|
From: Karthik S. <ks...@dr...> - 2015-10-27 11:03:21
|
Hi all, While running an ARMv8 benchmark, I found that an instruction Valgrind appears to choke on. For reference, the error message follows: "The instruction is legitimate but Valgrind doesn't handle it, i.e. it's > Valgrind's fault. If you think this is the case or you are not sure, > please let us know and we'll try to fix it." I tracked down the problem instruction: a vectorized form of the double precision 'fmaxnm' instruction to compare two double precision numbers. The scalar version of this instruction runs fine, and I can verify that the vectorized instruction only gets generated by gcc when using the tree-vectorize flag (and 03). 405720: 4e60c508 fmaxnm v8.2d, v8.2d, v0.2d Does anyone know if: 1) Vectorized double-precision instructions are unimplemented in valgrind, 2) This particular instruction has been unimplemented, or 3) The platform I'm using is doing something sticky with these instructions? Thanks, Karthik |
|
From: Julian S. <js...@ac...> - 2016-01-18 14:35:36
|
On 27/10/15 12:02, Karthik Sangaiah wrote: > 405720: 4e60c508 fmaxnm v8.2d, v8.2d, v0.2d I'm pretty sure that's implemented. Have you tried with 3.11.0? I suspect it fails because you are using an earlier version. J |
|
From: Karthik S. <ks...@dr...> - 2016-01-18 18:24:24
|
Hey Julian, Looking back, I believe our toolset was actually using Valgrind 3.10.1. I'll give Valgrind 3.11.0 a shot. Thanks, Karthik On Mon, Jan 18, 2016 at 9:35 AM, Julian Seward <js...@ac...> wrote: > On 27/10/15 12:02, Karthik Sangaiah wrote: > > > 405720: 4e60c508 fmaxnm v8.2d, v8.2d, v0.2d > > I'm pretty sure that's implemented. Have you tried with 3.11.0? > I suspect it fails because you are using an earlier version. > > J > > |