R600/SI: Fold fabs/fneg into src input modifier

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@208480 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/R600/SIISelLowering.cpp b/lib/Target/R600/SIISelLowering.cpp
index 466d24771ce34036cac39165fb58d59d3111a84f..da8aa5b7a981da459265ef9cfc4f33b28c543e07 100644 (file)
--- a/lib/Target/R600/SIISelLowering.cpp
+++ b/lib/Target/R600/SIISelLowering.cpp
@@ -1323,6 +1323,7 @@ SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
   // e64 version if available, -1 otherwise
   int OpcodeE64 = AMDGPU::getVOPe64(Opcode);
   const MCInstrDesc *DescE64 = OpcodeE64 == -1 ? nullptr : &TII->get(OpcodeE64);
+  int InputModifiers[3] = {0};
 
   assert(!DescE64 || DescE64->getNumDefs() == NumDefs);
 
@@ -1399,7 +1400,10 @@ SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
       }
     }
 
-    if (DescE64 && !Immediate) {
+    if (Immediate)
+      continue;
+
+    if (DescE64) {
 
       // Test if it makes sense to switch to e64 encoding
       unsigned OtherRegClass = DescE64->OpInfo[Op].RegClass;
@@ -1418,6 +1422,31 @@ SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
         DescE64 = nullptr;
       }
     }
+
+    if (!DescE64 && !Promote2e64)
+      continue;
+    if (!Operand.isMachineOpcode())
+      continue;
+    if (Operand.getMachineOpcode() == AMDGPU::FNEG_SI) {
+      Ops.pop_back();
+      Ops.push_back(Operand.getOperand(0));
+      InputModifiers[i] = 1;
+      Promote2e64 = true;
+      if (!DescE64)
+        continue;
+      Desc = DescE64;
+      DescE64 = 0;
+    }
+    else if (Operand.getMachineOpcode() == AMDGPU::FABS_SI) {
+      Ops.pop_back();
+      Ops.push_back(Operand.getOperand(0));
+      InputModifiers[i] = 2;
+      Promote2e64 = true;
+      if (!DescE64)
+        continue;
+      Desc = DescE64;
+      DescE64 = 0;
+    }
   }
 
   if (Promote2e64) {
@@ -1425,7 +1454,7 @@ SDNode *SITargetLowering::foldOperands(MachineSDNode *Node,
     Ops.clear();
     for (unsigned i = 0; i < OldOps.size(); ++i) {
       // src_modifier
-      Ops.push_back(DAG.getTargetConstant(0, MVT::i32));
+      Ops.push_back(DAG.getTargetConstant(InputModifiers[i], MVT::i32));
       Ops.push_back(OldOps[i]);
     }
     // Add the modifier flags while promoting

diff --git a/test/CodeGen/R600/fabs.ll b/test/CodeGen/R600/fabs.ll
index 2cd3a4f604f23db7a0a6d01440eb98a2f164235d..b87ce22540950b93e65a5a4f7f7b58a04e3327bb 100644 (file)
--- a/test/CodeGen/R600/fabs.ll
+++ b/test/CodeGen/R600/fabs.ll
@@ -49,6 +49,17 @@ entry:
   ret void
 }
 
+; SI-CHECK-LABEL: @fabs_fold
+; SI-CHECK-NOT: V_AND_B32_e32
+; SI-CHECK: V_MUL_F32_e64 v{{[0-9]+}}, s{{[0-9]+}}, |v{{[0-9]+}}|
+define void @fabs_fold(float addrspace(1)* %out, float %in0, float %in1) {
+entry:
+  %0 = call float @fabs(float %in0)
+  %1 = fmul float %0, %in1
+  store float %1, float addrspace(1)* %out
+  ret void
+}
+
 declare float @fabs(float ) readnone
 declare <2 x float> @llvm.fabs.v2f32(<2 x float> ) readnone
 declare <4 x float> @llvm.fabs.v4f32(<4 x float> ) readnone

diff --git a/test/CodeGen/R600/fneg.ll b/test/CodeGen/R600/fneg.ll
index 7ad760cadbb0b35bcd5e43559d81657eb649cdf7..4cddc73789562cfc7aa6d0f88054285b9d3248a2 100644 (file)
--- a/test/CodeGen/R600/fneg.ll
+++ b/test/CodeGen/R600/fneg.ll
@@ -59,3 +59,14 @@ entry:
   store float %1, float addrspace(1)* %out
   ret void
 }
+
+; SI-CHECK-LABEL: @fneg_fold
+; SI-CHECK-NOT: V_XOR_B32
+; SI-CHECK: V_MUL_F32_e64 v{{[0-9]+}}, s{{[0-9]+}}, -v{{[0-9]+}}
+define void @fneg_fold(float addrspace(1)* %out, float %in) {
+entry:
+  %0 = fsub float -0.0, %in
+  %1 = fmul float %0, %in
+  store float %1, float addrspace(1)* %out
+  ret void
+}