From b035a5740ce32a5bddfecb826845a350f589e377 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Thu, 29 Jan 2015 19:34:25 +0000
Subject: [PATCH] R600/SI: Add subtarget feature for if f32 fma is fast

llvm-svn: 227483
---
 llvm/lib/Target/R600/AMDGPU.td           |  6 ++++++
 llvm/lib/Target/R600/AMDGPUSubtarget.cpp |  3 ++-
 llvm/lib/Target/R600/AMDGPUSubtarget.h   |  5 +++++
 llvm/lib/Target/R600/Processors.td       | 12 +++++++++---
 llvm/lib/Target/R600/SIISelLowering.cpp  |  2 +-
 5 files changed, 23 insertions(+), 5 deletions(-)

diff --git a/llvm/lib/Target/R600/AMDGPU.td b/llvm/lib/Target/R600/AMDGPU.td
index 1df4448..be9a149 100644
--- a/llvm/lib/Target/R600/AMDGPU.td
+++ b/llvm/lib/Target/R600/AMDGPU.td
@@ -48,6 +48,12 @@ def FeatureFP64Denormals : SubtargetFeature<"fp64-denormals",
         "Enable double precision denormal handling",
         [FeatureFP64]>;
 
+def FeatureFastFMAF32 : SubtargetFeature<"fast-fmaf",
+        "FastFMAF32",
+        "true",
+        "Assuming f32 fma is at least as fast as mul + add",
+        []>;
+
 // Some instructions do not support denormals despite this flag. Using
 // fp32 denormals also causes instructions to run at the double
 // precision rate for the device.
diff --git a/llvm/lib/Target/R600/AMDGPUSubtarget.cpp b/llvm/lib/Target/R600/AMDGPUSubtarget.cpp
index 541dbab..e08a7dd 100644
--- a/llvm/lib/Target/R600/AMDGPUSubtarget.cpp
+++ b/llvm/lib/Target/R600/AMDGPUSubtarget.cpp
@@ -65,7 +65,8 @@ AMDGPUSubtarget::AMDGPUSubtarget(StringRef TT, StringRef GPU, StringRef FS,
     : AMDGPUGenSubtargetInfo(TT, GPU, FS), DevName(GPU), Is64bit(false),
       DumpCode(false), R600ALUInst(false), HasVertexCache(false),
       TexVTXClauseSize(0), Gen(AMDGPUSubtarget::R600), FP64(false),
-      FP64Denormals(false), FP32Denormals(false), CaymanISA(false),
+      FP64Denormals(false), FP32Denormals(false),
+      FastFMAF32(false), CaymanISA(false),
       FlatAddressSpace(false), EnableIRStructurizer(true),
       EnablePromoteAlloca(false), EnableIfCvt(true),
       EnableLoadStoreOpt(false), WavefrontSize(0), CFALUBug(false), LocalMemorySize(0),
diff --git a/llvm/lib/Target/R600/AMDGPUSubtarget.h b/llvm/lib/Target/R600/AMDGPUSubtarget.h
index 389cc8c..0c0145c 100644
--- a/llvm/lib/Target/R600/AMDGPUSubtarget.h
+++ b/llvm/lib/Target/R600/AMDGPUSubtarget.h
@@ -55,6 +55,7 @@ private:
   bool FP64;
   bool FP64Denormals;
   bool FP32Denormals;
+  bool FastFMAF32;
   bool CaymanISA;
   bool FlatAddressSpace;
   bool EnableIRStructurizer;
@@ -127,6 +128,10 @@ public:
     return FP64Denormals;
   }
 
+  bool hasFastFMAF32() const {
+    return FastFMAF32;
+  }
+
   bool hasFlatAddressSpace() const {
     return FlatAddressSpace;
   }
diff --git a/llvm/lib/Target/R600/Processors.td b/llvm/lib/Target/R600/Processors.td
index 6593016..fb5aa61 100644
--- a/llvm/lib/Target/R600/Processors.td
+++ b/llvm/lib/Target/R600/Processors.td
@@ -83,9 +83,13 @@ def : Proc<"cayman",     R600_VLIW4_Itin,
 // Southern Islands
 //===----------------------------------------------------------------------===//
 
-def : ProcessorModel<"SI",       SIFullSpeedModel, [FeatureSouthernIslands]>;
+def : ProcessorModel<"SI", SIFullSpeedModel,
+  [FeatureSouthernIslands, FeatureFastFMAF32]
+>;
 
-def : ProcessorModel<"tahiti",   SIFullSpeedModel, [FeatureSouthernIslands]>;
+def : ProcessorModel<"tahiti",   SIFullSpeedModel,
+  [FeatureSouthernIslands, FeatureFastFMAF32]
+>;
 
 def : ProcessorModel<"pitcairn", SIQuarterSpeedModel, [FeatureSouthernIslands]>;
 
@@ -105,7 +109,9 @@ def : ProcessorModel<"kabini",     SIQuarterSpeedModel, [FeatureSeaIslands]>;
 
 def : ProcessorModel<"kaveri",     SIQuarterSpeedModel, [FeatureSeaIslands]>;
 
-def : ProcessorModel<"hawaii",     SIFullSpeedModel, [FeatureSeaIslands]>;
+def : ProcessorModel<"hawaii", SIFullSpeedModel,
+  [FeatureSeaIslands, FeatureFastFMAF32]
+>;
 
 def : ProcessorModel<"mullins",    SIQuarterSpeedModel, [FeatureSeaIslands]>;
 
diff --git a/llvm/lib/Target/R600/SIISelLowering.cpp b/llvm/lib/Target/R600/SIISelLowering.cpp
index 6b2ea06..12677f0 100644
--- a/llvm/lib/Target/R600/SIISelLowering.cpp
+++ b/llvm/lib/Target/R600/SIISelLowering.cpp
@@ -655,7 +655,7 @@ bool SITargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
 
   switch (VT.getSimpleVT().SimpleTy) {
   case MVT::f32:
-    return false; /* There is V_MAD_F32 for f32 */
+    return Subtarget->hasFastFMAF32();
   case MVT::f64:
     return true;
   default:
-- 
2.7.4