4 ms·
Overview of cross-architecture portability problems
- hanikesn 2y agoLooks like non 4k page sizes are missing, which tripped out some software running on asahi Linux.
- saagarjha 2y agoMost software doesn’t really care. This list only has common issues (though the issues it goes over seem more like problems from 2010…)
- denotational 2y agoMissed my favourite one: differences in the memory consistency model. If you’re using stdatomic and the like correctly, then the library and compiler writers have your back, but if you aren’t (e.g. using relaxed ordering when acquire/release is required), or if you’re rolling your own synchronisation primitives (don’t do this unless you know what you’re doing!), then you’re going to have a bad day.
- twoodfin 2y agoWhen Alpha was a vaguely viable platform, it was fantastically helpful to throw it into a build farm to find these kinds of issues.
- Archit3ch 2y agoThe fun of discovering size_t is defined differently on Apple Silicon.
- saagarjha 2y agoDifferently from what?
- johnklos 2y agoOne of the larger problems is purely social. Some people unnecessarily resist the idea that code can run on something other than x86 (and perhaps now Arm). Interestingly, some apologists say that maintaining portability in code is a hinderance that costs time and money, as though the profit of a company or the productivity of a programmer will be dramatically affected if they need to program without making assumptions about the underlying architecture. In reality, writing without those assumptions usually makes code better, with fewer edge cases. It'd be nice if people wouldn't go out of their way to fight against portability, but some people do :(
- rqtwteye 2y agoI am for portability but if you decide to go that route, you need to do a ton of testing on different platforms to make sure it really works. That definitely costs time and effort. There are a lot of ways to do it wrong.
- DanielHB 2y agoEven setting up your automated tests on multiple platforms can cost a ton of money on CI infra alone. Besides increased server costs some architectures cost a lot more to rent on cloud-CI systems (cough macOS) and setting up the local infra to run tests on-prem can be a huge endeavor and each new platform added has fixed costs. Nevermind your cloud CI randomly dropping support for one of your archs.
- gwbas1c 2y agoI argue the opposite: It's important to know the purpose of the program, including the computer that the program will run on. I can understand polite disagreements about target architectures: You might write a windows desktop program and disagree with someone if you should support 32-bit or be 64-bit only. BUT: Outside of x64 and ARM, your program (Windows Desktop) is highly unlikely to run on any other instruction set. The reality is that it's very time consuming and labor-intense to ship correctly-working software. Choosing and limiting your target architecture is an important part of that decision. Either you increase development cost/time (and risk not getting a return on your investment,) or you ship something buggy and risk your reputation / support costs.
- magicalhippo 2y agoOne I recall, working on a C++ program that we distributed to Windows, Linux and PowerPC OSX at the time, was how some platforms had memory zero-initialized by the OS memory manager, and some didn't. Our code didn't mean to take advantage of this, but it sometimes meant buggy code would appear fine on one platform as pointers in structures would be zeroed out, but crash on others where they weren't. As I recall, it was mostly that and the endianess that caused the most grief. Not that there was many issues at all since we used Qt and Boost...
- microtherion 2y agoOn many platforms this can depend on the size and timing of the allocation. As soon as a new PAGE has to be allocated (either because the allocation is large, or because the rest of the heap became full), it is zero initialized (pretty sure all OSes do this, to prevent cross-process spying), but if it's just a malloc inside an existing page, it does not clear previous cruft.
- AStonesThrow 2y agoIn my misspent college years, I was coding game servers in C. A new iteration came about, and the project lead had coded it on a VAX/BSD system, where I had no access. Under whatever C implementation on VAX/BSD, a NULL pointer dereference returned "0". Yep, you read that right. There was no segfault, no error condition, just a simple nul-for-null! This was all fine and dandy until he handed it over for me to port to SunOS, Ultrix, NeXT Mach BSD, etc. (Interesting times!) I honestly didn't find a second implementation or architecture where NULL-dereference was OK. Whatever the standards at the time, we were on the cusp of ANSI-compliance and everyone was adopting gcc. Segmentation faults were "handled" by immediately kicking off 1-255 players, and crashing the server. Not a good user experience! So my first debugging task, and it went on a long time, was to find every pointer dereference (and it was nul-terminated strings) and wrap them in a conditional "if (ptr != NULL) { ... }" At the time, I had considered it crazy/lazy to omit those in the first place and code on such an assumption. But C was so cozy with the hardware, and we were just kids. And there was the genesis for the cynical expression "All The World's A VAX!"
- Pinus 2y agoAt least some versions of HP-UX running on the HPPA architecture did the same thing. Writing to a NULL pointer would SIGSEGV, but reading got you a zero. There was a linker flag to turn off this "feature", but when I tried that, I started getting SIGSEGV:s all over various libraries I had no control over...
- AStonesThrow 2y agoIf I recall correctly, we were always careful to allocate memory and assign it to the pointer when creating a new string. But the test for whether a string-typed property existed was just to dereference that pointer. Interestingly, the maintainer had written this as an expansion to an existing codebase, which started on a NeXT and was already handling those NULL pointers properly, so it was a mismatch of coding style; a regression, I suppose. Our next challenge was the database being read entirely into RAM, and as the players expanded the virtual world, the Resident Set Size went out of control and ate all resources on the workstation. Then a fellow came along and wrote a "port concentrator" add-on, because at the time, process limits permitted only 64 open file descriptors, which capped the number of inbound TCP connections for active players. The port concentrator spawned 4 workers and muxed a total of 256 connections, for a pioneering MMORPG environment!
- benchloftbrunch 2y agoSomething not mentioned is that on Windows `long` is always 32 bits (same size as `int`), even on 64-bit architectures.
- malkia 2y agoIs it still the case that wasm is still 32-bit?
- melchizedek6809 2y agoMemory64 is supported by a lot of runtimes now, although it isn't fully standardized yet (see https://github.com/WebAssembly/proposals https://github.com/WebAssembly/proposals), so no idea how reliable the implementations actually are, since I haven't had a need for that much memory yet
- milabogdanova 2y ago[dead]