This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: (R5900) Implementing Vector Support
- From: Woon yung Liu <ysai187 at yahoo dot com>
- To: Matthew Fortune <Matthew dot Fortune at imgtec dot com>, Richard Henderson <rth at redhat dot com>, Gcc Mailing List <gcc at gcc dot gnu dot org>
- Date: Sat, 25 Jun 2016 10:23:24 +0000 (UTC)
- Subject: Re: (R5900) Implementing Vector Support
- Authentication-results: sourceware.org; auth=none
- References: <23a57920-3e9e-05f9-e428-a7e5e89d4de9 at redhat dot com> <133091800 dot 362759 dot 1462978450117 dot JavaMail dot yahoo at mail dot yahoo dot com> <93d40024-8baf-f571-765e-3f3ae59961df at redhat dot com> <687383190 dot 1940916 dot 1463221310394 dot JavaMail dot yahoo at mail dot yahoo dot com> <850bcd56-a219-a153-b467-8414fa19c207 at redhat dot com> <300510214 dot 326837 dot 1463573792037 dot JavaMail dot yahoo at mail dot yahoo dot com> <e7ccb4e0-6b37-ea2c-01bb-4fff2742ca32 at redhat dot com> <1735082185 dot 4758122 dot 1463655944237 dot JavaMail dot yahoo at mail dot yahoo dot com> <1809107277 dot 905923 dot 1464508754638 dot JavaMail dot yahoo at mail dot yahoo dot com> <1ad4a4a6-72ee-3048-0b9d-53ed16c5a4c8 at redhat dot com> <904764586 dot 674015 dot 1464958440737 dot JavaMail dot yahoo at mail dot yahoo dot com> <6D39441BF12EF246A7ABCE6654B023537E43748E at HHMAIL01 dot hh dot imgtec dot org> <966846598 dot 974749 dot 1465009060396 dot JavaMail dot yahoo at mail dot yahoo dot com>
- Reply-to: Woon yung Liu <ysai187 at yahoo dot com>
Hi again,
I have managed to fix the problems mentioned earlier. Part of the problem (segmentation fault) was caused by logical operations with register modes (i.e. QImode, DImode) other than SImode.
The other issue was the use of gen_lowpart. Replacing all calls to gen_lowpart within the affected pattern with calls to simplify_gen_subreg, solved the ICEs.
I have also attempted to implement the PMADDH and PMSUBH patterns (parallel MADD and MSUB, for 8x16-bit values), but they don't work. I believe that it's because GCC can't find a use for them, given that parallel-multiplication is currently done with both inputs and outputs to only GPRs (doesn't expose the HI+LO registers).
Unfortunately, when I attempted to change the design of the mulv8hi3 pattern to be like the MIPS mulsidi3 pattern (output to HI+LO), GCC begun spewing assertion fails within expand_mult because of expand_binop when it tries to compile any multiplication between vectors. I haven't managed to find that function yet, but I guess that this means that I can't do that (probably due to an internal check against UNITS_PER_WORD?).
Regarding the auto-vectorization problems that I mentioned before, it seems like the message "conversion not supported by target" is caused by targetm.vectorize.builtin_conversion not ever being set. Unfortunately, it seems like that target hook was deprecated in all existing ports (i.e. rs6000 and i386), sometime in 2011.
What are the new requirements for supporting vectorization?
Thanks and regards,
-W Y