AMDGPU/EG: Add a new FeatureFMA and use it to selectively enable FMA instruction

Only used by pre-GCN targets
v2: fix predicate setting for FMA_Common

Differential Revision: https://reviews.llvm.org/D40692

llvm-svn: 319712

Author: jvesely

llvm/lib/Target/AMDGPU/AMDGPU.td

@@ -19,6 +19,12 @@ def FeatureFP64 : SubtargetFeature<"fp64",
   "Enable double precision operations"
 >;
 
+def FeatureFMA : SubtargetFeature<"fmaf",
+  "FMA",
+  "true",
+  "Enable single precision FMA (not as fast as mul+add, but fused)"
+>;
+
 def FeatureFastFMAF32 : SubtargetFeature<"fast-fmaf",
   "FastFMAF32",
   "true",

llvm/lib/Target/AMDGPU/AMDGPUInstructions.td

@@ -49,6 +49,7 @@ def NoFP16Denormals : Predicate<"!Subtarget->hasFP16Denormals()">;
 def NoFP32Denormals : Predicate<"!Subtarget->hasFP32Denormals()">;
 def NoFP64Denormals : Predicate<"!Subtarget->hasFP64Denormals()">;
 def UnsafeFPMath : Predicate<"TM.Options.UnsafeFPMath">;
+def FMA : Predicate<"Subtarget->hasFMA()">;
 
 def InstFlag : OperandWithDefaultOps <i32, (ops (i32 0))>;
 def ADDRIndirect : ComplexPattern<iPTR, 2, "SelectADDRIndirect", [], []>;

llvm/lib/Target/AMDGPU/AMDGPUSubtarget.h

@@ -140,6 +140,7 @@ class AMDGPUSubtarget : public AMDGPUGenSubtargetInfo {
 
   // Subtarget statically properties set by tablegen
   bool FP64;
+  bool FMA;
   bool IsGCN;
   bool GCN3Encoding;
   bool CIInsts;
@@ -348,6 +349,10 @@ class AMDGPUSubtarget : public AMDGPUGenSubtargetInfo {
     return CaymanISA;
   }
 
+  bool hasFMA() const {
+    return FMA;
+  }
+
   TrapHandlerAbi getTrapHandlerAbi() const {
     return isAmdHsaOS() ? TrapHandlerAbiHsa : TrapHandlerAbiNone;
   }

llvm/lib/Target/AMDGPU/R600ISelLowering.cpp

@@ -211,6 +211,11 @@ R600TargetLowering::R600TargetLowering(const TargetMachine &TM,
   setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
   setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
 
+  if (!Subtarget->hasFMA()) {
+    setOperationAction(ISD::FMA, MVT::f32, Expand);
+    setOperationAction(ISD::FMA, MVT::f64, Expand);
+  }
+
   setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
 
   const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };

llvm/lib/Target/AMDGPU/R600Instructions.td

@@ -989,7 +989,10 @@ class MULADD_IEEE_Common <bits<5> inst> : R600_3OP <
 class FMA_Common <bits<5> inst> : R600_3OP <
   inst, "FMA",
   [(set f32:$dst, (fma f32:$src0, f32:$src1, f32:$src2))], VecALU
->;
+>
+{
+  let OtherPredicates = [FMA];
+}
 
 class CNDE_Common <bits<5> inst> : R600_3OP <
   inst, "CNDE",

llvm/lib/Target/AMDGPU/R600Processors.td

@@ -53,7 +53,7 @@ def : Processor<"cedar", R600_VLIW5_Itin,
 >;
 
 def : Processor<"cypress", R600_VLIW5_Itin,
-  [FeatureEvergreen, FeatureWavefrontSize64, FeatureVertexCache]
+  [FeatureEvergreen, FeatureWavefrontSize64, FeatureVertexCache, FeatureFMA]
 >;
 
 def : Processor<"juniper", R600_VLIW5_Itin,
@@ -82,7 +82,7 @@ def : Processor<"caicos", R600_VLIW5_Itin,
 >;
 
 def : Processor<"cayman", R600_VLIW4_Itin,
-  [FeatureNorthernIslands, FeatureCaymanISA]
+  [FeatureNorthernIslands, FeatureCaymanISA, FeatureFMA]
 >;
 
 def : Processor<"turks", R600_VLIW5_Itin,

llvm/test/CodeGen/AMDGPU/fma.ll

@@ -1,5 +1,12 @@
 ; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=amdgcn -mcpu=tahiti -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
-; XUN:  llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
+; RUN:  llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=cypress -verify-machineinstrs < %s | FileCheck -check-prefix=EG -check-prefix=FUNC %s
+; RUN:  not llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=cedar -verify-machineinstrs < %s
+; RUN:  not llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=juniper -verify-machineinstrs < %s
+; RUN:  not llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=redwood -verify-machineinstrs < %s
+; RUN:  not llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=sumo -verify-machineinstrs < %s
+; RUN:  not llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=barts -verify-machineinstrs < %s
+; RUN:  not llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=caicos -verify-machineinstrs < %s
+; RUN:  not llc -amdgpu-scalarize-global-loads=false  -march=r600 -mcpu=turks -verify-machineinstrs < %s
 
 declare float @llvm.fma.f32(float, float, float) nounwind readnone
 declare <2 x float> @llvm.fma.v2f32(<2 x float>, <2 x float>, <2 x float>) nounwind readnone