I was just looking at a few projects that want to generate code quickly but
in reasonable quality.

This post here about a jit compiler project (as yet unfinished), says that
for GCC there are about 300 passes, but that most of those don't buy you
much.

But he wrote "Recently, I did an experiment by switching on only a fast and
simple RA and combiner in GCC. There are no options to do this, I needed to
modify GCC. (I can provide a patch if somebody is interested.) Compared to
hundreds of optimizations in GCC-9.0 with -O2, these two optimizations
achieve almost 80% performance on an Intel i7-9700K machine under Fedora
Core 29 for real-world programs through SpecCPU, one of the most credible
compiler benchmark suites..."

So, turning on only two, admittedly involved, optimization passes got him
70% of the speed of GCC -O2 on spec cpu.

So maybe a well chosen optimization layer would be worthwhile.  His project
is interesting, it intends to implement a compiler orders of magnitude
faster than LLVM's own jit, which could probably gain real world speed ups
of several times just by getting rid of the start-up lag (it takes a lot of
book keeping in LLVM before it can start a compilation).



On Sun, Dec 27, 2020 at 3:32 AM Samir Ribić <[email protected]> wrote:

> Well, tcc is one of the fastest C compilers when we speak about
> *compilation speed*. However, this is achieved by generating code almost
> simultaneously with syntax analysis, without intermediate steps that use
> some other compilers. Compilers like gcc do the following:
> - Perform syntax analysis and generate abstract syntax tree
> - Perform semantic analysis over the tree
> - Optimize the tree
> - Generate intermediate code
> - Optimize intermediate code
> - Generate assembly code
> - Optimize assembly code
> - Generate machine code
> - Link to executable using external linker
>
> Tcc does not generate intermediate tree and platform neutral code. It has
> workflow more similar to the following
> - While analysing the syntax check the semantics and generate the required
> tables. If the syntax is correct, then generate machine code. immediately
> - Use internal linker and link with simple startup routine
>
> The result is much faster compilation and often shorter executable due to
> simple startup routine.
>
> However, this also means that tcc can not be the fastest C compiler when
> we speak about *execution speed*. For example. it is impossible to detect
> dead code. Look at this code
> a=200;
> b=10;
> a=b;
> Remark that first instruction is completely useless. The compilers who
> generate intermediate formats can detect it and remove this instruction
> from generated code.
>
> Similarly, sometimes variables do not need to be stored in  memory like
> this loop counter "i":
> for (int i=0;i<n;i++) pow=pow*b;
> After we remarked that the loop counter is used just to repeat
> multiplication n times, it is possible to generate machine code LOOP
> instruction with register counter, which is quite fast.
> Unfortunately, the compiler can not remark it without analysing the whole
> source file before generating assembly code.
>
> Should we make tcc to perform more like gcc and generate faster code? No!
> Adding intermediate formats and infinite optimizations will nullify two
> good tcc properties, fast compilation and small compiler size. However, I
> would like to see some simple code generating improvements, which are
> possible without changes in compiler design. For example, operations
> between variable and constant like a=5;  a>5;  a+5  etc  on 80386 can be
> done using one instruction while tcc generates two instructions.
>
>
> On Sat, Dec 26, 2020 at 9:05 PM Joshua Scholar <[email protected]>
> wrote:
>
>> I'm having fun getting to know the code, and writing features for tcc
>> that I make sure don't require large changes to the project and that my
>> additions themselves compile in tcc.
>>
>> But when I tested a hash function, I was disheartened to find that spooky
>> hash runs about 1/10 the speed compiled in tcc as compiled in gcc.
>>  Although gcc compiled with a no optimization setting makes code that runs
>> near the same speed as code compiled under tcc.
>>
>> So I started thinking that for the project I'm working on maybe I'd do
>> better to be learning to use LLVM than working on this.
>>
>> But since I wondered, well maybe tcc is bad at heavily optimizable loops,
>> but maybe it's ok on general code.
>>
>> So I took a copy of tcc that had been compiled with the microsoft
>> compiler and used it to compile itself, which of course goes MUCH faster
>> than compiling it with microsoft.
>>
>> Then I took the version of tcc that had been compiled under tcc and used
>> it to compile itself.
>>
>> And it still compiled so fast that I can hardly feel the difference.  I
>> don't THINK it was 10 times slower.
>>
>> So strange to be both impressed at how slow tcc's code is and how fast
>> tcc's code is within an hour.
>>
>> Joshua Scholar
>> _______________________________________________
>> Tinycc-devel mailing list
>> [email protected]
>> https://lists.nongnu.org/mailman/listinfo/tinycc-devel
>>
> _______________________________________________
> Tinycc-devel mailing list
> [email protected]
> https://lists.nongnu.org/mailman/listinfo/tinycc-devel
>
_______________________________________________
Tinycc-devel mailing list
[email protected]
https://lists.nongnu.org/mailman/listinfo/tinycc-devel

Reply via email to