Cameron Zwarich
2013-Feb-17 05:09 UTC
[LLVMdev] keeping instructions in order and hidden dependencies
AFAIK, You have two choices: use a pseudo that is lowered into separate instructions later as part of asm emission, or use MI bundles. The former is generally what existing targets use for this sort of thing, but perhaps the second would work better for you. Cameron On Feb 16, 2013, at 8:37 PM, Reed Kotler <rkotler at mips.com> wrote:> Some of my pseudos do conditional branch .+4 and such. > > I don't want the instruction scheduler to get creative on me. > > On 02/16/2013 07:20 PM, reed kotler wrote: >> I have some pseudos that I am expanding in the Mips 16 port. Currently >> they are blasted in one chunk as a multi line instruction sequence but I >> am changing the code now to expand them >> after register allocation. >> >> They are essentially macros and I need to make sure, at this time at >> least, that the individual instructions are not reordered or moved around. >> >> There are dependencies sometimes between the instructions that I'm not >> sure how to tell LLVM about. >> >> For example, this first one is a two instruction macro where register T8 >> is implicitly set by the first instruction and used by the second >> instruction. >> >> T8 is not a mips16 registers but some instructions use it implicitly and >> it can function as >> a condition code register. >> >> In this first case, I do a compare (CMP) and the result sets T8 and then >> the branch instruction following it uses this as if it were a condition >> code register. >> >> Maybe I can just set first instruction as defining T8 and the second as >> using it for the last time. >> >> Without this expansion, this is not an issue. >> >> At some future time it might be possible to reuse this condition >> register later and move it around but for now I'm not needing that >> optimization. >> >> Tia. >> >> Reed > > > _______________________________________________ > LLVM Developers mailing list > LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu > http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
Reed Kotler
2013-Feb-17 05:20 UTC
[LLVMdev] keeping instructions in order and hidden dependencies
One of my reasons for lowering things early is that I need to get an accurate count of the size of things. Some of the pseudos even have instructions like compare immediate, which in Mips 16 has two forms depending on the magnitude of the immediate field. Though I suppose it's possible to leave things as a pseudo and calculate their size, though I'm not sure where I could store the result. I need the size for long/short jump optimization, constant islands, delay slot filling (on mips16 you can't put a 32 bit instruction in a delay slot) and other things. I have not used bundles. Is the idea that I could create a bundle and insert the instructions in the bundle and then they are guaranteed to stay together? On 02/16/2013 09:09 PM, Cameron Zwarich wrote:> AFAIK, You have two choices: use a pseudo that is lowered into separate instructions later as part of asm emission, or use MI bundles. The former is generally what existing targets use for this sort of thing, but perhaps the second would work better for you. > > Cameron > > On Feb 16, 2013, at 8:37 PM, Reed Kotler <rkotler at mips.com> wrote: > >> Some of my pseudos do conditional branch .+4 and such. >> >> I don't want the instruction scheduler to get creative on me. >> >> On 02/16/2013 07:20 PM, reed kotler wrote: >>> I have some pseudos that I am expanding in the Mips 16 port. Currently >>> they are blasted in one chunk as a multi line instruction sequence but I >>> am changing the code now to expand them >>> after register allocation. >>> >>> They are essentially macros and I need to make sure, at this time at >>> least, that the individual instructions are not reordered or moved around. >>> >>> There are dependencies sometimes between the instructions that I'm not >>> sure how to tell LLVM about. >>> >>> For example, this first one is a two instruction macro where register T8 >>> is implicitly set by the first instruction and used by the second >>> instruction. >>> >>> T8 is not a mips16 registers but some instructions use it implicitly and >>> it can function as >>> a condition code register. >>> >>> In this first case, I do a compare (CMP) and the result sets T8 and then >>> the branch instruction following it uses this as if it were a condition >>> code register. >>> >>> Maybe I can just set first instruction as defining T8 and the second as >>> using it for the last time. >>> >>> Without this expansion, this is not an issue. >>> >>> At some future time it might be possible to reuse this condition >>> register later and move it around but for now I'm not needing that >>> optimization. >>> >>> Tia. >>> >>> Reed >> >> >> _______________________________________________ >> LLVM Developers mailing list >> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
Reed Kotler
2013-Feb-17 05:31 UTC
[LLVMdev] keeping instructions in order and hidden dependencies
At some future time I will need to teach LLVM more about what I'm doing so that it can schedule the individual instructions but now I'm mostly concerned about not having bugs from subtle optimization issues. On 02/16/2013 09:20 PM, Reed Kotler wrote:> One of my reasons for lowering things early is that I need to get an > accurate count of the size of things. > > Some of the pseudos even have instructions like compare immediate, which > in Mips 16 has two forms depending on the magnitude of the immediate field. > > Though I suppose it's possible to leave things as a pseudo and calculate > their size, though I'm not sure where I could store the result. > > I need the size for long/short jump optimization, constant islands, > delay slot filling (on mips16 you can't put a 32 bit instruction in a > delay slot) and other things. > > I have not used bundles. > > Is the idea that I could create a bundle and insert the instructions in > the bundle and then they are guaranteed to stay together? > > On 02/16/2013 09:09 PM, Cameron Zwarich wrote: >> AFAIK, You have two choices: use a pseudo that is lowered into >> separate instructions later as part of asm emission, or use MI >> bundles. The former is generally what existing targets use for this >> sort of thing, but perhaps the second would work better for you. >> >> Cameron >> >> On Feb 16, 2013, at 8:37 PM, Reed Kotler <rkotler at mips.com> wrote: >> >>> Some of my pseudos do conditional branch .+4 and such. >>> >>> I don't want the instruction scheduler to get creative on me. >>> >>> On 02/16/2013 07:20 PM, reed kotler wrote: >>>> I have some pseudos that I am expanding in the Mips 16 port. Currently >>>> they are blasted in one chunk as a multi line instruction sequence >>>> but I >>>> am changing the code now to expand them >>>> after register allocation. >>>> >>>> They are essentially macros and I need to make sure, at this time at >>>> least, that the individual instructions are not reordered or moved >>>> around. >>>> >>>> There are dependencies sometimes between the instructions that I'm not >>>> sure how to tell LLVM about. >>>> >>>> For example, this first one is a two instruction macro where >>>> register T8 >>>> is implicitly set by the first instruction and used by the second >>>> instruction. >>>> >>>> T8 is not a mips16 registers but some instructions use it implicitly >>>> and >>>> it can function as >>>> a condition code register. >>>> >>>> In this first case, I do a compare (CMP) and the result sets T8 and >>>> then >>>> the branch instruction following it uses this as if it were a condition >>>> code register. >>>> >>>> Maybe I can just set first instruction as defining T8 and the second as >>>> using it for the last time. >>>> >>>> Without this expansion, this is not an issue. >>>> >>>> At some future time it might be possible to reuse this condition >>>> register later and move it around but for now I'm not needing that >>>> optimization. >>>> >>>> Tia. >>>> >>>> Reed >>> >>> >>> _______________________________________________ >>> LLVM Developers mailing list >>> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev
Cameron Zwarich
2013-Feb-17 07:49 UTC
[LLVMdev] keeping instructions in order and hidden dependencies
You are trying to do a few different things here, and a uniform solution may not work for all of them. For a fixed instruction sequence, e.g. a special kind of move-and-branch sequence used for tail calls, you probably want a pseudo. If you are trying to combine arbitrary instructions together, e.g. Thumb IT blocks, you probably want to use bundles, even if the sequences are a fixed length. I think bundles might work well for your delay slots. Bundles will force instructions to stay together in sequence, and there is really no other way to do this with arbitrary instructions. If you have a compare instruction that generates a flag and a branch instruction that uses it, you want to model this dependency with an unallocatable physical register, e.g. the CPSR on ARM: // Condition code registers. def CCR : RegisterClass<"ARM", [i32], 32, (add CPSR)> { let CopyCost = -1; // Don't allow copying of status registers. let isAllocatable = 0; } This will require you to mark defs and uses (and kills), but if you're using it purely inside of a single basic block you won't have to worry about it too much. If you have instructions that may only generate this dependency in certain circumstances, you can use implicit def/use operands, which can be added to any MachineInstr even if the register is not part of the normal def/use list. If you need the instructions to always be adjacent and not simply satisfy a physical register dependency, then you'll have to use pseudos or bundles. For long/short jump optimizations you probably want to optimistically assume that all jumps are short and add a pass to relax them prior to asm emission, but I don't know the exact constraints of MIPS. On ARM this is done as part of the constant island pass. I CC'd Jim because he's the expert on the ARM constant island formation and he can give you some advice from experience there. I have always avoided that code. :-) Cameron On Feb 16, 2013, at 9:20 PM, Reed Kotler <rkotler at mips.com> wrote:> One of my reasons for lowering things early is that I need to get an accurate count of the size of things. > > Some of the pseudos even have instructions like compare immediate, which in Mips 16 has two forms depending on the magnitude of the immediate field. > > Though I suppose it's possible to leave things as a pseudo and calculate their size, though I'm not sure where I could store the result. > > I need the size for long/short jump optimization, constant islands, delay slot filling (on mips16 you can't put a 32 bit instruction in a delay slot) and other things. > > I have not used bundles. > > Is the idea that I could create a bundle and insert the instructions in the bundle and then they are guaranteed to stay together? > > On 02/16/2013 09:09 PM, Cameron Zwarich wrote: >> AFAIK, You have two choices: use a pseudo that is lowered into separate instructions later as part of asm emission, or use MI bundles. The former is generally what existing targets use for this sort of thing, but perhaps the second would work better for you. >> >> Cameron >> >> On Feb 16, 2013, at 8:37 PM, Reed Kotler <rkotler at mips.com> wrote: >> >>> Some of my pseudos do conditional branch .+4 and such. >>> >>> I don't want the instruction scheduler to get creative on me. >>> >>> On 02/16/2013 07:20 PM, reed kotler wrote: >>>> I have some pseudos that I am expanding in the Mips 16 port. Currently >>>> they are blasted in one chunk as a multi line instruction sequence but I >>>> am changing the code now to expand them >>>> after register allocation. >>>> >>>> They are essentially macros and I need to make sure, at this time at >>>> least, that the individual instructions are not reordered or moved around. >>>> >>>> There are dependencies sometimes between the instructions that I'm not >>>> sure how to tell LLVM about. >>>> >>>> For example, this first one is a two instruction macro where register T8 >>>> is implicitly set by the first instruction and used by the second >>>> instruction. >>>> >>>> T8 is not a mips16 registers but some instructions use it implicitly and >>>> it can function as >>>> a condition code register. >>>> >>>> In this first case, I do a compare (CMP) and the result sets T8 and then >>>> the branch instruction following it uses this as if it were a condition >>>> code register. >>>> >>>> Maybe I can just set first instruction as defining T8 and the second as >>>> using it for the last time. >>>> >>>> Without this expansion, this is not an issue. >>>> >>>> At some future time it might be possible to reuse this condition >>>> register later and move it around but for now I'm not needing that >>>> optimization. >>>> >>>> Tia. >>>> >>>> Reed >>> >>> >>> _______________________________________________ >>> LLVM Developers mailing list >>> LLVMdev at cs.uiuc.edu http://llvm.cs.uiuc.edu >>> http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev > >
Maybe Matching Threads
- [LLVMdev] keeping instructions in order and hidden dependencies
- [LLVMdev] keeping instructions in order and hidden dependencies
- [LLVMdev] keeping instructions in order and hidden dependencies
- [LLVMdev] keeping instructions in order and hidden dependencies
- [LLVMdev] keeping instructions in order and hidden dependencies