<div dir="ltr">That actually sounds like a generally useful thing to do for any kind of code where relative frequency of execution can be predicted.<div><br></div><div>You don't need to put those cold and frozen areas very far apart. Certainly they should be within range of a normal conditional branch instruction! They only need to be in different memory pages, which usually these days means 4KB block. 386 can do 64KB jcc, so that should be fine.</div><div><br></div><div>Depending on how big your typical tracelets are you may want to make each area several blocks in size, to avoid too marge a percentage unused at the end of them, but probably not very many.</div></div><div class="gmail_extra"><br><div class="gmail_quote">On Wed, Nov 19, 2014 at 5:21 PM, Brett Simmers <span dir="ltr"><<a href="mailto:bsimmers@fb.com" target="_blank">bsimmers@fb.com</a>></span> wrote:<br><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">I'm part of a team working on adding an llvm codegen backend to HHVM (PHP JIT, <a href="http://hhvm.com" target="_blank">http://hhvm.com</a>) using MCJIT. We have a code layout problem and I'm looking for opinions on good ways to solve it.<br>
<br>
The short version is that the memory we emit code into is split into a few different areas, and we'd like a way to control which area each BasicBlock ends up in during codegen. I know this probably sounds pretty odd, so here's a much more detailed explanation:<br>
<br>
Our translation unit is a "tracelet" which is typically a single basic block of PHP code. We emit code one tracelet at a time as it's needed for execution, so the physical layout of the generated code looks like the execution path taken the first time the code is run (we're actually a little smarter than that now, but this description is good enough for the problem at hand).<br>
<br>
We're under pretty severe icache/iTLB pressure, so we do whatever we can to keep the hot path as compact as possible. One of the ways we do this is by dividing our code cache into three fixed-size areas: main, cold, and frozen. Our current, non-llvm codegen backend has one area tag per basic block, and most tracelets we compile will span all three areas. This means that if we emit code for tracelets A and then B, A's main code will be followed immediately by B's main code (and the same for cold/frozen). It ends up looking something like this:<br>
<br>
good layout: | A_main B_main <unused> | A_cold B_cold <unused> |<br>
A_frozen... |<br>
<br>
Main is the primary execution path. Cold contains code that we expect to happen rarely. Conditional branches from main to cold are generally taken <5% of the time. Frozen contains code that we expect to almost never run, but must have for correctness (mostly for handling exceptions). Conditional branches from main to frozen are generally taken <0.01% of the time. The fact that we have three areas isn't really relevant to the problem itself; I think any solution should allow an arbitrary number of areas.<br>
<br>
We've done some experiments where we collapse all three areas into one, so A's main code is followed by A's cold code, which is followed by A's frozen code, and so on for B and other tracelets, like this:<br>
<br>
bad layout: | A_main A_cold A_frozen B_main B_cold B_frozen... |<br>
<br>
This causes an unacceptable performance regression, so for our experimental llvm backend we need to come up with a solution.<br>
<br>
We're using a custom calling convention to model each tracelet as a single llvm function. I know there are intrinsics in llvm that allow us to give blocks relative probabilities, and if I'm understanding things correctly that should allow us to achieve the bad layout above, since the code for a given llvm function will be emitted in one contiguous chunk, regardless of how the blocks are organized within it. I'd like to figure out a way to achieve the good layout.<br>
<br>
The best idea I have so far is to annotate each BasicBlock with an area tag like we do in our existing codegen backend, and then teach the llvm codegen backend how to place each of those blocks in the appropriate part of our code cache. We have a custom subclass of RTDyldMemoryManager to get it emitting all code in the main area right now; I assume any solution would impact the API of this class to get at the different areas.<br>
<br>
If we have to split the main/cold/frozen parts of each tracelet into different llvm functions we might be able to make that work but I'm skeptical. My main concern is that cold/frozen code nearly always uses a bunch of values from the main code path that branched to it, and if the overhead of getting from main to cold/frozen is more than a single jmp/jcc instruction (on x86) that's going to be a dealbreaker.<br>
<br>
Does this sound doable, and is it something you'd be ok having in llvm? If we can come up with a good design my team is happy to do the actual work, though if anyone else is interested in doing it we certainly won't complain :).<br>
<br>
Thanks!<br>
Brett<br>
______________________________<u></u>_________________<br>
LLVM Developers mailing list<br>
<a href="mailto:LLVMdev@cs.uiuc.edu" target="_blank">LLVMdev@cs.uiuc.edu</a> <a href="http://llvm.cs.uiuc.edu" target="_blank">http://llvm.cs.uiuc.edu</a><br>
<a href="http://lists.cs.uiuc.edu/mailman/listinfo/llvmdev" target="_blank">http://lists.cs.uiuc.edu/<u></u>mailman/listinfo/llvmdev</a><br>
</blockquote></div><br></div>