| 1 |
On Wednesday 01 August 2007, Arturo Garcia wrote: |
| 2 |
> I am looking for resources that explain |
| 3 |
|
| 4 |
*shrug* if you find any, please post to the list for others |
| 5 |
|
| 6 |
> the differences between glibc, glibceabi, eabi, |
| 7 |
> softfloat, as you can see... I'm a bit lost. |
| 8 |
|
| 9 |
there are two big libc players: |
| 10 |
- glibc: full featured (all of POSIX and then some) |
| 11 |
- uclibc: meant to be small, but lacks some features (critically so for some) |
| 12 |
|
| 13 |
in the arm world, there are two ABIs: |
| 14 |
- oabi: "Old ABI": the classic ARM ABI |
| 15 |
- eabi: "Embedded ABI": the new ARM ABI |
| 16 |
EABI is "better" in terms of size and being able to address design decisions |
| 17 |
that ended up being problematic in OABI ... something that you really only |
| 18 |
find out about after using the same ABI for years and years. unfortunately, |
| 19 |
the support status of EABI can be a bit tipsy: you have to have very recent |
| 20 |
versions of everything (binutils, gcc, libc, kernel) or you can forget it. |
| 21 |
|
| 22 |
in the Linux world, there are a couple of ways to tackle the FPU issue: |
| 23 |
- put a hardware FPU on the chip/board and generate floating point |
| 24 |
instructions in userspace, thus increasing BOM while decreasing software |
| 25 |
development. |
| 26 |
- include software FPU emulation in the kernel and generate floating point |
| 27 |
instructions in userspace, thus (significantly) decreasing runtime |
| 28 |
performance will emulating these floating point operations in kernel space, |
| 29 |
but again you save time/money on software development. |
| 30 |
- build userspace with a softfloat toolchain, thus replacing floating point |
| 31 |
math with fixed point math and taking a runtime performance (but not nearly |
| 32 |
as bad as the kernel emulation) and continue to save time/money on software |
| 33 |
development. |
| 34 |
- dont use floating point. sometimes easy, sometimes hard, sometimes already |
| 35 |
done for you. depending on the project, this can be a huge sink in software |
| 36 |
development costs. typically audio/video projects get screwed here: everyone |
| 37 |
*wants* fixed point algorithms to do audio/video processing, but the large |
| 38 |
majority of people are fuckers and dont want to (1) help or (2) contribute |
| 39 |
their work back. |
| 40 |
|
| 41 |
so for ARM, the default environment is hardfloat userland with FPU emulation |
| 42 |
enabled in the kernel. if you want to go the softfloat route (which actually |
| 43 |
shouldnt be too shabby nowadays), just build a *-softfloat-* toolchain. |
| 44 |
|
| 45 |
HTH |
| 46 |
-mike |
Archives