_mode)
+    return NULL_RTX;
+
   return reg;
 }

--cut here--

Uros.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]