Hi everyone,

I have to apologise for my recent absence and lack of progress fixing the bugs I've caused.  I'm currently fixing the failure in webtbs/tw41865c.pp which is centred around the division of two Currency types.  I've got it to work on x86_64-win64 (which internally stores it as an Int64) but my fix doesn't work on x86_64-linux (which internally stores it as an Extended). However, while writing this e-mail, the fact I said "Extended" just clued me in... the test itself may be incorrect because the storage formats between integral and floating-point values are fundamentally different.  Back to work I go!

Gareth aka. Kit

Hi Gareth,

tw41865c comes from my reproducers for #41865, so here are some measurements that may help.

On x86_64-linux (cross compiler from current main 7a7ff6e731, run under qemu-x86_64) the storage is not the issue: Currency is 8 bytes and holds the scaled Int64 - Currency(10) is raw 100000, as on AArch64. What fails is the compilation:

  tw41865c.pp(12,10) Error: Operator is not overloaded: "Currency" div "Currency"

There Currency is a floating type (s64currency, computed on the x87), and div/mod are not defined for floating types. Would guarding the test with {$ifdef FPC_CURRENCY_IS_INT64} match what you intend, or should div/mod work there too?

On aarch64-win64, !1560 already makes tw41865c pass, and !1576 fixes your tw41865g/h at -O4:
https://gitlab.com/freepascal.org/fpc/source/-/merge_requests/1560 https://gitlab.com/freepascal.org/fpc/source/-/merge_requests/1576 Could you check whether they overlap with your changes?

Sven

Ah yes, that makes sense.  My fix-in-progress for div and mod is very similar to 1560 but I use a distinct method to create the Currency scalar that handles the flags etc (although it's otherwise the same).  I like the look of your i and j tests though.

For 1576 I want to avoid using a recursive node search if possible because the "nf_is_currency" flag should be propagated.

Gareth aka. Kit