[LLVMdev] Question on optimizeThumb2JumpTables

[Jakob Stoklund Olesen]

On Jul 29, 2013, at 6:50 AM, Chad Rosier <chad.rosier at gmail.com> wrote:

> Hi Jakob,
> You're the unfortunate soul who last touched the constant island pass, right?  Do you happen to have any insight for Daniel?

Sorry, no. I don't remember working with that particular bit of code. You could try digging through the commit logs.

Thanks,
/jakob



> On Tue, Jul 23, 2013 at 9:55 AM, Daniel Stewart <stewartd at codeaurora.org> wrote:
> In looking at the code in ARMConstantislandPass.cpp::optimizeThumb2JumpTables(), I see that there is the following condition for not creating tbb-based jump tables:
> 
>  
> 
>       // The instruction should be a tLEApcrel or t2LEApcrelJT; we want
> 
>       // to delete it as well.
> 
>       MachineInstr *LeaMI = PrevI;
> 
>       if ((LeaMI->getOpcode() != ARM::tLEApcrelJT &&
> 
>            LeaMI->getOpcode() != ARM::t2LEApcrelJT) ||
> 
>           LeaMI->getOperand(0).getReg() != BaseReg)
> 
>         OptOk = false;
> 
>  
> 
>       if (!OptOk)
> 
>         continue;
> 
>  
> 
> I am trying to figure out why the restriction of LeaMI->getOperand(0).getReg() != BaseReg is there. It seems this is overly restrictive. For example, here is a case where it succeeds:
> 
>  
> 
> 8944B   BB#53: derived from LLVM BB %172
> 
>             Live Ins: %R4 %R6 %D8 %Q5 %R9 %R7 %R8 %R10 %R5 %R11
> 
>             Predecessors according to CFG: BB#52
> 
> 8976B           %R1<def> = t2LEApcrelJT <jt#2>, 2, pred:14, pred:%noreg
> 
> 8992B           %R1<def> = t2ADDrs %R1<kill>, %R10, 18, pred:14, pred:%noreg, opt:%noreg
> 
> 9004B           %LR<def> = t2MOVi 1, pred:14, pred:%noreg, opt:%noreg
> 
> 9008B           t2BR_JT %R1<kill>, %R10<kill>, <jt#2>, 2
> 
>  
> 
> Shrink JT: t2BR_JT %R1<kill>, %R10<kill>, <jt#2>, 2
> 
>      addr: %R1<def> = t2ADDrs %R1<kill>, %R10, 18, pred:14, pred:%noreg, opt:%noreg
> 
>       lea: %R1<def> = t2LEApcrelJT <jt#2>, 2, pred:14, pred:%noreg
> 
>  
> 
>  
> 
> From this we see that the BaseReg = R1. R1 also happens to be the register used in the t2ADDrs calculation as well as defined by the t2LEApcrelJT operation. Because R1 is defined by t2LEApcrelJT, the restriction is met.
> 
>  
> 
> However, in the next example, it fails:
> 
>  
> 
> 5808B   BB#30: derived from LLVM BB %105
> 
>             Live Ins: %R4 %R6 %D8 %Q5 %R9 %R7 %R8 %R10 %R5 %R11
> 
>             Predecessors according to CFG: BB#29
> 
> 5840B           %R3<def> = t2LEApcrelJT <jt#1>, 1, pred:14, pred:%noreg
> 
> 5856B           %R2<def> = t2ADDrs %R3<kill>, %R7, 18, pred:14, pred:%noreg, opt:%noreg
> 
> 5872B           t2BR_JT %R2<kill>, %R7<kill>, <jt#1>, 1
> 
>             Successors according to CFG: BB#90(17) BB#31(17) BB#32(17) BB#33(17) BB#34(17) BB#51(17)
> 
>  
> 
> Here we see that the BaseReg = R2. But the t2LEApcrelJT instruction defines R3, not R2. But this is should be fine, because the t2ADDrs instruction takes R3 and defines R2, which is the real base address.
> 
>  
> 
> So my question is why is the restriction LeaMI->getOperand(0).getReg() != BaseReg there? Shouldn’t the restriction be trying to ensure that the register defined by t2LEApcrelJT also be the register used by the t2ADDrs instruction? It seems this test is being overly restrictive.
> 
>  
> 
> Daniel
> 
> 
> _______________________________________________
> LLVM Developers mailing list
> LLVMdev at cs.uiuc.edu         http://llvm.cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20130729/b2f895ea/attachment.html>