On Sat, Oct 21, 2023 at 7:08 PM Andres Freund <andres@anarazel.de> wrote:
> I've attached a patch revision that I spent the last couple hours working
> on. It's very very roughly based on a patch Tom Stellard had written (which I
> think a few rpm packages use). But instead of encoding details about specific
> layout details, I made the code check if the data layout works and fall back
> to the cpu / features used for llvmjit_types.bc. This way it's not s390x
> specific, future odd architecture behaviour would "automatically" be handled
> the same
The explanation makes sense and this seems like a solid plan to deal
with it. I didn't try on a s390x, but I tested locally on our master
branch with LLVM 7, 10, 17, 18, and then I hacked your patch to take
the fallback path as if a layout mismatch had been detected, and it
worked fine:
2023-10-22 11:49:55.663 NZDT [12000] DEBUG: detected CPU "skylake",
with features "...", resulting in layout
"e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
2023-10-22 11:49:55.664 NZDT [12000] DEBUG: detected CPU / features
yield incompatible data layout, using values from module instead
2023-10-22 11:49:55.664 NZDT [12000] DETAIL: module CPU "x86-64",
features "...", resulting in layout
"e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+ To deal with that, check if data layouts match during JIT
initialization. If
+ the runtime detected cpu / features result in a different layout,
try if the
+ cpu/features recorded in in llvmjit_types.bc work.
s|try |check |
s| in in | in |
+ errmsg_internal("could not
determine working CPU / feature comination for JIT compilation"),
s|comination|combination|
s| / |/|g