GCC vs. LLVM Clang Is Mixed On The Ivy Bridge Extreme

[XorEaxEax]

One possibility (which may or may not be the case) is that vectorization is at fault here.

Yes it's quite obvious that something in the heuristics failed for this code atleast, there's no reason why less precise floating point math should result in much slower code.

So it likely ties back to what you said earlier that Clang/LLVM will only try to vectorize with -ffast-math enabled, and that it is indeed the vectorization that fails.

The architecture specific stuff may be linked to the inaccurate cost model issue?

You mean -march=native ? I don't see how, both '-Ofast -funroll-loops' and '-Ofast -funroll-loops -march=native' where equally slow on Clang.

On the other hand '-march=native' didn't seem to do anything for GCC either, I think results around 50 milliseconds or less can be discared as noise.

It's funny though that Michael obviously knew about this problem with -ffast-math on Clang/LLVM, as the upstream (original) version of C-Ray 1.1 ships with '-O3 -ffast-math' in the makefile, and thus Michael modified the makefile to remove -ffast-math for his tests so as to make Clang/LLVM not look so bad. Good old agenda-biased Michael.

This is why I take all 'tests' done here on Phoronix with a large grain of salt, particularly those between GCC and Clang/LLVM as I know he is extremely pro Clang/LLVM and has an agenda against FSF (which seems to spill over on GCC and other FSF/GNU software).

I guess 3.4 will be released in the next month or two, and it would be interesting to revisit this at that point.

Indeed, the competition between GCC and Clang/LLVM is a perfect situation for us as end users (and I think for the projects themselves). I use both, though GCC is definitely what I use for release binaries due to it's better optimizations, if this changes and Clang/LLVM delivers better code performance, I will use that toolchain for release builds.

So here's looking forward to Clang 3.4 and GCC 4.9 and the advances they bring.

[name99]

I did a quick rundown test on C-Ray using GCC and Clang with and without -ffast-math:

GCC version: 4.8.1 20130725
Clang version: 3.3 (tags/RELEASE_33/final)
Arch Linux 64-bit, core i5
Benchmark: cat scene | ./c-ray-mt -t 4 -s 7500x3500 > foo.ppm

results are in milliseconds, and is the average of 5 benchmark-runs (exluding a varm-up run)

gcc -O3
5840

gcc -O3 -funroll-loops
5704

gcc -O3 -ffast-math -funroll-loops
4374

gcc -Ofast -funroll-loops
4368

gcc -Ofast -funroll-loops -march=native
4351

On GCC we can see that -ffast-math greatly improves the result, now let's look at Clang:

clang -O3
6403

clang -O3 -funroll-loops
6396

clang -O3 -ffast-math -funroll-loops
7137

clang -Ofast -funroll-loops
7122

clang -Ofast -funroll-loops -march=native
7153

On Clang however, we see that -ffast-math _degrades_ performance markedly on C-Ray, so had Michael used it for his Phoronix C-Ray test then Clang would have come out looking MUCH worse than it does now since GCC got a great boost from -ffast-math.

Apart from that it seems that -funroll-loops does nothing performance-wise on Clang, and same goes for -march=native.

Interesting (and remarkable) that we get such a regression from -ffast-math. It'd be interesting (if it's not a hassle) to learn why.
One possibility (which may or may not be the case) is that vectorization is at fault here. A big push in 3.3 was to ensure that the vectorization cost model was accurate, so that your vectorized code didn't make things worse by spending so much time just shuffling data. But the hope of having an accurate cost model doesn't mean that you ACTUALLY have one. It's possible that there's something severely broken in the cost model (at least for FP vectors) which is giving us these results.

The unroll-loops does not surprise me. The LLVM guys probably believe they have good heuristics for when (or not) to do this and are likely correct.
The architecture specific stuff may be linked to the inaccurate cost model issue? It would be amusing if we learned there was an off-by-one error or something in the micro-architecture specs table that drove the compiler!

I guess 3.4 will be released in the next month or two, and it would be interesting to revisit this at that point.

[XorEaxEax]

Just noting, gcc accepts -O[any positive number], it's just that high numbers get clamped to 3. It's been this way for ages, gcc 4.2 accepts -O666 just fine.

Good to know, I thought it just ignored anything above -O3 and used the default -O0, clamping at -O3 makes more sense though as the user likely wanted aggressive optimization when attempting a higher value than -O3.

[curaga]

Just noting, gcc accepts -O[any positive number], it's just that high numbers get clamped to 3. It's been this way for ages, gcc 4.2 accepts -O666 just fine.

[XorEaxEax]

Oh, one thing to add to my earlier comment.
LLVM (and maybe GCC, but I don't know there) will not automatically vectorize many FP loops if fast-math is not enabled because getting the loop to vectorize requires re-ordering FP operations. This means that using fast-math, if your code allows it, can affect performance by quite a bit more than you might imagine.

I did a quick rundown test on C-Ray using GCC and Clang with and without -ffast-math:

GCC version: 4.8.1 20130725
Clang version: 3.3 (tags/RELEASE_33/final)
Arch Linux 64-bit, core i5
Benchmark: cat scene | ./c-ray-mt -t 4 -s 7500x3500 > foo.ppm

results are in milliseconds, and is the average of 5 benchmark-runs (exluding a varm-up run)

gcc -O3
5840

gcc -O3 -funroll-loops
5704

gcc -O3 -ffast-math -funroll-loops
4374

gcc -Ofast -funroll-loops
4368

gcc -Ofast -funroll-loops -march=native
4351

On GCC we can see that -ffast-math greatly improves the result, now let's look at Clang:

clang -O3
6403

clang -O3 -funroll-loops
6396

clang -O3 -ffast-math -funroll-loops
7137

clang -Ofast -funroll-loops
7122

clang -Ofast -funroll-loops -march=native
7153

On Clang however, we see that -ffast-math _degrades_ performance markedly on C-Ray, so had Michael used it for his Phoronix C-Ray test then Clang would have come out looking MUCH worse than it does now since GCC got a great boost from -ffast-math.

Apart from that it seems that -funroll-loops does nothing performance-wise on Clang, and same goes for -march=native.

[XorEaxEax]

LLVM (and so XCode) also has link time (ie whole program) optimization, enabled by -O4. I imagine GCC has the same.

GCC activates link time optimization using -flto (which is also supported as a flag by Clang), -O4 is not recognized at all on GCC afaik.

Apple slides showed that LTO made a substantial difference (5% to 20%) in performance, but of course that is against real world code that is split over a large number of files; it may have much less impact on these sorts of microbenchmarks.

Yes it's extremely code-base dependent, you can basically achieve the same effect by manually declaring non exported functions as static (or like sqlite did, join together all source files into one big file before compiling), personally I've seen little performance gain from LTO on my own code and the code I've benchmarked, but again it really depends on the code in question, also the binary often ends up quite a bit smaller with LTO which is of course nice.

What's not clear to me is the extent to which either LLVM or GCC have fully optimized their LTO pass. Apple had (PPC specific) tools fifteen years ago that could run whole program optimization and rearrange the function layout so that functions that called each other were packed together (and so took up less TLB coverage and shared overlapping cache lines).

Well GCC has a '-fwhole-program' option which enables more aggressive interprocedural optimizations (as in, moving blocks of code around to improve cache use, eliminate/consolidate code blocks etc).

but could be run with a profiling pass to get a better understanding of the hot call chains. But as far as I know, the LLVM LTO does not (yet?) do this sort of thing, and I have zero idea about GCC.

Yes, this (profile guided optimization) is the by far best performance giving optimization I've used which is outside of the -On levels, GCC has this implemented as -fprofile-generate/-fprofile-use and it can deliver some great improvements, typically I get between ~4-8% on performance oriented code, sometimes up to 20%.

I know there was a Google Summer of Code project to implement profile guided optimization into LLVM but I haven't heard anything about it since so I fear it didn't amount to anything.

[XorEaxEax]

Where in the wild do we actually see O3 though?

Practically all performance oriented software, like encoders, games, archivers/compressors, emulators, 3d renderers etc

[WorBlux]

What's really bad with the Botan tests is that they use -O2 which makes those tests worthless when it comes to comparing compiler vs compiler code performance. As I've explained to Michael over and over, there are no rules for which optimizations a compiler should add at -O2, meaning that if compiler A decides to enable more optimizations than compiler B at that level, A will win which in turn says nothing of how the compilers compare when set to their HIGHEST optimization level which is what you use when you want the FASTEST code, which again is what all tests except the 'build time' test here was measuring.

Again, this is why you compare the compilers at their highest optimization level (-O3, and when it was used GCC won all performance tests) because at this level the compilers compete on fair ground, which is to create the fastest binary they can. Now -O2 can be interesting as sometimes -O2 beats -O3 due to optimization heuristics failing, but only in the context of actually having -O3 results to compare with.

I don't know why Michael persist at doing this, unless he is consciously trying to cook the results in order to get some pointless Clang/LLVM wins.

Where in the wild do we actually see O3 though?

[name99]

Oh, one thing to add to my earlier comment.
LLVM (and maybe GCC, but I don't know there) will not automatically vectorize many FP loops if fast-math is not enabled because getting the loop to vectorize requires re-ordering FP operations. This means that using fast-math, if your code allows it, can affect performance by quite a bit more than you might imagine.

[name99]

Don't forget, for gcc there's also:
-Ofast
Disregard strict standards compliance. -Ofast enables all -O3 optimizations. It also enables optimizations that are not valid for all standard-compliant programs. It turns on -ffast-math and the Fortran-specific -fno-protect-parens and -fstack-arrays.

XCode 5 also has a -Ofast (whose main effect is fast math, but also sets one or two other settings that I forget). I don't know if that has made it into mainline LLVM yet.

[XCode's defaults are also now for new projects to be very aggressive about assuming that pointers don't alias. It's likely that this is an Apple developer mandate (they provide guidelines about how you should write your code to not make use of aliased pointers, eg through use of unions rather than pointer casting), and that mainline LLVM is not as aggressive. So this is an interesting point, but may not be relevant to the Phoronix community which, I am guessing, is more interested in having crappy old code compile properly than in following a company mandate about how to write better performant new code.]

LLVM (and so XCode) also has link time (ie whole program) optimization, enabled by -O4. I imagine GCC has the same. It seems only sensible that this should be activated when both compilers are tested against each other. Apple slides showed that LTO made a substantial difference (5% to 20%) in performance, but of course that is against real world code that is split over a large number of files; it may have much less impact on these sorts of microbenchmarks.
What's not clear to me is the extent to which either LLVM or GCC have fully optimized their LTO pass. Apple had (PPC specific) tools fifteen years ago that could run whole program optimization and rearrange the function layout so that functions that called each other were packed together (and so took up less TLB coverage and shared overlapping cache lines). They used heuristics in the absence of anything better, but could be run with a profiling pass to get a better understanding of the hot call chains. But as far as I know, the LLVM LTO does not (yet?) do this sort of thing, and I have zero idea about GCC.
You can do even better with LTO (either heuristics or profile-directed). Rather than rearranging and repacking the code based on functions, you can do so based on basic-blocks so that if(!error){...}else(/*handle error*/...} moves the error handling code far away to the end of your binary. Now your actual binary (on disk and in RAM) looks like a series of basic blocks that jump between each other, not a series of functions --- looks weird yes, and more difficult to reverse-engineer from binary, but perfectly legit and not in any way fragile. Obviously this gives you even better utilization of both your I$ and your TLB. This has been done academically but I don't know if any commercial products (ICC? Dev Studio?) use it. You can also do the same sort of repacking with global/static data layout to try to get data that is frequently used together on the same cache line.