search

Quuxplusone / LLVMBugzillaTest

0 stars 0 forks source link

LoopUnroll pass generates some redundency not being cleaned up #23525

Open Quuxplusone opened 9 years ago

Quuxplusone commented 9 years ago
Bugzilla Link PR23524
Status NEW
Importance P normal
Reported by Wei Mi (wmi@google.com)
Reported on 2015-05-14 13:06:39 -0700
Last modified on 2015-05-14 13:16:22 -0700
Version trunk
Hardware PC Linux
CC hfinkel@anl.gov, llvm-bugs@lists.llvm.org
Fixed by commit(s)
Attachments 1.cc (171 bytes, text/x-c++src)
Blocks
Blocked by
See also 
Created attachment 14328
testcase 1.cc

For testcase 1.cc:

~/workarea/llvm-r236608/build/bin/clang++ -O2 -S 1.cc

The kernel loop in assembly:
.LBB0_9:                                # %for.body3
                                        #   Parent Loop BB0_3 Depth=1
                                        # =>  This Inner Loop Header: Depth=2
        decl    %edi
        cmpb    $0, (%rsi)
        jne     .LBB0_11
# BB#10:                                # %if.then
                                        #   in Loop: Header=BB0_9 Depth=2
        movl    $0, (%rdx)
        movw    $0, 4(%rdx)
.LBB0_11:                               # %for.inc
                                        #   in Loop: Header=BB0_9 Depth=2
        leaq    6(%rdx), %rbx
        decl    %edi          # could have been merged with decl %edi above
        cmpb    $0, 1(%rsi)
        jne     .LBB0_13
# BB#12:                                # %if.then.1
                                        #   in Loop: Header=BB0_9 Depth=2
        movl    $0, 6(%rdx)
        movw    $0, 10(%rdx)
.LBB0_13:                               # %for.inc.1
                                        #   in Loop: Header=BB0_9 Depth=2
        addq    $6, %rbx
        addq    $2, %rsi
        testl   %edi, %edi
        movq    %rbx, %rdx
        jne     .LBB0_9

We can see that the decl %edi in LBB0_9 and LBB0_11 could have been merged. The
latter decl %edi is generated in LoopUnroll pass. We need another InstCombine
pass to cleanup some redundency generated in LoopUnroll pass.

With another InstCombine pass added after LoopUnroll, we get:
.LBB0_10:                               # %for.body3
                                        #   Parent Loop BB0_3 Depth=1
                                        # =>  This Inner Loop Header: Depth=2
        cmpb    $0, (%rbx)
        jne     .LBB0_12
# BB#11:                                # %if.then
                                        #   in Loop: Header=BB0_10 Depth=2
        movl    $0, -10(%rdx)
        movw    $0, -6(%rdx)
.LBB0_12:                               # %for.inc
                                        #   in Loop: Header=BB0_10 Depth=2
        addl    $-2, %edi    # redundency removed.
        cmpb    $0, 1(%rbx)
        jne     .LBB0_14
# BB#13:                                # %if.then.1
                                        #   in Loop: Header=BB0_10 Depth=2
        movl    $0, -4(%rdx)
        movw    $0, (%rdx)
.LBB0_14:                               # %for.inc.1
                                        #   in Loop: Header=BB0_10 Depth=2
        addq    $12, %rdx
        addq    $2, %rbx
        testl   %edi, %edi
        jne     .LBB0_10
Quuxplusone commented 9 years ago

Attached 1.cc (171 bytes, text/x-c++src): testcase 1.cc