On 06/09/12 20:24, Chris Lattner wrote:> On Sep 6, 2012, at 8:51 AM, Richard Osborne <richard at xmos.com> wrote: > >> I recently noticed that all globals bigger than 16 bytes are being 16 byte aligned by LLVM (assuming there isn't an explicitly requested alignment). I'd really rather avoid this, at least for the XCore backend. I tracked this down to the following code in TargetData.cpp: >> >> if (GV->hasInitializer() && GVAlignment == 0) { >> if (Alignment < 16) { >> // If the global is not external, see if it is large. If so, give it a >> // larger alignment. >> if (getTypeSizeInBits(ElemType) > 128) >> Alignment = 16; // 16-byte alignment. >> } >> } >> >> I was a bit surprised to see these numbers hardcoded in TargetData since everything else is taken from the datalayout string. I was wondering what the logic was behind the number 16. Would it make sense to derive this number from the other alignments somehow (e.g. the maximum preferred alignment across all types). Alternatively would it make sense to make it configurable in the datalayout string? > I don't think that there is any specific logic to it. It was just a heuristic that "made sense at the time", not based on any scientific evaluation. If you want to raise it to some other arbitrary limit (say 32 bytes) that would be fine given some performance analysis. If you want to design a way to express this in target data, please propose a specific way to model it. > > -ChrisI want it to be configurable. On my target there is no advantage to aligning anything by more than a 4 bytes - it just wastes space. I'll try to put together a proposal for making it possible to set this per target. -- Richard Osborne | XMOS http://www.xmos.com
On Sep 7, 2012, at 8:02 AM, Richard Osborne <richard at xmos.com> wrote:>>> I was a bit surprised to see these numbers hardcoded in TargetData since everything else is taken from the datalayout string. I was wondering what the logic was behind the number 16. Would it make sense to derive this number from the other alignments somehow (e.g. the maximum preferred alignment across all types). Alternatively would it make sense to make it configurable in the datalayout string? >> I don't think that there is any specific logic to it. It was just a heuristic that "made sense at the time", not based on any scientific evaluation. If you want to raise it to some other arbitrary limit (say 32 bytes) that would be fine given some performance analysis. If you want to design a way to express this in target data, please propose a specific way to model it. >> >> -Chris > I want it to be configurable. On my target there is no advantage to aligning anything by more than a 4 bytes - it just wastes space. I'll try to put together a proposal for making it possible to set this per target.Ok, after considering this a bit more, how about we take a different approach: despite its name, TargetData is really mostly for consumption by the mid-level optimizers. getPreferredAlignment is really something that only TargetLowering[ObjectFile] should be using. What do you think about sinking getPreferredAlignment down to it? This will make it much easier to target-hookize this. -Chris
On 07/09/12 18:13, Chris Lattner wrote:> On Sep 7, 2012, at 8:02 AM, Richard Osborne <richard at xmos.com> wrote: >>>> I was a bit surprised to see these numbers hardcoded in TargetData since everything else is taken from the datalayout string. I was wondering what the logic was behind the number 16. Would it make sense to derive this number from the other alignments somehow (e.g. the maximum preferred alignment across all types). Alternatively would it make sense to make it configurable in the datalayout string? >>> I don't think that there is any specific logic to it. It was just a heuristic that "made sense at the time", not based on any scientific evaluation. If you want to raise it to some other arbitrary limit (say 32 bytes) that would be fine given some performance analysis. If you want to design a way to express this in target data, please propose a specific way to model it. >>> >>> -Chris >> I want it to be configurable. On my target there is no advantage to aligning anything by more than a 4 bytes - it just wastes space. I'll try to put together a proposal for making it possible to set this per target. > Ok, after considering this a bit more, how about we take a different approach: despite its name, TargetData is really mostly for consumption by the mid-level optimizers. getPreferredAlignment is really something that only TargetLowering[ObjectFile] should be using. What do you think about sinking getPreferredAlignment down to it? This will make it much easier to target-hookize this. > > -ChrisHi Chris, Sorry for not replying sooner but I've haven't had a chance to look at this until now. I think your suggestion makes a lot of sense. The preferred alignment will be a lot easier to change if it is not embedded in the IR in the form of the target data string. I started trying to moving the functions to get preferred alignment to TargetLowering but found there are a few mid-level optimizers that make use of the preferred alignment. In particular: * lib/Analysis/ValueTracking.cpp - uses getPreferredAlignment() to determine number of trailing zeros of the address of a global defined in the current module. * lib/Transforms/Utils/InlineFunction.cpp - Uses getPrefTypeAlignment() to set the alignment of the alloca it introduces when a byval argument is inlined * lib/Transforms/IPO/ConstantMerge.cpp - Uses getPreferredAlignment() to determine the alignment of the merged constant. * lib/Transforms/Vectorize/BBVectorize.cpp - Depending on what options are set this pass avoids vectorization if it would introduce loads and store of less than the alignment returned by getPreferredAlignment(). * lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp - getPrefTypeAlignment() is used to explictly set the alignment of allocas with unspecified alignment. Also instcombine tries to enforce preferred alignment for loads and stores if possible. Does it seem reasonable to change these to use ABI alignment? -- Richard Osborne | XMOS http://www.xmos.com