Default to turning on architecture specific optimizations in the codegen

LLVM defaults to turning on whatever architecture specific optimizations are supported on the user's machine (e.g. -msse4.2). This makes programs run faster during testing and generally gives users a better perception of the compiler. If the user wants to compile for some more portable subset (e.g. when cross-compiling) he/she adds -march, -mcpu, etc flags on the command line.

To implement this we'd have to detect what CPU type is used etc and initialize our DynFlags based on that.

Trac metadata

Trac field	Value
Version	7.6.3
Type	Bug
TypeOfFailure	OtherFailure
Priority	normal
Resolution	Unresolved
Component	Compiler
Test case
Differential revisions
BlockedBy
Related
Blocking
CC
Operating system
Architecture

changed weight to 5

added Tbug Trac import labels

LLVM documentation about their flags and autodetection behavior here: http://llvm.org/docs/CommandGuide/llc.html

mentioned in issue #9777 (closed)

Trac metadata

Trac field	Value
Type	Bug → FeatureRequest
TypeOfFailure	OtherFailure → RuntimePerformance
Component	Compiler → Compiler (NCG)

added Pnormal label

I stumbled upon this while trying to document some of the -msse flags (#18641).

First off, I've seen this commit, which make SSE2 the minimum version and enables it by default on x86 and x86_64 - and that's great!

While looking at the flags, I see that the -msse flag is now useless (which is fine), but that -msse3 and -msse4 don't seem to do anything either, bc. I can only find checks for SSE2 and SSE4.2. That's not bad per se, if it is documented. (Note that the AVX flags also do nothing, but we're really not using them)

But we check for SSE4.2 to use the POPCNT instruction (see also comment in Instr.hs). While it was introduced with SSE4, this instruction has it's own feature bit. I don't think there is a CPU which supports SSE4.1/SSE4a and doesn't support POPCNT. SSE4.2 only introduced 7 additional instructions, mostly for string processing and we're not using a single one (at least they are not in Instr.hs). So maybe we could relax that requirement from SSE4.2 to SSE4.

Using more advanced instructions by default sounds good to me, but detecting the users CPU is not enough. We might be dealing with a cross-compiler. So we need to look at the targetPlatform. Since SSE4.2 was first shipped more than 10 years ago, I think it wouldn't be a bad default for x86_64. We'd have to document this default and let the user override it with -msseX if their target platform is, e.g., a Pentium 4.

EDIT: Now I found an article saying that "Intel considers POPCNT to be part of SSE4.2". While the original statement still stands, namely that POPCNT has its own feature bit and that I don't know any CPU implementing SSE4.1/4a without POPCNT. Nevertheless, we can leave it in -msse4. Furthermore NCG uses BMI2 instructions, which came out the same time as SSE4, so we should activate them by default too.

mentioned in issue #21560

mentioned in issue #24073 (closed)

marked #24073 (closed) as a duplicate of this issue

marked this issue as related to #24073 (closed)

mentioned in issue #25450