Hello,
I'm trying to cross compile the MiBench benchmarks using LLC and ARM cross
compiler the binaries generated will be used on pandaboard or the qemu-arm
emulator.
I have faced multiple issues during the compilation process, here are the
commands used.
# generate LLVM-IR for dijkstra_large.c
$gcc -S -flto -o dijkstra_large.bc -fplugin=${DRAGONEGG} dijkstra_large.c
# generate TARGET object file for dijkstra_large.bc
$./llc -march=arm -mcpu=cortex-a9 dijkstra_large.bc -o dijkstra_large.o
-filetype=obj -O3
# Generate executalbe using Target's toolchain
$arm-linux-gcc -static dijkstra_large.o -o dijkstra_large
The first issue is that the arguments that I use as an input to dijkstra
benchmark do not get inserted in their appropriate position in the [argv] array.
In order to run the benchmark the command is ./dijkstra_large input.dat. using
this command I got the argv[1] = null which should be input.dat. If I add
another extra argument between the benchmark binary name and the input.dat
arugment like (./dijkstra_large xxx input.dat) the benchmark works without any
error and the argv[1] equals the input.dat. This issue is preventing me from
using other benchmarks which needs multiple arguments to run.
The second issue is that when I use "arm-linux-gnueabihf-gcc" cross
compiler instead of the arm-linux-gcc I get the following error
"/usr/lib/gcc/arm-linux-gnueabihf/4.6/../../../../arm-linux-gnueabihf/bin/ld:
error: a.out uses VFP register arguments, dijkstra_large.o does not
/usr/lib/gcc/arm-linux-gnueabihf/4.6/../../../../arm-linux-gnueabihf/bin/ld:
failed to merge target specific data of file dijkstra_large.o
collect2: ld returned 1 exit status"
I have tried to change the -mfloat-abi to hard or soft but still getting the
same error message. Is there any other options that forces the llvm to use VFP
registers?
Also it would be nice if anyone can point me to a tutorial which shows how to
build LLVM cross compiler using the clang.
Thanks,
Najem.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20131217/f33fea9a/attachment.html>
Hi Najem,> # generate LLVM-IR for dijkstra_large.c > $gcc -S -flto -o dijkstra_large.bc -fplugin=${DRAGONEGG} dijkstra_large.cIs your GCC configured as an arm-linux-gnueabihf compiler? If not, it's almost certainly going to be getting ABI details wrong. I'm slightly surprised that it affects argv, but it's certainly the first thing to think about.> # generate TARGET object file for dijkstra_large.bc > $./llc -march=arm -mcpu=cortex-a9 dijkstra_large.bc -o dijkstra_large.o -filetype=obj -O3If Dragonegg *is* generating for ARM then the "-march=arm" option to llc should be unnecessary. Also, GCC should probably have "-O3" as well, since llc is the backend code generator and doesn't run the mid-end phases.> # Generate executalbe using Target's toolchain > $arm-linux-gcc -static dijkstra_large.o -o dijkstra_largeI've not seen an arm-linux-gcc before, so I'm slightly worried about what ABI it's using (-gnu, -gnueabi or -gnueabihf).> The second issue is that when I use "arm-linux-gnueabihf-gcc" cross compiler > instead of the arm-linux-gcc I get the following error> Also it would be nice if anyone can point me to a tutorial which shows how > to build LLVM cross compiler using the clang.That sounds like a very good idea. For a start there are far more people here who know how to get clang to cooperate than DragonEgg (I've never even built DragonEgg, for example). We've got a couple of documents on that topic: http://llvm.org/docs/HowToCrossCompileLLVM.html http://clang.llvm.org/docs/CrossCompilation.html They're not perfect by any means yet (still fairly new), but it'd be great to hear your experiences trying to follow them so we can improve things. Cheers. Tim.