public inbox for gcc-bugs@sourceware.org
help / color / mirror / Atom feed
* [Bug rtl-optimization/59880] New: Improve REE for implicit SI->DI zero-extend
@ 2014-01-19 16:14 hjl.tools at gmail dot com
  2014-01-19 17:09 ` [Bug rtl-optimization/59880] " jakub at gcc dot gnu.org
                   ` (10 more replies)
  0 siblings, 11 replies; 12+ messages in thread
From: hjl.tools at gmail dot com @ 2014-01-19 16:14 UTC (permalink / raw)
  To: gcc-bugs

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59880

            Bug ID: 59880
           Summary: Improve REE for implicit SI->DI zero-extend
           Product: gcc
           Version: 4.9.0
            Status: UNCONFIRMED
          Severity: enhancement
          Priority: P3
         Component: rtl-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: hjl.tools at gmail dot com

After r206774, on Linux/x86-64, the following code

---foo.c---
extern __thread unsigned int __bid_IDEC_glbflags;
typedef unsigned long long UINT64;
typedef __attribute__ ((aligned(16))) struct
{
  UINT64 w[2];
} UINT128;
extern UINT64 __bid64_from_uint64 (UINT64);
extern void __bid_round64_2_18 (int q,
                int x,
                UINT64 C,
                UINT64 * ptr_Cstar,
                int *delta_exp,
                int *ptr_is_midpoint_lt_even,
                int *ptr_is_midpoint_gt_even,
                int *ptr_is_inexact_lt_midpoint,
                int *ptr_is_inexact_gt_midpoint);
extern void __bid_round128_19_38 (int q,
                  int x,
                  UINT128 C,
                  UINT128 * ptr_Cstar,
                  int *delta_exp,
                  int *ptr_is_midpoint_lt_even,
                  int *ptr_is_midpoint_gt_even,
                  int *ptr_is_inexact_lt_midpoint,
                  int *ptr_is_inexact_gt_midpoint);
UINT64
__bid64_from_uint64 (UINT64 x)
{
  UINT64 res;
  UINT128 x128, res128;
  unsigned int q, ind;
  int incr_exp = 0;
  int is_midpoint_lt_even = 0, is_midpoint_gt_even = 0;
  int is_inexact_lt_midpoint = 0, is_inexact_gt_midpoint = 0;
  if (x <= 0x002386F26FC0ffffull) {
    if (x < 0x0020000000000000ull) {
      res = 0x31c0000000000000ull | x;
    } else {
      res = 0x6c70000000000000ull | (x & 0x0007ffffffffffffull);
    }
  }
  else
    {
      if (x < 0x16345785d8a0000ull) {
    q = 17;
    ind = 1;
      } else if (x < 0xde0b6b3a7640000ull) {
    q = 18;
    ind = 2;
      } else if (x < 0x8ac7230489e80000ull) {
    q = 19;
    ind = 3;
      } else {
    q = 20;
    ind = 4;
      }
      if (q <= 19) {
    __bid_round64_2_18 (
                q, ind, x, &res, &incr_exp,
                &is_midpoint_lt_even, &is_midpoint_gt_even,
                &is_inexact_lt_midpoint, &is_inexact_gt_midpoint);
      }
      else {
    x128.w[1] = 0x0;
    x128.w[0] = x;
    __bid_round128_19_38 (q, ind, x128, &res128, &incr_exp,
                  &is_midpoint_lt_even, &is_midpoint_gt_even,
                  &is_inexact_lt_midpoint, &is_inexact_gt_midpoint);
    res = res128.w[0];
      }
      if (incr_exp)
    ind++;
      if (is_inexact_lt_midpoint || is_inexact_gt_midpoint ||
      is_midpoint_lt_even || is_midpoint_gt_even)
    *&__bid_IDEC_glbflags |= 0x00000020;
      if (res < 0x0020000000000000ull) {
    res = (((UINT64) ind + 398) << 53) | res;
      } else
    {
      res = 0x6000000000000000ull | (((UINT64) ind + 398) << 51) |
        (res & 0x0007ffffffffffffull);
    }
    }
  return(res);;
}
-----------

contains 2 extra SI->DI zero-extend when compiled with

-O2 -march=corei7 -mtune=slm -fPIC

        movl    %ebp, %edx      # 311   *movsi_internal/1       [length = 2]
        ^^^^^^^^^^^^^^^^^^^ Implicit SI->DI zero-extend
        leaq    88(%rsp), %rsp  # 267   pro_epilogue_adjust_stack_di_add/1     
[length = 5]
        .cfi_remember_state
        .cfi_def_cfa_offset 24
        movabsq $2251799813685247, %rax # 101   *movdi_internal/5       [length
= 10]
        movl    %edx, %edx      # 312   *zero_extendsidi2/4     [length = 2]
        ^^^^^^^^^^^^^^^^^^^ Unnecessary


        movl    %ebp, %eax      # 308   *movsi_internal/1       [length = 2]
        ^^^^^^^^^^^^^^^^^^^ Implicit SI->DI zero-extend
        leaq    88(%rsp), %rsp  # 287   pro_epilogue_adjust_stack_di_add/1     
[length = 5]
        .cfi_remember_state
        .cfi_def_cfa_offset 24
        movl    %eax, %eax      # 309   *zero_extendsidi2/4     [length = 2]
        ^^^^^^^^^^^^^^^^^^^^ Unnecessary

REE pass should remove them.


^ permalink raw reply	[flat|nested] 12+ messages in thread
end of thread, other threads:[~2014-01-22 22:01 UTC | newest]

Thread overview: 12+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2014-01-19 16:14 [Bug rtl-optimization/59880] New: Improve REE for implicit SI->DI zero-extend hjl.tools at gmail dot com
2014-01-19 17:09 ` [Bug rtl-optimization/59880] " jakub at gcc dot gnu.org
2014-01-19 17:34 ` [Bug rtl-optimization/59880] ix86_avoid_lea_for_addr is buggy hjl.tools at gmail dot com
2014-01-19 18:25 ` ubizjak at gmail dot com
2014-01-19 18:41 ` ubizjak at gmail dot com
2014-01-19 19:41 ` jakub at gcc dot gnu.org
2014-01-20  9:52 ` [Bug target/59880] " jakub at gcc dot gnu.org
2014-01-20  9:55 ` jakub at gcc dot gnu.org
2014-01-22 18:29 ` uros at gcc dot gnu.org
2014-01-22 19:58 ` uros at gcc dot gnu.org
2014-01-22 19:59 ` ubizjak at gmail dot com
2014-01-22 22:01 ` hjl at gcc dot gnu.org

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).