[llvm-bugs] [Bug 27827] New: partial redundent mov

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=27827

            Bug ID: 27827
           Summary: partial redundent mov
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Register Allocator
          Assignee: unassignedbugs at nondot.org
          Reporter: wmi at google.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

Created attachment 16390
  --> https://llvm.org/bugs/attachment.cgi?id=16390&action=edit
testcase 1.cc

Two testcases which may have the same problem:

For 1.cc, we saw the following code snippet in the generated code:

~/workarea/llvm-r270096/dbuild/bin/clang -O2 -S 1.cc -fno-omit-frame-pointer -o
1.s

.LBB0_43:                               # %if.end.i161
                                        #   in Loop: Header=BB0_8 Depth=1
        movl    %r14d, %ecx
        subq    %rcx, %r12
        movl    $64, %r8d
        subl    %r15d, %r8d
        movq    %rbx, %rdx
        movl    %r8d, %ecx
        shrq    %cl, %rdx
        addq    %r12, %rdx
        movl    %r15d, %ecx
        shlq    %cl, %rbx
        jmp     .LBB0_44
        ...
# BB#35:                                # %if.then.i123
                                        #   in Loop: Header=BB0_8 Depth=1
        movl    $4, %ecx
        subl    %r15d, %ecx
        movq    %r12, %rdx
        shrq    %cl, %rdx
        movl    %r15d, %ecx
        shlq    %cl, %r12
        movq    %r12, %rbx
.LBB0_44:                               # %if.end31.i165
                                        #   in Loop: Header=BB0_8 Depth=1
        movq    %rbx, %r12

For "movq    %rbx, %r12" in .LBB0_44, if it is promoted to the end of its
predecessors BB#35 and LBB0_43, the copy promoted to BB#35 can be removed and
the execution path from BB#35 to .LBB0_44 is accelerated.

For 2.c, to reproduce the problem, http://reviews.llvm.org/D20315?id=57666
should be applied.

~/workarea/llvm-r270096-with-D20315/dbuild/bin/clang -O2 -S 2.c

# BB#0:                                 # %entry
        pxor    %xmm0, %xmm0
        movdqa  .LCPI0_0(%rip), %xmm4   # xmm4 = [0,1,2,3]
        movl    $1000, %eax             # imm = 0x3E8
        movdqa  .LCPI0_1(%rip), %xmm1   # xmm1 = [4,4,4,4]
        movdqa  .LCPI0_2(%rip), %xmm2   # xmm2 = [2,2,2,2]
        pxor    %xmm3, %xmm3
        .p2align        4, 0x90
.LBB0_1:                                # %vector.body
                                        # =>This Inner Loop Header: Depth=1
        movdqa  %xmm4, %xmm5
        paddd   %xmm1, %xmm5
        pxor    %xmm2, %xmm4
        pshufd  $78, %xmm4, %xmm6       # xmm6 = xmm4[2,3,0,1]
        movdqa  %xmm6, %xmm7
        psrad   $31, %xmm7
        punpckldq       %xmm7, %xmm6    # xmm6 =
xmm6[0],xmm7[0],xmm6[1],xmm7[1]
        movdqa  %xmm4, %xmm7
        psrad   $31, %xmm7
        punpckldq       %xmm7, %xmm4    # xmm4 =
xmm4[0],xmm7[0],xmm4[1],xmm7[1]
        paddq   %xmm4, %xmm0
        paddq   %xmm6, %xmm3
        addl    $-4, %eax
        movdqa  %xmm5, %xmm4
        jne     .LBB0_1

It is beneficial to mov "movdqa  %xmm4, %xmm5" at the start of LBB0_1 to the
end of all its predecessors: the end of BB#0 and the end of LBB0_1.

# BB#0:                                 # %entry
        pxor    %xmm0, %xmm0
        movdqa  .LCPI0_0(%rip), %xmm4   # xmm4 = [0,1,2,3]
        movl    $1000, %eax             # imm = 0x3E8
        movdqa  .LCPI0_1(%rip), %xmm1   # xmm1 = [4,4,4,4]
        movdqa  .LCPI0_2(%rip), %xmm2   # xmm2 = [2,2,2,2]
        pxor    %xmm3, %xmm3
        movdqa  %xmm4, %xmm5  ==> promoted to preheader
        .p2align        4, 0x90
.LBB0_1:                                # %vector.body
                                        # =>This Inner Loop Header: Depth=1
        paddd   %xmm1, %xmm5
        pxor    %xmm2, %xmm4
        pshufd  $78, %xmm4, %xmm6       # xmm6 = xmm4[2,3,0,1]
        movdqa  %xmm6, %xmm7
        psrad   $31, %xmm7
        punpckldq       %xmm7, %xmm6    # xmm6 =
xmm6[0],xmm7[0],xmm6[1],xmm7[1]
        movdqa  %xmm4, %xmm7
        psrad   $31, %xmm7
        punpckldq       %xmm7, %xmm4    # xmm4 =
xmm4[0],xmm7[0],xmm4[1],xmm7[1]
        paddq   %xmm4, %xmm0
        paddq   %xmm6, %xmm3
        addl    $-4, %eax
        movdqa  %xmm5, %xmm4
        movdqa  %xmm4, %xmm5     ==> apparently redundent and will be deleted.
        jne     .LBB0_1

I think register coalescer can be improved to remove such partial redundent
mov.

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20160520/7d355a5a/attachment.html>