AMDGPU: Add fp legacy instruction intrinsics

This could use some additional optimization work
to use mad/mac legacy.

llvm-svn: 276764

5 files changed, 21 insertions, 2 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 7bf414facda..8435551c3c5 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -2699,7 +2699,9 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(TRIG_PREOP)
   NODE_NAME_CASE(RCP)
   NODE_NAME_CASE(RSQ)
+  NODE_NAME_CASE(RCP_LEGACY)
   NODE_NAME_CASE(RSQ_LEGACY)
+  NODE_NAME_CASE(FMUL_LEGACY)
   NODE_NAME_CASE(RSQ_CLAMP)
   NODE_NAME_CASE(LDEXP)
   NODE_NAME_CASE(FP_CLASS)
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index e79bb6724b7..f1fde8069a1 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -249,7 +249,9 @@ enum NodeType : unsigned {
   //            For f64, max error 2^29 ULP, handles denormals.
   RCP,
   RSQ,
+  RCP_LEGACY,
   RSQ_LEGACY,
+  FMUL_LEGACY,
   RSQ_CLAMP,
   LDEXP,
   FP_CLASS,
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td b/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
index 85675c800cf..f16ea8de429 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstrInfo.td
@@ -67,6 +67,7 @@ def AMDGPUrcp : SDNode<"AMDGPUISD::RCP", SDTFPUnaryOp>;
 def AMDGPUrsq : SDNode<"AMDGPUISD::RSQ", SDTFPUnaryOp>;
 
 // out = 1.0 / sqrt(a)
+def AMDGPUrcp_legacy : SDNode<"AMDGPUISD::RCP_LEGACY", SDTFPUnaryOp>;
 def AMDGPUrsq_legacy : SDNode<"AMDGPUISD::RSQ_LEGACY", SDTFPUnaryOp>;
 
 // out = 1.0 / sqrt(a) result clamped to +/- max_float.
@@ -84,6 +85,10 @@ def AMDGPUfmax_legacy : SDNode<"AMDGPUISD::FMAX_LEGACY", SDTFPBinOp,
   []
 >;
 
+def AMDGPUfmul_legacy : SDNode<"AMDGPUISD::FMUL_LEGACY", SDTFPBinOp,
+  [SDNPCommutative, SDNPAssociative]
+>;
+
 def AMDGPUclamp : SDNode<"AMDGPUISD::CLAMP", SDTFPTernaryOp, []>;
 
 // out = max(a, b) a and b are signed ints
diff --git a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
index 323525f2151..6746c8a97f1 100644
--- a/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/SIISelLowering.cpp
@@ -1996,6 +1996,11 @@ SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
 
     return DAG.getNode(AMDGPUISD::RSQ_LEGACY, DL, VT, Op.getOperand(1));
   }
+  case Intrinsic::amdgcn_rcp_legacy: {
+    if (Subtarget->getGeneration() >= SISubtarget::VOLCANIC_ISLANDS)
+      return emitRemovedIntrinsicError(DAG, DL, VT);
+    return DAG.getNode(AMDGPUISD::RCP_LEGACY, DL, VT, Op.getOperand(1));
+  }
   case Intrinsic::amdgcn_rsq_clamp: {
     if (Subtarget->getGeneration() < SISubtarget::VOLCANIC_ISLANDS)
       return DAG.getNode(AMDGPUISD::RSQ_CLAMP, DL, VT, Op.getOperand(1));
@@ -2208,6 +2213,9 @@ SDValue SITargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
     return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, Op->getVTList(), Src0,
                        Denominator, Numerator);
   }
+  case Intrinsic::amdgcn_fmul_legacy:
+    return DAG.getNode(AMDGPUISD::FMUL_LEGACY, DL, VT,
+                       Op.getOperand(1), Op.getOperand(2));
   case Intrinsic::amdgcn_sffbh:
   case AMDGPUIntrinsic::AMDGPU_flbit_i32: // Legacy name.
     return DAG.getNode(AMDGPUISD::FFBH_I32, DL, VT, Op.getOperand(1));
@@ -3356,6 +3364,7 @@ SDValue SITargetLowering::PerformDAGCombine(SDNode *N,
   case AMDGPUISD::FRACT:
   case AMDGPUISD::RCP:
   case AMDGPUISD::RSQ:
+  case AMDGPUISD::RCP_LEGACY:
   case AMDGPUISD::RSQ_LEGACY:
   case AMDGPUISD::RSQ_CLAMP:
   case AMDGPUISD::LDEXP: {
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index 944248a7fdd..61460057803 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -1402,7 +1402,8 @@ defm V_MOV_FED_B32 : VOP1InstSI <vop1<0x9>, "v_mov_fed_b32", VOP_I32_I32>;
 defm V_LOG_CLAMP_F32 : VOP1InstSI <vop1<0x26>, "v_log_clamp_f32",
   VOP_F32_F32, int_amdgcn_log_clamp>;
 defm V_RCP_CLAMP_F32 : VOP1InstSI <vop1<0x28>, "v_rcp_clamp_f32", VOP_F32_F32>;
-defm V_RCP_LEGACY_F32 : VOP1InstSI <vop1<0x29>, "v_rcp_legacy_f32", VOP_F32_F32>;
+defm V_RCP_LEGACY_F32 : VOP1InstSI <vop1<0x29>, "v_rcp_legacy_f32",
+  VOP_F32_F32, AMDGPUrcp_legacy>;
 defm V_RSQ_CLAMP_F32 : VOP1InstSI <vop1<0x2c>, "v_rsq_clamp_f32",
   VOP_F32_F32, AMDGPUrsq_clamp
 >;
@@ -1496,7 +1497,7 @@ defm V_SUBREV_F32 : VOP2Inst <vop2<0x5, 0x3>, "v_subrev_f32",
 let isCommutable = 1 in {
 
 defm V_MUL_LEGACY_F32 : VOP2Inst <vop2<0x7, 0x4>, "v_mul_legacy_f32",
-  VOP_F32_F32_F32
+  VOP_F32_F32_F32, AMDGPUfmul_legacy
 >;
 
 defm V_MUL_F32 : VOP2Inst <vop2<0x8, 0x5>, "v_mul_f32",