[PATCH] D116583: Change the default optimisation level of PTXAS from -O0 to -O3. This makes the optimisation levels of PTXAS and the ptxjitcompiler equal (ptxjitcompiler defaults to -O3).

[Artem Belevich via Phabricator via cfe-commits]

tra added inline comments.


================
Comment at: clang/lib/Driver/ToolChains/Cuda.cpp:433
   } else {
-    // If no -O was passed, pass -O0 to ptxas -- no opt flag should correspond
-    // to no optimizations, but ptxas's default is -O3.
-    CmdArgs.push_back("-O0");
+    // If no -O was passed, pass -O3 to ptxas -- this makes ptxas's
+    // optimization level the same as the ptxjitcompiler.
----------------
hdelan wrote:
> tra wrote:
> > I think this would be contrary to the expectation that lack of `-O` in clang means - `do not optimize` and it generally implies the whole compilation chain, including assembler. Matching whatever nvidia tools do is an insufficient reason for breaking this assumption, IMO. 
> > 
> > If you do want do run optimized ptxas on unoptimized PTX, you can use `-Xcuda-ptxas -O3`.
> I think for the average user, consistency across the `ptxjitcompiler` and `ptxas` is far more important than assuming that no `-O` means no optimization. I think most users will assume that no `-O` will assume that whatever tools being used will take their default optimization level, which in the case of clang is `-O0` and in the case of `ptxas` is `-O3`. 
> 
> We have had a few bugs with `ptxas`/`ptxjitcompiler` at higher optimization levels, which were quite hard to pin down since offline `ptxas` and `ptxjitcompiler` were using different optimisation levels, making bugs appear in one and not the other. Of course we are aware of this now but this inconsistency can result in bugs that are difficult to diagnose. Having consistency between the `ptxjitcompiler` and `ptxas` is therefore of practical benefit. Whereas if we are to leave it as is, with `ptxas` defaulting to `-O0`, the benefit is purely semantic and not practical.
> I think for the average user, consistency across the ptxjitcompiler and ptxas is far more important than assuming that no -O means no optimization. 

The default is intended to provide the least amount of surprises for the most users. There are more users of clang as a CUDA compiler than users of clang as a cuda compiler who care about consistency with ptxjitcompiler. My point is that the improvements for a subset of users should be balanced vs usability in the common case. In this case the benefit does not justify the downsides, IMO.

Please add me as a reviewer when the patch is ready for public review and we'll discuss it in a wider LLVM community.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D116583/new/

https://reviews.llvm.org/D116583