3 ms·

If the compiler takes a full machine description into account for each of the target architectures (AFAIK, LLVM does that. ), it won't be based around a "lowest

If the compiler takes a full machine description into account for each of the target architectures (AFAIK, LLVM does that. ), it won't be based around a "lowest common denominator". The machine description is a list of instructions and what they do and things like timing and scheduling constraints. The task of the compiler is to find the most optimized instruction sequence that do a certain thing for the architecture it is compiling for. This principle is the same for every architecture, so you don't need a specific optimizer per platform.

_delirium 15y ago
That's the theory, but you do run into problems if the entire parameterization misses some architectural differences that didn't exist when it was first developed (or weren't important). That's one problem gcc has, that its architecture description is aged and doesn't have a good way of specifying some things that are very important for 2010s performance, so has had to undergo some major surgery (still in progress). LLVM currently avoids that partly by being newer. It might also be better designed; we'll see how it ages over the next 20 years. =]
- wladimir 15y ago
  As LLVM is used for everything from x86 to GPU code generation, I have some trust in it being able to hold up. But indeed, only time will tell.
  - cliffbean 15y ago
    LLVM itself knows almost nothing about GPU-specific architectural features. Emitting PTX from LLVM, for example, is mostly just a translation from one compiler IR to another compiler IR.
- froydnj 15y ago
  > That's one problem gcc has, that its architecture description is aged and doesn't have a good way of specifying some things that are very important for 2010s performance, so has had to undergo some major surgery (still in progress). LLVM currently avoids that partly by being newer. LLVM also currently avoids that by not generating code that's as good for those 2010 architectures. I would also be interested in hearing you expound upon those things that are "very important for 2010s performance".
  - _delirium 15y ago
    > I would also be interested in hearing you expound upon those things that are "very important for 2010s performance". I'm not particularly an expert in this area, but my understanding from articles/mailing-lists/etc. was that there were inadequacies in the previous machine-description and IR frameworks that led to the revamping with Gimple and MD-RTL over the past few years. Some of that was just maintainability, but I thought there were also optimization-related issues with the old internals. Is that not the case?
    - froydnj 15y ago
      It's hard to say without reference to specifics; certainly the MD files have been around since time-immemorial, even if they've been augmented with new features over the years. And the gimple bits were primarily aimed at making the in-memory representation more efficient, a goal which they accomplished. In any event, as you say, things have been revamped, so whatever issues there were with the old internals no longer apply, right? So there's no point in dragging out old chestnuts like "GCC can't support modern architectures or modern optimization techniques", because that's not true anymore, right? (I apologize if this comes across as harsh; I'm just tired of seeing LLVM articles where commenters appear to be drinking the LLVM Kool-Aid without having any idea what the LLVM folks are grousing about. Sometimes the LLVM folks have a point, sometimes they're just asserting their engineering decisions are superior, which is debatable, and sometimes they're just grousing because they don't seem to like GCC. It's hard to say exactly what's in view from commenters and from the LLVM folks themselves. GCC currently supports 8-bit microcontrollers, 32/64-bit desktop chips, a few "nonstandard" VLIW and DSP architectures, and lots of other chips in between. I, for one, am impressed with how much Clang and LLVM have done, but I'll also be more impressed if, in a decade and a half, LLVM's architecture hasn't acquired some warts and it seriously supports more than two architectures. After all, GCC was, in many ways, state-of-the-art when it first came out too...)
      - _delirium 15y ago
        Ah yeah, that's fair. I'd also be impressed if LLVM doesn't have warts in a decade or two, which is sort of what I was trying to get at with the end of my comment--- that to the extent that LLVM has a cleaner architecture representation (if it does), it's mostly a function if it being a newer "clean 1.0", and not having yet had to adapt as architecture and optimization techniques change over time.

_delirium 15y ago

That's the theory, but you do run into problems if the entire parameterization misses some architectural differences that didn't exist when it was first developed (or weren't important). That's one problem gcc has, that its architecture description is aged and doesn't have a good way of specifying some things that are very important for 2010s performance, so has had to undergo some major surgery (still in progress). LLVM currently avoids that partly by being newer. It might also be better designed; we'll see how it ages over the next 20 years. =]

wladimir 15y ago
As LLVM is used for everything from x86 to GPU code generation, I have some trust in it being able to hold up. But indeed, only time will tell.
- cliffbean 15y ago
  LLVM itself knows almost nothing about GPU-specific architectural features. Emitting PTX from LLVM, for example, is mostly just a translation from one compiler IR to another compiler IR.
froydnj 15y ago
> That's one problem gcc has, that its architecture description is aged and doesn't have a good way of specifying some things that are very important for 2010s performance, so has had to undergo some major surgery (still in progress). LLVM currently avoids that partly by being newer. LLVM also currently avoids that by not generating code that's as good for those 2010 architectures. I would also be interested in hearing you expound upon those things that are "very important for 2010s performance".
- _delirium 15y ago
  > I would also be interested in hearing you expound upon those things that are "very important for 2010s performance". I'm not particularly an expert in this area, but my understanding from articles/mailing-lists/etc. was that there were inadequacies in the previous machine-description and IR frameworks that led to the revamping with Gimple and MD-RTL over the past few years. Some of that was just maintainability, but I thought there were also optimization-related issues with the old internals. Is that not the case?
  - froydnj 15y ago
    It's hard to say without reference to specifics; certainly the MD files have been around since time-immemorial, even if they've been augmented with new features over the years. And the gimple bits were primarily aimed at making the in-memory representation more efficient, a goal which they accomplished. In any event, as you say, things have been revamped, so whatever issues there were with the old internals no longer apply, right? So there's no point in dragging out old chestnuts like "GCC can't support modern architectures or modern optimization techniques", because that's not true anymore, right? (I apologize if this comes across as harsh; I'm just tired of seeing LLVM articles where commenters appear to be drinking the LLVM Kool-Aid without having any idea what the LLVM folks are grousing about. Sometimes the LLVM folks have a point, sometimes they're just asserting their engineering decisions are superior, which is debatable, and sometimes they're just grousing because they don't seem to like GCC. It's hard to say exactly what's in view from commenters and from the LLVM folks themselves. GCC currently supports 8-bit microcontrollers, 32/64-bit desktop chips, a few "nonstandard" VLIW and DSP architectures, and lots of other chips in between. I, for one, am impressed with how much Clang and LLVM have done, but I'll also be more impressed if, in a decade and a half, LLVM's architecture hasn't acquired some warts and it seriously supports more than two architectures. After all, GCC was, in many ways, state-of-the-art when it first came out too...)
    - _delirium 15y ago
      Ah yeah, that's fair. I'd also be impressed if LLVM doesn't have warts in a decade or two, which is sort of what I was trying to get at with the end of my comment--- that to the extent that LLVM has a cleaner architecture representation (if it does), it's mostly a function if it being a newer "clean 1.0", and not having yet had to adapt as architecture and optimization techniques change over time.