[llvm-dev] Proposal to remove MMX support.

[Eli Friedman via llvm-dev]

Broadly speaking, I see two problems with implicitly enabling MMX emulation on a target that has SSE2:


  1.  The interaction with inline asm.  Inline asm can still have MMX operands/results/clobbers, and can still put the processor in MMX mode.  If code is mixing MMX intrinsics and inline asm, there could be a significant penalty to moving values across register files. And it’s not clear what we want to do with _mm_empty(): under full emulation, it should be a no-op, but if there’s MMX asm, we need to actually clear the register file.
  2.  The calling convention problem; your description covers this.

If we add an explicit opt-in, I don’t see any problem with emulation (even on non-x86 targets, if someone implements that).

If we’re going make changes that could break people’s software, I’d prefer it to break with a compile-time error: require the user affirmatively state that there aren’t any interactions with assembly code.  For the MMX intrinsics in particular, code that’s using intrinsics is likely to also be using inline asm, so the interaction is important.


In terms of whether it’s okay to assume people don’t need support for MMX intrinsics on targets without SSE2, that’s probably fine.  Wikipedia says Intel stopped manufacturing PC chips without SSE2 around 2007.  And supported versions of Windows require SSE2.

-Eli

From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of James Y Knight via llvm-dev
Sent: Sunday, August 30, 2020 4:11 PM
To: llvm-dev <llvm-dev at lists.llvm.org>
Subject: [EXT] [llvm-dev] Proposal to remove MMX support.

I recently diagnosed a bug in someone else's software, which turned out to be due to incorrect MMX intrinsics usage: if you use any of the x86 intrinsics that accept or return __m64 values, then you, the programmer are required to call _mm_empty() before using any x87 floating point instructions or leaving the function. I was aware that this was required at the assembly-level, but not that the compiler forced users to deal with this when using intrinsics.

This is a real nasty footgun -- if you get this wrong, your program doesn't crash -- no, that would be too easy! Instead, every x87 instruction will simply result in a NaN value.

Even more unfortunately than all that, it is currently impossible to correctly use _mm_empty() to resolve the problem, because the compiler has no restrictions against placing x87 FPU operations between an MMX instruction and the EMMS instruction.

Of course, I didn't discover any of this -- it was already well-known...just not to me. But let's actually fix it.

Existing bugs:
llvm.org/PR35982<https://bugs.llvm.org/show_bug.cgi?id=35982> -- POSTRAScheduler disarrange emms and mmx instruction
llvm.org/PR41029<https://bugs.llvm.org/show_bug.cgi?id=41029> -- The __m64 not passed according to i386 ABI
llvm.org/PR42319<https://bugs.llvm.org/show_bug.cgi?id=42319> -- Add pass to insert EMMS/FEMMS instructions to separate MMX and X87 states
llvm.org/PR42320<https://bugs.llvm.org/show_bug.cgi?id=42320> -- Implement MMX intrinsics with SSE equivalents

Proposal
We should re-implement all the currently-MMX intrinsics in Clang's *mmintrin.h headers by using the existing SSE/SSE2 compiler builtins, on both x86-32 and x86-64, and then delete the MMX implementation of these intrinsics. We would thus stop supporting the use of these intrinsics, without SSE2 also enabled. I've created a preliminary patch for these header changes, https://reviews.llvm.org/D86855.

Sometime later, we should then remove the MMX intrinsics in LLVM IR. (Only the intrinsics -- the machine-instruction and register definitions for MMX should be kept indefinitely for use by assembly code.) That raises the question of bitcode compat. Maybe we do something to prevent new use of the intrinsics, but keep the implementations around for bitcode compatibility for a while longer?

We might also consider defaulting to -mno-mmx for new compilations in x86-64, which would have the additional effect of disabling the "y" constraint in inline-asm. (MMX instructions could still exist in the binary, but they'd need to be entirely contained within an inline-asm blob).

Unfortunately, given the ABI requirement in x86-32 to use MMX registers for 8-byte-vector arguments and returns -- which we've been violating for 7 years -- we probably cannot simply use -mno-mmx by default on x86-32. Unless, of course, we decide that we might as well just continue violating the ABI indefinitely. (Why bother to be correct, after the better part of a decade being incorrect...)

Impact
- No more %mm* register usage on x86-64, other than via inline-asm. No more %mm* register usage on x86-32, other than inline-asm and when calling a function that takes/returns 8-byte vectors (assuming we fix the ABI-compliance issue).
- Since the default compiler flags include SSE2, most code will switch to using SSE2 instructions instead of MMX instructions when using intrinsics, and continue to compile fine. It'll also likely be faster, since MMX instructions are legacy, and not optimized in CPUs anymore.
- Code explicitly disabling SSE2 (e.g. -mno-sse2 or -march=penium2) will stop compiling if it requires MMX intrinsics.
- Code using the intrinsics will run faster, especially on x86-64, where the vectors are passed around in xmm registers, and is being copied to mm registers just to run a legacy instruction. But even without that, the mmx instructions also just have less throughput than the sse2 variants on modern CPUs.

Alternatives
We could keep both implementations of the functions in mmintrin.h, in order to preserve the ability to use the intrinsics when compiling for a CPU without SSE2.

However, this doesn't seem worthwhile to me -- we're talking about dropping the ability to generate vectorized code using compiler intrinsics for Intel Pentium MMX, Pentium, II, and Pentium III (released 1997-1999), as well as AMD K6 and K7 series chips of around the same timeframe.

We could also keep the clang headers mostly-unmodified, and make the llvm IR builtins themselves expand to SSE2 instructions. I believe GCC has effectively chosen this option. That seems less desirable; it'll be more complex to implement at that level, versus in the headers, and doesn't leave a path towards eliminating the builtins in the future.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200831/8032d1e7/attachment.html>