Add fast-math.

sletz · Sep 9, 2017 · 1aa96a2 · 1aa96a2
1 parent cf60110
commit 1aa96a2
Show file tree

Hide file tree

Showing 2 changed files with 21 additions and 2 deletions.
diff --git a/Source/Runtime/LLVMEmitIR.cpp b/Source/Runtime/LLVMEmitIR.cpp
@@ -122,7 +122,12 @@ namespace LLVMJIT
 		, functionInstance(inFunctionInstance)
 		, llvmFunction(inLLVMFunction)
 		, irBuilder(context)
-        {}
+        {
+            // SL : fast-math a IR level
+            llvm::FastMathFlags FMF;
+            FMF.setUnsafeAlgebra();
+            irBuilder.setFastMathFlags(FMF);
+        }
 
 		void emit();
 
@@ -776,6 +781,8 @@ namespace LLVMJIT
 		//
 		// Load/store operators
 		//
+
+        // SL : removed setVolatile
 
 		#define EMIT_LOAD_OP(valueTypeId,name,llvmMemoryType,naturalAlignmentLog2,conversionOp) \
 			void valueTypeId##_##name(LoadOrStoreImm<naturalAlignmentLog2> imm) \

diff --git a/Source/Runtime/LLVMJIT.cpp b/Source/Runtime/LLVMJIT.cpp
@@ -752,7 +752,19 @@ namespace LLVMJIT
 			// our symbols can't be found in the JITed object file.
 			targetTriple += "-elf";
 		#endif
-		targetMachine = llvm::EngineBuilder().selectTarget(
+
+        // SL : fast-math a native level
+        llvm::TargetOptions targetOptions;
+
+        targetOptions.AllowFPOpFusion = llvm::FPOpFusion::Fast;
+        targetOptions.UnsafeFPMath = true;
+        targetOptions.NoInfsFPMath = true;
+        targetOptions.NoNaNsFPMath = true;
+        targetOptions.GuaranteedTailCallOpt = true;
+        targetOptions.NoTrappingFPMath = true;
+        targetOptions.FPDenormalMode = llvm::FPDenormal::IEEE;
+
+        targetMachine = llvm::EngineBuilder().setTargetOptions(targetOptions).selectTarget(
 			llvm::Triple(targetTriple),"",llvm::sys::getHostCPUName(),
 			#if defined(_WIN32) && !defined(_WIN64)
 				// Use SSE2 instead of the FPU on x86 for more control over how intermediate results are rounded.