[llvm] r314568 - [AMDGPU] Set fast-math flags on functions given the options

[Matt Arsenault via llvm-commits]

> On Sep 29, 2017, at 16:40, Stanislav Mekhanoshin via llvm-commits <llvm-commits at lists.llvm.org> wrote:
> 
> Author: rampitec
> Date: Fri Sep 29 16:40:19 2017
> New Revision: 314568
> 
> URL: http://llvm.org/viewvc/llvm-project?rev=314568&view=rev
> Log:
> [AMDGPU] Set fast-math flags on functions given the options
> 
> We have a single library build without relaxation options.
> When inlined library functions remove fast math attributes
> from the functions they are integrated into.
> 
> This patch sets relaxation attributes on the functions after
> linking provided corresponding relaxation options are given.
> Math instructions inside the inlined functions remain to have
> no fast flags, but inlining does not prevent fast math
> transformations of a surrounding caller code anymore.
> 
> Differential Revision: https://reviews.llvm.org/D38325
> 
> Added:
>    llvm/trunk/test/CodeGen/AMDGPU/inline-attr.ll
> Modified:
>    llvm/trunk/lib/Target/AMDGPU/AMDGPU.h
>    llvm/trunk/lib/Target/AMDGPU/AMDGPULibCalls.cpp
>    llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
> 
> Modified: llvm/trunk/lib/Target/AMDGPU/AMDGPU.h
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/AMDGPU.h?rev=314568&r1=314567&r2=314568&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Target/AMDGPU/AMDGPU.h (original)
> +++ llvm/trunk/lib/Target/AMDGPU/AMDGPU.h Fri Sep 29 16:40:19 2017
> @@ -23,6 +23,7 @@ class ModulePass;
> class Pass;
> class Target;
> class TargetMachine;
> +class TargetOptions;
> class PassRegistry;
> class Module;
> 
> @@ -52,7 +53,7 @@ FunctionPass *createSIDebuggerInsertNops
> FunctionPass *createSIInsertWaitsPass();
> FunctionPass *createSIInsertWaitcntsPass();
> FunctionPass *createSIFixWWMLivenessPass();
> -FunctionPass *createAMDGPUSimplifyLibCallsPass();
> +FunctionPass *createAMDGPUSimplifyLibCallsPass(const TargetOptions &);
> FunctionPass *createAMDGPUUseNativeCallsPass();
> FunctionPass *createAMDGPUCodeGenPreparePass();
> FunctionPass *createAMDGPUMachineCFGStructurizerPass();
> 
> Modified: llvm/trunk/lib/Target/AMDGPU/AMDGPULibCalls.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/AMDGPULibCalls.cpp?rev=314568&r1=314567&r2=314568&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Target/AMDGPU/AMDGPULibCalls.cpp (original)
> +++ llvm/trunk/lib/Target/AMDGPU/AMDGPULibCalls.cpp Fri Sep 29 16:40:19 2017
> @@ -30,6 +30,7 @@
> #include "llvm/IR/ValueSymbolTable.h"
> #include "llvm/Support/Debug.h"
> #include "llvm/Support/raw_ostream.h"
> +#include "llvm/Target/TargetOptions.h"
> #include <vector>
> #include <cmath>
> 
> @@ -168,10 +169,13 @@ namespace {
> 
>   AMDGPULibCalls Simplifier;
> 
> +  const TargetOptions Options;
> +
>   public:
>     static char ID; // Pass identification
> 
> -    AMDGPUSimplifyLibCalls() : FunctionPass(ID) {
> +    AMDGPUSimplifyLibCalls(const TargetOptions &Opt = TargetOptions())
> +      : FunctionPass(ID), Options(Opt) {
>       initializeAMDGPUSimplifyLibCallsPass(*PassRegistry::getPassRegistry());
>     }
> 
> @@ -1680,14 +1684,34 @@ bool AMDGPULibCalls::evaluateCall(CallIn
> }
> 
> // Public interface to the Simplify LibCalls pass.
> -FunctionPass *llvm::createAMDGPUSimplifyLibCallsPass() {
> -  return new AMDGPUSimplifyLibCalls();
> +FunctionPass *llvm::createAMDGPUSimplifyLibCallsPass(const TargetOptions &Opt) {
> +  return new AMDGPUSimplifyLibCalls(Opt);
> }
> 
> FunctionPass *llvm::createAMDGPUUseNativeCallsPass() {
>   return new AMDGPUUseNativeCalls();
> }
> 
> +static bool setFastFlags(Function &F, const TargetOptions &Options) {
> +  AttrBuilder B;
> +
> +  if (Options.UnsafeFPMath || Options.NoInfsFPMath)
> +    B.addAttribute("no-infs-fp-math", "true");
> +  if (Options.UnsafeFPMath || Options.NoNaNsFPMath)
> +    B.addAttribute("no-nans-fp-math", "true");
> +  if (Options.UnsafeFPMath) {
> +    B.addAttribute("less-precise-fpmad", "true”);
less-precise-fpmad was removed a while ago (although maybe there is a use for it)


> +    B.addAttribute("unsafe-fp-math", "true");
> +  }
> +
> +  if (!B.hasAttributes())
> +    return false;
> +
> +  F.addAttributes(AttributeList::FunctionIndex, B);
> +
> +  return true;
> +}
> +
> bool AMDGPUSimplifyLibCalls::runOnFunction(Function &F) {
>   if (skipFunction(F))
>     return false;
> @@ -1699,6 +1723,9 @@ bool AMDGPUSimplifyLibCalls::runOnFuncti
>         F.printAsOperand(dbgs(), false, F.getParent());
>         dbgs() << '\n';);
> 
> +  if (!EnablePreLink)
> +    Changed |= setFastFlags(F, Options);
> +
>   for (auto &BB : F) {
>     for (BasicBlock::iterator I = BB.begin(), E = BB.end(); I != E; ) {
>       // Ignore non-calls.
> 
> Modified: llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp?rev=314568&r1=314567&r2=314568&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp (original)
> +++ llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp Fri Sep 29 16:40:19 2017
> @@ -370,17 +370,18 @@ void AMDGPUTargetMachine::adjustPassMana
>         PM.add(createAMDGPUAlwaysInlinePass(false));
>   });
> 
> +  const auto &Opt = Options;

Since TargetOptions is an ugly, stateful hack that is mutated on each function based on the attributes I don’t think I trust this. This is essentially setting the attributes based on the presence of the very attributes you are setting. 

I also think there needs to be a way to restrict this to specifically handle the known builtin functions. I think doing this correctly requires a Module or SCC pass that knows about the special builtins which don’t have the specified attributes.

-Matt