Recently HIP toolchain made a change to use clang instead of opt/llc to do compilation
(https://reviews.llvm.org/D81861). The intention is to make HIP toolchain canonical like
other toolchains.
However, this change introduced an unintentional change regarding backend fp fuse
option, which caused regressions in some HIP applications.
Basically before the change, HIP toolchain used clang to generate bitcode, then use
opt/llc to optimize bitcode and generate ISA. As such, the amdgpu backend takes
the default fp fuse mode which is 'Standard'. This mode respect contract flag of
fmul/fadd instructions and do not fuse fmul/fadd instructions without contract flag.
However, after the change, HIP toolchain now use clang to generate IR, do optimization,
and generate ISA as one process. Now amdgpu backend fp fuse option is determined
by -ffp-contract option, which is 'fast' by default. And this -ffp-contract=fast language option
is translated to 'Fast' fp fuse option in backend. Suddenly backend starts to fuse fmul/fadd
instructions without contract flag.
This causes wrong result for some device library functions, e.g. tan(-1e20), which should
return 0.8446, now returns -0.933. What is worse is that since backend with 'Fast' fp fuse
option does not respect contract flag, there is no way to use #pragma clang fp contract
directive to enforce fp contract requirements.
This patch fixes the regression by introducing a new value 'fast-honor-pragmas' for -ffp-contract
and use it for HIP by default. 'fast-honor-pragmas' is equivalent to 'fast' in frontend but
let the backend to use 'Standard' fp fuse option. 'fast-honor-pragmas' is useful since 'Fast'
fp fuse option in backend does not honor contract flag, it is of little use to HIP
applications since all code with #pragma STDC FP_CONTRACT or any IR from a
source compiled with -ffp-contract=on is broken.
Differential Revision: https://reviews.llvm.org/D90174
FPContractModeKind is the codegen option flag which is already ternary (off,
on, fast). This makes it universally the type for the contractable info
across the front-end:
* In FPOptions (i.e. in the Sema + in the expression nodes).
* In LangOpts::DefaultFPContractMode which is the option that initializes
FPOptions in the Sema.
Another way to look at this change is that before fp-contractable on/off were
the only states handled to the front-end:
* For "on", FMA folding was performed by the front-end
* For "fast", we simply forwarded the flag to TargetOptions to handle it in
LLVM
Now off/on/fast are all exposed because for fast we will generate
fast-math-flags during CodeGen.
This is toward moving fp-contraction=fast from an LLVM TargetOption to a
FastMathFlag in order to fix PR25721.
---
This is a recommit of r299027 with an adjustment to the test
CodeGenCUDA/fp-contract.cu. The test assumed that even
though -ffp-contract=on is passed FE-based folding of FMA won't happen.
This is obviously wrong since the user is asking for this explicitly with the
option. CUDA is different that -ffp-contract=fast is on by default.
The test used to "work" because contract=fast and contract=on were maintained
separately and we didn't fold in the FE because contract=fast was on due to
the target-default. This patch consolidates the contract=on/fast/off state
into a ternary state hence the change in behavior.
---
Differential Revision: https://reviews.llvm.org/D31167
llvm-svn: 299033
This matches default nvcc behavior and gives substantial
performance boost on GPU where fmad is much cheaper compared to add+mul.
Differential Revision: http://reviews.llvm.org/D20341
llvm-svn: 270094
Yaxun (Sam) Liu