From 657d1bed2368ab0355ddf69acd737a43d0bfabcf Mon Sep 17 00:00:00 2001
From: Justin Holewinski <justin.holewinski@gmail.com>
Date: Wed, 18 May 2011 15:42:23 +0000
Subject: [PATCH] PTX: add flag to disable mad/fma selection

Patch by Dan Bailey

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@131537 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/PTX/PTX.td             |  3 +++
 lib/Target/PTX/PTXInstrInfo.td    | 10 ++++++++--
 lib/Target/PTX/PTXSubtarget.cpp   |  3 ++-
 lib/Target/PTX/PTXSubtarget.h     |  7 ++++++-
 test/CodeGen/PTX/mad-disabling.ll | 16 ++++++++++++++++
 5 files changed, 35 insertions(+), 4 deletions(-)
 create mode 100644 test/CodeGen/PTX/mad-disabling.ll

diff --git a/lib/Target/PTX/PTX.td b/lib/Target/PTX/PTX.td
index cf10f0c2f59..231866a0895 100644
--- a/lib/Target/PTX/PTX.td
+++ b/lib/Target/PTX/PTX.td
@@ -24,6 +24,9 @@ include "llvm/Target/Target.td"
 def FeatureDouble : SubtargetFeature<"double", "SupportsDouble", "true",
                                      "Do not demote .f64 to .f32">;
 
+def FeatureNoFMA  : SubtargetFeature<"no-fma","SupportsFMA", "false",
+                                     "Disable Fused-Multiply Add">;
+
 //===- PTX Version --------------------------------------------------------===//
 
 def FeaturePTX20 : SubtargetFeature<"ptx20", "PTXVersion", "PTX_VERSION_2_0",
diff --git a/lib/Target/PTX/PTXInstrInfo.td b/lib/Target/PTX/PTXInstrInfo.td
index e90d4fffb0d..d5d08bed224 100644
--- a/lib/Target/PTX/PTXInstrInfo.td
+++ b/lib/Target/PTX/PTXInstrInfo.td
@@ -39,6 +39,10 @@ def DoesNotSupportPTX22 : Predicate<"!getSubtarget().supportsPTX22()">;
 def SupportsPTX23       : Predicate<"getSubtarget().supportsPTX23()">;
 def DoesNotSupportPTX23 : Predicate<"!getSubtarget().supportsPTX23()">;
 
+// Fused-Multiply Add
+def SupportsFMA         : Predicate<"getSubtarget().supportsFMA()">;
+def DoesNotSupportFMA   : Predicate<"!getSubtarget().supportsFMA()">;
+
 //===----------------------------------------------------------------------===//
 // Instruction Pattern Stuff
 //===----------------------------------------------------------------------===//
@@ -629,8 +633,8 @@ def FDIVri64SM10 : InstPTX<(outs RRegf64:$d),
 // In the short term, mad is supported on all PTX versions and we use a
 // default rounding mode no matter what shader model or PTX version.
 // TODO: Allow the rounding mode to be selectable through llc.
-defm FMADSM13 : PTX_FLOAT_4OP<"mad.rn", fmul, fadd>, Requires<[SupportsSM13]>;
-defm FMAD : PTX_FLOAT_4OP<"mad", fmul, fadd>, Requires<[DoesNotSupportSM13]>;
+defm FMADSM13 : PTX_FLOAT_4OP<"mad.rn", fmul, fadd>, Requires<[SupportsSM13, SupportsFMA]>;
+defm FMAD : PTX_FLOAT_4OP<"mad", fmul, fadd>, Requires<[DoesNotSupportSM13, SupportsFMA]>;
 
 ///===- Floating-Point Intrinsic Instructions -----------------------------===//
 
@@ -667,6 +671,8 @@ def FCOS64 : InstPTX<(outs RRegf64:$d),
 
 ///===- Comparison and Selection Instructions -----------------------------===//
 
+// .setp
+
 // Compare u16
 
 defm SETPEQu16 : PTX_SETP_I<RRegu16, "u16", i16imm, SETEQ,  "eq">;
diff --git a/lib/Target/PTX/PTXSubtarget.cpp b/lib/Target/PTX/PTXSubtarget.cpp
index e7542908e24..e8a1dfecd00 100644
--- a/lib/Target/PTX/PTXSubtarget.cpp
+++ b/lib/Target/PTX/PTXSubtarget.cpp
@@ -21,7 +21,8 @@ PTXSubtarget::PTXSubtarget(const std::string &TT, const std::string &FS,
   : PTXShaderModel(PTX_SM_1_0),
     PTXVersion(PTX_VERSION_2_0),
     SupportsDouble(false),
-    Is64Bit(is64Bit) {
+    SupportsFMA(true),
+    Is64Bit(is64Bit) {	
   std::string TARGET = "generic";
   ParseSubtargetFeatures(FS, TARGET);
 }
diff --git a/lib/Target/PTX/PTXSubtarget.h b/lib/Target/PTX/PTXSubtarget.h
index eebb284f375..59fa6965bba 100644
--- a/lib/Target/PTX/PTXSubtarget.h
+++ b/lib/Target/PTX/PTXSubtarget.h
@@ -49,7 +49,10 @@ namespace llvm {
 
       // The native .f64 type is supported on the hardware.
       bool SupportsDouble;
-
+      
+      // Support the fused-multiply add (FMA) and multiply-add (MAD) instructions
+      bool SupportsFMA;
+      
       // Use .u64 instead of .u32 for addresses.
       bool Is64Bit;
 
@@ -64,6 +67,8 @@ namespace llvm {
 
       bool is64Bit() const { return Is64Bit; }
 
+      bool supportsFMA() const { return SupportsFMA; }
+      
       bool supportsSM13() const { return PTXShaderModel >= PTX_SM_1_3; }
 
       bool supportsSM20() const { return PTXShaderModel >= PTX_SM_2_0; }
diff --git a/test/CodeGen/PTX/mad-disabling.ll b/test/CodeGen/PTX/mad-disabling.ll
new file mode 100644
index 00000000000..ad7b3417ec4
--- /dev/null
+++ b/test/CodeGen/PTX/mad-disabling.ll
@@ -0,0 +1,16 @@
+; RUN: llc < %s -march=ptx32 -mattr=+ptx20,+sm20 | grep "mad"
+; RUN: llc < %s -march=ptx32 -mattr=+ptx20,+sm20,+no-fma | grep -v "mad"
+
+define ptx_device float @test_mul_add_f(float %x, float %y, float %z) {
+entry:
+  %a = fmul float %x, %y
+  %b = fadd float %a, %z
+  ret float %b
+}
+
+define ptx_device double @test_mul_add_d(double %x, double %y, double %z) {
+entry:
+  %a = fmul double %x, %y
+  %b = fadd double %a, %z
+  ret double %b
+}
-- 
2.34.1