[LLVMdev] better code for IV

Andrew Trick atrick at apple.com
Thu Feb 27 16:43:54 PST 2014


On Feb 26, 2014, at 5:23 AM, Shemer, Anat <anat.shemer at intel.com> wrote:

> Hi Andy,
>  
> Thanks a lot for looking into this.
>  
> I work on a backend for OpenCL.
> (1)   This pattern of 64bit loop limit and 32bit IV appears in many OpenCL kernels, so I have to deal with what I get and changing the source code is not an option.
> (2)   I guess that you wrote LSR for a reason and that it has value so I didn’t consider this option.
>  
> I have a solution that implements your option (4). However I think that this was done in order to do all operations in the same size, so this is not my first choice. It seems to me more reasonable not to take apart the ashr(shl()) in the LSR pass.
>  
> Can you please advise how to do your proposed option (3)? I looked in the pass code and I couldn’t figure this out.

I can think of one potential quick hack. In IVUsers.cpp, isInteresting() you could return false for a shl that looks like the first half of a sign-extend. Maybe you can see if its only user is ashr. There are other ways, but this seems quick and easy.

Please you file a PR with your test case and move the discussion there.

Thanks.

-Andy


>  
> Thanks, Anat
>  
>  
> From: Andrew Trick [mailto:atrick at apple.com] 
> Sent: Wednesday, February 26, 2014 06:48
> To: Shemer, Anat
> Cc: llvmdev at cs.uiuc.edu
> Subject: Re: better code for IV
>  
>  
> On Feb 18, 2014, at 11:32 PM, Shemer, Anat <anat.shemer at intel.com> wrote:
> 
> 
> Hi Andrew,
> The issue below refers to LSR, so I’ll appreciate your feedback. It also refers to instruction combining and might impact backends other than X86, so if you know of others that might be interested you are more than welcome to add them.
> Thanks, Anat
>  
> I'm not sure how you ended up with that IR. You have a 64-bit index that is 32-bit sign-extended within the loop. That's not something I've seen. Normally you have a pure 64-bit index or a 32-bit sign extended index. In this case, the sign extend would be hoisted out of the loop.
>  
> That said, LSR has a lot of trouble with sext/zext/trunc.
>  
> In this case, SCEV is getting it right:
>  
> %idxprom = ashr exact i64 %sext, 32
>   -->  (sext i32 {0,+,1}<%L_entry> to i64) Exits: (sext i32 (trunc i64 (-1 + %iNumElements) to i32) to i64)
>  
> The presense of the shl does lose the nuw/nsw flags, but I don't think that's your problem.
>  
> I had to debug this a bit to understand the issue, hence my delay in answering. Some possibilities:
>  
> (1) Generate IR that is normally generated for C code, or change the source to match conventional patterns.
>  
> (2) -mllvm -disable-lsr. You'll still have two increments, but no crazy shifting. You could rewrite the code with a single counter/index and get perfect codegen.
>  
> (3) Teach IV-Users to see through the ashr instruction. Currently it thinks the ashr is not interesting so tells LSR to generate an expression for its operands. That's why you get the weird shifted IV. This is probably the right fix, but I can't say for sure without trying it.
>  
> (4) Don't let instcombine bitwise expand the trunc/sext when it feeds a gep. I'm not a big fan of the bitwise expansion but realize it was done for a reason. I don't know all of the optimizations that it exposes without doing some investigation. Before you could do this you would need to send a specific proposal titled "InstCombine should not expand trunc/sext into shl/ashr" and get the ok from other people on the list. You would also need to do performance analysis of llvm test-suite and get someone to do the same for arm.
>  
> -Andy
>  
>  
>  
> _____________________________________________
> From: Shemer, Anat 
> Sent: Tuesday, February 18, 2014 15:07
> To: 'llvmdev at cs.uiuc.edu'
> Subject: better code for IV
>  
>  
> Hi,
>  
> I will be glad if you can give me some advice on how to make LLVM generate better code for IV in a simple loop case. There is IR sample below, I know of one way to do it, and I’m looking for advice on another way.
>  
> When compiling a simple C loop (c[i]=a[i]+b[i], a, b, c are float*), the IR starts as follows:
>  
>         define void @ArrayAdd1(float* nocapture %a, float* nocapture %b, float* nocapture %c, i64 %iNumElements) {
>         Entry:
>           br label %L_pre_head
>  
>         L_pre_head:                                       ; preds = %Entry
>           br label %L_entry
>  
>         L_entry:                                          ; preds = %L_entry, %L_pre_head
>           %L_ind_var = phi i64 [ 0, %L_pre_head ], [ %L_inc_ind_var, %L_entry ]
>           %L_tid = phi i64 [ 0, %L_pre_head ], [ %L_inc_tid, %L_entry ]
>           %trunc = trunc i64 %L_tid to i32
>           %idxprom = sext i32 %trunc to i64
>           %arrayidx = getelementptr inbounds float* %a, i64 %idxprom
>           %0 = load float* %arrayidx, align 4
>           %arrayidx2 = getelementptr inbounds float* %b, i64 %idxprom
>           %1 = load float* %arrayidx2, align 4
>           %add = fadd float %0, %1
>           %arrayidx4 = getelementptr inbounds float* %c, i64 %idxprom
>           store float %add, float* %arrayidx4, align 4
>           %L_inc_ind_var = add nuw nsw i64 %L_ind_var, 1
>           %L_cmp.to.max = icmp eq i64 %L_inc_ind_var, %iNumElements
>           %L_inc_tid = add nuw nsw i64 %L_tid, 1
>           br i1 %L_cmp.to.max, label %L_exit, label %L_entry
>  
>         L_exit:                                           ; preds = %L_entry
>           br label %2
>  
>         ; <label>:2                                       ; preds = %L_exit
>           ret void
>         }
>  
> And after going through all passes before code generation it becomes:
>  
>         define void @ArrayAdd1(float * nocapture %a, float * nocapture %b, float * nocapture %c, i64 %iNumElements)
>         {
>         Entry:
>           br label %L_pre_head
>  
>         L_pre_head:                                   ; preds = %Entry
>           br label %L_entry
>  
>         L_entry:                              ; preds = %L_entry, %L_pre_head
>           %L_ind_var = phi i64 [ 0, %L_pre_head ], [ %L_inc_ind_var, %L_entry ]
>           %L_tid = phi i64 [ 0, %L_pre_head ], [ %L_inc_tid, %L_entry ]
>           %sext = shl i64 %L_tid, 32
>           %idxprom = ashr exact i64 %sext, 32
>           %arrayidx = getelementptr inbounds float * %a, i64 %idxprom
>           %0 = load float * %arrayidx, align 4
>           %arrayidx2 = getelementptr inbounds float * %b, i64 %idxprom
>           %1 = load float * %arrayidx2, align 4
>           %add = fadd float %0, %1
>           %arrayidx4 = getelementptr inbounds float * %c, i64 %idxprom
>           store float %add, float * %arrayidx4, align 4
>           %L_inc_ind_var = add nuw nsw i64 %L_ind_var, 1
>           %L_cmp.to.max = icmp eq i64 %L_inc_ind_var, %iNumElements
>           %L_inc_tid = add nuw nsw i64 %L_tid, 1
>           br i1 %L_cmp.to.max, label %L_exit, label %L_entry
>  
>         L_exit:                                       ; preds = %L_entry
>           br label %2
>  
>         ; <label>:2                                       ; preds = %L_exit
>           ret void
>         }
>  
> On the way to code gen the IR goes through some additional passes, among them the LSR in LoopStrengthReduce.cpp. This pass promotes the shl operation out of the loop. As a result the generated code is awkward and uses 4 registers for IV management instead of 2 (see transformed IR and result asm below).
>  
> One way to deal with this is to prevent the transformation of
>       %trunc = trunc i64 %L_tid to i32
>       %idxprom = sext i32 %trunc to i64
> to
>       %sext = shl i64 %L_tid, 32
>       %idxprom = ashr exact i64 %sext, 32
>  
> Sext(trunk()) remains nice until code gen and one asm instruction of the sort “move %eax, %rax” is generated, as I want (see asm sample at the end of this message). I implemented this locally in InstCombineCasts.cpp, in InstCombiner::visitSExt(). The code looks for the specific pattern of sext32to64(trunk64to32()) and doesn’t let it become ashr(shl()).
>  
> Another way that I can think of, is to keep ashr32(shl32()) together and then it’s possible to replace both instructions with one asm instruction (a la  “move %eax, %rax”) and avoid an additional register for monitoring end of loop. I’m not sure what is the right point in LSRInstance::LSRInstance to avoid this optimization.
>  
> Do you think that the first way is ok? If you think that this should be done in the second way or a third way I will appreciate your guidance.
>  
> Thanks, Anat
>  
>  
> The IR after LSR:
>         define void @ArrayAdd1(float* nocapture %a, float* nocapture %b, float* nocapture %c, i64 %iNumElements) {
>         Entry:
>           br label %L_pre_head
>  
>         L_pre_head:                                       ; preds = %Entry
>           br label %L_entry
>  
>         L_entry:                                          ; preds = %L_entry, %L_pre_head
>           %lsr.iv1 = phi i64 [ %lsr.iv.next2, %L_entry ], [ 0, %L_pre_head ]
>           %lsr.iv = phi i64 [ %lsr.iv.next, %L_entry ], [ %iNumElements, %L_pre_head ]
>           %idxprom = ashr exact i64 %lsr.iv1, 32
>           %arrayidx = getelementptr inbounds float* %a, i64 %idxprom
>           %0 = load float* %arrayidx, align 4
>           %arrayidx2 = getelementptr inbounds float* %b, i64 %idxprom
>           %1 = load float* %arrayidx2, align 4
>           %add = fadd float %0, %1
>           %arrayidx4 = getelementptr inbounds float* %c, i64 %idxprom
>           store float %add, float* %arrayidx4, align 4
>           %lsr.iv.next = add i64 %lsr.iv, -1
>           %lsr.iv.next2 = add i64 %lsr.iv1, 4294967296
>           %L_cmp.to.max = icmp eq i64 %lsr.iv.next, 0
>           br i1 %L_cmp.to.max, label %L_exit, label %L_entry
>  
>         L_exit:                                           ; preds = %L_entry
>           br label %2
>  
>         ; <label>:2                                       ; preds = %L_exit
>           ret void
>         }
>  
> Asm code:
>     ArrayAdd1:                              # @ArrayAdd1
>             .cfi_startproc
>     # BB#0:                                 # %Entry
>             xorl    %r9d, %r9d
>             movabsq $4294967296, %r8        # imm = 0x100000000
>             .align  16, 0x90
>     .LBB0_1:                                # %L_entry
>                                             # =>This Inner Loop Header: Depth=1
>             movq    %r9, %rax
>             sarq    $32, %rax
>             movss   (%rdi,%rax,4), %xmm0
>             addss   (%rsi,%rax,4), %xmm0
>             movss   %xmm0, (%rdx,%rax,4)
>             addq    %r8, %r9
>             decq    %rcx
>             jne     .LBB0_1
>     # BB#2:
>             Ret
>  
> This is what I want to get:
>     ArrayAdd2:                              # @ArrayAdd2
>             .cfi_startproc
>     # BB#0:                                 # %Entry
>             xorl    %eax, %eax
>             .align  16, 0x90
>     .LBB1_1:                                # %L_entry
>                                             # =>This Inner Loop Header: Depth=1
>             movslq  %eax, %r8
>             movss   (%rdi,%r8,4), %xmm0
>             addss   (%rsi,%r8,4), %xmm0
>             movss   %xmm0, (%rdx,%r8,4)
>             incq    %rax
>             cmpq    %rax, %rcx
>             jne     .LBB1_1
>     # BB#2:
>             Ret
>  
>  
> ---------------------------------------------------------------------
> Intel Israel (74) Limited
> 
> This e-mail and any attachments may contain confidential material for
> the sole use of the intended recipient(s). Any review or distribution
> by others is strictly prohibited. If you are not the intended
> recipient, please contact the sender and delete all copies.
> 
>  
> ---------------------------------------------------------------------
> Intel Israel (74) Limited
> 
> This e-mail and any attachments may contain confidential material for
> the sole use of the intended recipient(s). Any review or distribution
> by others is strictly prohibited. If you are not the intended
> recipient, please contact the sender and delete all copies.
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20140227/60181d2a/attachment.html>


More information about the llvm-dev mailing list