[HIP] Fix regressions due to fp contract change

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

Author: yxsamliu

clang/docs/LanguageExtensions.rst

@@ -3209,8 +3209,9 @@ statements in C).
 
 The pragma can also be used with ``off`` which turns FP contraction off for a
 section of the code. This can be useful when fast contraction is otherwise
-enabled for the translation unit with the ``-ffp-contract=fast`` flag.
-
+enabled for the translation unit with the ``-ffp-contract=fast-honor-pragmas`` flag.
+Note that ``-ffp-contract=fast`` will override pragmas to fuse multiply and
+addition across statements regardless of any controlling pragmas.
 
 ``#pragma clang fp exceptions`` specifies floating point exception behavior. It
 may take one the the values: ``ignore``, ``maytrap`` or ``strict``. Meaning of

clang/docs/UsersManual.rst

@@ -1336,15 +1336,16 @@ are listed below.
    The C standard permits intermediate floating-point results within an
    expression to be computed with more precision than their type would
    normally allow. This permits operation fusing, and Clang takes advantage
-   of this by default. This behavior can be controlled with the
-   ``FP_CONTRACT`` pragma. Please refer to the pragma documentation for a
-   description of how the pragma interacts with this option.
+   of this by default. This behavior can be controlled with the ``FP_CONTRACT``
+   and ``clang fp contract`` pragmas. Please refer to the pragma documentation
+   for a description of how the pragmas interact with this option.
 
    Valid values are:
 
-   * ``fast`` (everywhere)
-   * ``on`` (according to FP_CONTRACT pragma, default)
+   * ``fast`` (fuse across statements disregarding pragmas, default for CUDA)
+   * ``on`` (fuse in the same statement unless dictated by pragmas, default for languages other than CUDA/HIP)
    * ``off`` (never fuse)
+   * ``fast-honor-pragmas`` (fuse across statements unless dictated by pragmas, default for HIP)
 
 .. _opt_fhonor-infinities:
 

clang/include/clang/Basic/LangOptions.h

@@ -187,8 +187,11 @@ class LangOptions : public LangOptionsBase {
     // Enable the floating point pragma
     FPM_On,
 
-    // Aggressively fuse FP ops (E.g. FMA).
-    FPM_Fast
+    // Aggressively fuse FP ops (E.g. FMA) disregarding pragmas.
+    FPM_Fast,
+
+    // Aggressively fuse FP ops and honor pragmas.
+    FPM_FastHonorPragmas
   };
 
   /// Alias for RoundingMode::NearestTiesToEven.
@@ -417,7 +420,13 @@ class FPOptions {
   }
   explicit FPOptions(const LangOptions &LO) {
     Value = 0;
-    setFPContractMode(LO.getDefaultFPContractMode());
+    // The language fp contract option FPM_FastHonorPragmas has the same effect
+    // as FPM_Fast in frontend. For simplicity, use FPM_Fast uniformly in
+    // frontend.
+    auto LangOptContractMode = LO.getDefaultFPContractMode();
+    if (LangOptContractMode == LangOptions::FPM_FastHonorPragmas)
+      LangOptContractMode = LangOptions::FPM_Fast;
+    setFPContractMode(LangOptContractMode);
     setRoundingMode(LO.getFPRoundingMode());
     setFPExceptionMode(LO.getFPExceptionMode());
     setAllowFPReassociate(LO.AllowFPReassoc);

clang/include/clang/Driver/Options.td

@@ -1276,9 +1276,13 @@ def fno_rounding_math : Flag<["-"], "fno-rounding-math">, Group<f_Group>, Flags<
 def ftrapping_math : Flag<["-"], "ftrapping-math">, Group<f_Group>, Flags<[CC1Option]>;
 def fno_trapping_math : Flag<["-"], "fno-trapping-math">, Group<f_Group>, Flags<[CC1Option]>;
 def ffp_contract : Joined<["-"], "ffp-contract=">, Group<f_Group>,
-  Flags<[CC1Option]>, HelpText<"Form fused FP ops (e.g. FMAs): fast (everywhere)"
-  " | on (according to FP_CONTRACT pragma) | off (never fuse). Default"
-  " is 'fast' for CUDA/HIP and 'on' otherwise.">, Values<"fast,on,off">;
+  Flags<[CC1Option]>, HelpText<"Form fused FP ops (e.g. FMAs):"
+  " fast (fuses across statements disregarding pragmas)"
+  " | on (only fuses in the same statement unless dictated by pragmas)"
+  " | off (never fuses)"
+  " | fast-honor-pragmas (fuses across statements unless diectated by pragmas)."
+  " Default is 'fast' for CUDA, 'fast-honor-pragmas' for HIP, and 'on' otherwise.">,
+  Values<"fast,on,off,fast-honor-pragmas">;
 
 defm strict_float_cast_overflow : OptOutFFlag<"strict-float-cast-overflow",
   "Assume that overflowing float-to-int casts are undefined (default)",

clang/lib/CodeGen/BackendUtil.cpp

@@ -481,6 +481,7 @@ static bool initTargetOptions(DiagnosticsEngine &Diags,
     Options.AllowFPOpFusion = llvm::FPOpFusion::Standard;
     break;
   case LangOptions::FPM_On:
+  case LangOptions::FPM_FastHonorPragmas:
     Options.AllowFPOpFusion = llvm::FPOpFusion::Standard;
     break;
   case LangOptions::FPM_Fast:

clang/lib/Frontend/CompilerInvocation.cpp

@@ -2424,9 +2424,20 @@ void CompilerInvocation::setLangDefaults(LangOptions &Opts, InputKind IK,
 
   Opts.HIP = IK.getLanguage() == Language::HIP;
   Opts.CUDA = IK.getLanguage() == Language::CUDA || Opts.HIP;
-  if (Opts.CUDA)
-    // Set default FP_CONTRACT to FAST.
+  if (Opts.HIP) {
+    // HIP toolchain does not support 'Fast' FPOpFusion in backends since it
+    // fuses multiplication/addition instructions without contract flag from
+    // device library functions in LLVM bitcode, which causes accuracy loss in
+    // certain math functions, e.g. tan(-1e20) becomes -0.933 instead of 0.8446.
+    // For device library functions in bitcode to work, 'Strict' or 'Standard'
+    // FPOpFusion options in backends is needed. Therefore 'fast-honor-pragmas'
+    // FP contract option is used to allow fuse across statements in frontend
+    // whereas respecting contract flag in backend.
+    Opts.setDefaultFPContractMode(LangOptions::FPM_FastHonorPragmas);
+  } else if (Opts.CUDA) {
+    // Allow fuse across statements disregarding pragmas.
     Opts.setDefaultFPContractMode(LangOptions::FPM_Fast);
+  }
 
   Opts.RenderScript = IK.getLanguage() == Language::RenderScript;
   if (Opts.RenderScript) {
@@ -3343,6 +3354,8 @@ static void ParseLangArgs(LangOptions &Opts, ArgList &Args, InputKind IK,
       Opts.setDefaultFPContractMode(LangOptions::FPM_On);
     else if (Val == "off")
       Opts.setDefaultFPContractMode(LangOptions::FPM_Off);
+    else if (Val == "fast-honor-pragmas")
+      Opts.setDefaultFPContractMode(LangOptions::FPM_FastHonorPragmas);
     else
       Diags.Report(diag::err_drv_invalid_value) << A->getAsString(Args) << Val;
   }

clang/lib/Sema/SemaAttr.cpp

@@ -966,6 +966,8 @@ void Sema::ActOnPragmaFPContract(SourceLocation Loc,
   case LangOptions::FPM_Off:
     NewFPFeatures.setDisallowFPContract();
     break;
+  case LangOptions::FPM_FastHonorPragmas:
+    llvm_unreachable("Should not happen");
   }
   FpPragmaStack.Act(Loc, Sema::PSK_Set, StringRef(), NewFPFeatures);
   CurFPFeatures = NewFPFeatures.applyOverrides(getLangOpts());