Resolving sizeof's; target triples; type optimizations

Hi all,

A few different though somewhat related questions here. I’m really grateful for your answers!

  1. From a previous question I know that sizeof’s are resolved into literals early in the front-end before IR is even emitted. It seems that they are resolved into whatever value is correct for the host machine. But if one wishes to then take the IR and emit assembly code for some other platform (in other words, to cross-compile), what’s the correct way to go about it? How to make the front-end resolve the sizeof’s according to some other target triple? For example, the sizeof of a C struct of one int and one char is resolved into 8 and not 5, but for some platforms out there 5 might be the correct answer. Will the “target datalayout” information in the IR file affect this in any way?

  2. What is that “target triple” as appears in the IR code? Who inserts it? Who takes note of it, if at all? Can the front-end be made to use some other triple than the native one? Does it affect the backends at all?

  3. In my code, I see an array of 8 x i8 being optimized into i64 (so that memcpy is then optimized into mov) by some optimization transformation(s). I want to prevent this from happening. What optimization does this kind of type replacement? How can one prevent it alone from running (but keeping all other optmizations, i. e. not using -O0 or something similar).

Many thanks!!

Harel Cain

Hi all,

A few different though somewhat related questions here. I'm really grateful
for your answers!

1. From a previous question I know that sizeof's are resolved into literals
early in the front-end before IR is even emitted. It seems that they are
resolved into whatever value is correct for the host machine. But if one
wishes to then take the IR and emit assembly code for some other platform
(in other words, to cross-compile), what's the correct way to go about it?
How to make the front-end resolve the sizeof's according to some other
target triple? For example, the sizeof of a C struct of one int and one char
is resolved into 8 and not 5, but for some platforms out there 5 might be
the correct answer. Will the "target datalayout" information in the IR file
affect this in any way?

See http://llvm.org/docs/FAQ.html#platformindependent . If you're
dealing with C code, it isn't too hard to add a new target to clang;
send an email to cfe-dev if you need help with that. The "target
datalayout" information is purely a hint to the optimizers.

2. What is that "target triple" as appears in the IR code? Who inserts it?
Who takes note of it, if at all? Can the front-end be made to use some other
triple than the native one? Does it affect the backends at all?

In the workflow for compiling C, the clang frontend sets it, and the
backend uses it to pick the correct target the generate assembly for.

3. In my code, I see an array of 8 x i8 being optimized into i64 (so that
memcpy is then optimized into mov) by some optimization transformation(s). I
want to prevent this from happening. What optimization does this kind of
type replacement? How can one prevent it alone from running (but keeping all
other optmizations, i. e. not using -O0 or something similar).

IIRC, instcombine does this; I'm not sure why you would want to
disable it, though.

-Eli

Thanks for the answers.

See http://llvm.org/docs/FAQ.html#platformindependent . If you’re
dealing with C code, it isn’t too hard to add a new target to clang;
send an email to cfe-dev if you need help with that. The “target
datalayout” information is purely a hint to the optimizers.

My experience this far has been mainly with using llvm-gcc and llvmc, especially using llvmc as a cross-compiler, with passing the correct flags (mainly -march) to llc via llvmc so as to pick for example ARM, MIPS and C backends.
In doing this, I’m essentially overriding any “target triple” existing in the IR, breaking up the process into a multi-stage operation (llvm-gcc, opt and llc running independently) and picking whatever target I want, right?

What’s the proper way to use clang as a cross-compiler (for example with the ARM or C backends) then?

The reason I want to avoid optimizing 8 x i8 into i64 is that I use the C backend (with the resultant C files as input to legacy compilers on esoteric platforms), and i64s turn into long longs that don’t exist on those platforms.

Thanks again,
Harel Cain

Hello

In doing this, I'm essentially overriding any "target triple" existing in
the IR, breaking up the process into a multi-stage operation (llvm-gcc, opt
and llc running independently) and picking whatever target I want, right?

No. This won't work since, as indicated in the FAQ entry IR derived
from C/C++ is not target neutral, so you cannot 'just' override the
triple.
You have to build/use llvm-gcc/clang as cross-compiler.

I guess that involves some special flags when running ‘configure’ before building the front-end. Right? Is there any handy documentation on how to build/use llvm-gcc as a cross-compiler?

Harel