R600/SI: Add intrinsics for various math instructions.

These will be used for custom lowering and for library
implementations of various math functions, so it's useful
to expose these as builtins.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@211247 91177308-0d34-0410-b5e6-96231b3b80d8

20 files changed, 393 insertions, 32 deletions

diff --git a/include/llvm/IR/IntrinsicsR600.td b/include/llvm/IR/IntrinsicsR600.td
index ecb5668d8e..09607d5834 100644
--- a/include/llvm/IR/IntrinsicsR600.td
+++ b/include/llvm/IR/IntrinsicsR600.td
@@ -33,4 +33,40 @@ defm int_r600_read_tgid : R600ReadPreloadRegisterIntrinsic_xyz <
                                        "__builtin_r600_read_tgid">;
 defm int_r600_read_tidig : R600ReadPreloadRegisterIntrinsic_xyz <
                                        "__builtin_r600_read_tidig">;
+
 } // End TargetPrefix = "r600"
+
+let TargetPrefix = "AMDGPU" in {
+def int_AMDGPU_div_scale :
+  Intrinsic<[llvm_anyfloat_ty, llvm_i1_ty],
+            [LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>,
+            GCCBuiltin<"__builtin_amdgpu_div_scale">;
+
+def int_AMDGPU_div_fmas :
+  Intrinsic<[llvm_anyfloat_ty],
+            [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>],
+            [IntrNoMem]>,
+            GCCBuiltin<"__builtin_amdgpu_div_fmas">;
+
+def int_AMDGPU_div_fixup :
+  Intrinsic<[llvm_anyfloat_ty],
+            [LLVMMatchType<0>, LLVMMatchType<0>, LLVMMatchType<0>], [IntrNoMem]>,
+            GCCBuiltin<"__builtin_amdgpu_div_fixup">;
+
+def int_AMDGPU_trig_preop :
+  Intrinsic<[llvm_anyfloat_ty],
+            [LLVMMatchType<0>, llvm_i32_ty], [IntrNoMem]>,
+            GCCBuiltin<"__builtin_amdgpu_trig_preop">;
+
+def int_AMDGPU_rcp :
+  Intrinsic<[llvm_anyfloat_ty],
+            [LLVMMatchType<0>], [IntrNoMem]>,
+            GCCBuiltin<"__builtin_amdgpu_rcp">;
+
+def int_AMDGPU_rsq :
+  Intrinsic<[llvm_anyfloat_ty],
+            [LLVMMatchType<0>], [IntrNoMem]>,
+            GCCBuiltin<"__builtin_amdgpu_rsq">;
+
+
+} // End TargetPrefix = "AMDGPU"
diff --git a/lib/Target/R600/AMDGPUISelLowering.cpp b/lib/Target/R600/AMDGPUISelLowering.cpp
index 34c2b2bf61..1aa92fadbe 100644
--- a/lib/Target/R600/AMDGPUISelLowering.cpp
+++ b/lib/Target/R600/AMDGPUISelLowering.cpp
@@ -842,6 +842,28 @@ SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
       return DAG.getNode(AMDGPUISD::CLAMP, DL, VT,
                          Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
 
+    case Intrinsic::AMDGPU_div_scale:
+      return DAG.getNode(AMDGPUISD::DIV_SCALE, DL, VT,
+                         Op.getOperand(1), Op.getOperand(2));
+
+    case Intrinsic::AMDGPU_div_fmas:
+      return DAG.getNode(AMDGPUISD::DIV_FMAS, DL, VT,
+                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
+
+    case Intrinsic::AMDGPU_div_fixup:
+      return DAG.getNode(AMDGPUISD::DIV_FIXUP, DL, VT,
+                         Op.getOperand(1), Op.getOperand(2), Op.getOperand(3));
+
+    case Intrinsic::AMDGPU_trig_preop:
+      return DAG.getNode(AMDGPUISD::TRIG_PREOP, DL, VT,
+                         Op.getOperand(1), Op.getOperand(2));
+
+    case Intrinsic::AMDGPU_rcp:
+      return DAG.getNode(AMDGPUISD::RCP, DL, VT, Op.getOperand(1));
+
+    case Intrinsic::AMDGPU_rsq:
+      return DAG.getNode(AMDGPUISD::RSQ, DL, VT, Op.getOperand(1));
+
     case AMDGPUIntrinsic::AMDGPU_imax:
       return DAG.getNode(AMDGPUISD::SMAX, DL, VT, Op.getOperand(1),
                                                   Op.getOperand(2));
@@ -2042,6 +2064,14 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(FMIN)
   NODE_NAME_CASE(SMIN)
   NODE_NAME_CASE(UMIN)
+  NODE_NAME_CASE(URECIP)
+  NODE_NAME_CASE(DIV_SCALE)
+  NODE_NAME_CASE(DIV_FMAS)
+  NODE_NAME_CASE(DIV_FIXUP)
+  NODE_NAME_CASE(TRIG_PREOP)
+  NODE_NAME_CASE(RCP)
+  NODE_NAME_CASE(RSQ)
+  NODE_NAME_CASE(DOT4)
   NODE_NAME_CASE(BFE_U32)
   NODE_NAME_CASE(BFE_I32)
   NODE_NAME_CASE(BFI)
@@ -2051,8 +2081,6 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(MUL_I24)
   NODE_NAME_CASE(MAD_U24)
   NODE_NAME_CASE(MAD_I24)
-  NODE_NAME_CASE(URECIP)
-  NODE_NAME_CASE(DOT4)
   NODE_NAME_CASE(EXPORT)
   NODE_NAME_CASE(CONST_ADDRESS)
   NODE_NAME_CASE(REGISTER_LOAD)
diff --git a/lib/Target/R600/AMDGPUISelLowering.h b/lib/Target/R600/AMDGPUISelLowering.h
index b2bb2579dc..e2000a04ba 100644
--- a/lib/Target/R600/AMDGPUISelLowering.h
+++ b/lib/Target/R600/AMDGPUISelLowering.h
@@ -175,6 +175,9 @@ enum {
   DWORDADDR,
   FRACT,
   CLAMP,
+
+  // SIN_HW, COS_HW - f32 for SI, 1 ULP max error, valid from -100 pi to 100 pi.
+  // Denormals handled on some parts.
   COS_HW,
   SIN_HW,
   FMAX,
@@ -184,6 +187,15 @@ enum {
   SMIN,
   UMIN,
   URECIP,
+  DIV_SCALE,
+  DIV_FMAS,
+  DIV_FIXUP,
+  TRIG_PREOP, // 1 ULP max error for f64
+
+  // RCP, RSQ - For f32, 1 ULP max error, no denormal handling.
+  //            For f64, max error 2^29 ULP, handles denormals.
+  RCP,
+  RSQ,
   DOT4,
   BFE_U32, // Extract range of bits with zero extension to 32-bits.
   BFE_I32, // Extract range of bits with sign extension to 32-bits.
diff --git a/lib/Target/R600/AMDGPUInstrInfo.td b/lib/Target/R600/AMDGPUInstrInfo.td
index 942a9e8ff3..d0ee40a678 100644
--- a/lib/Target/R600/AMDGPUInstrInfo.td
+++ b/lib/Target/R600/AMDGPUInstrInfo.td
@@ -19,6 +19,14 @@ def AMDGPUDTIntTernaryOp : SDTypeProfile<1, 3, [
   SDTCisSameAs<0, 1>, SDTCisSameAs<0, 2>, SDTCisInt<0>, SDTCisInt<3>
 ]>;
 
+def AMDGPUTrigPreOp : SDTypeProfile<1, 2,
+  [SDTCisSameAs<0, 1>, SDTCisFP<0>, SDTCisInt<2>]
+>;
+
+def AMDGPUDivScaleOp : SDTypeProfile<2, 3,
+  [SDTCisFP<0>, SDTCisInt<1>, SDTCisSameAs<0, 2>, SDTCisSameAs<0, 3>, SDTCisSameAs<0, 4>]
+>;
+
 //===----------------------------------------------------------------------===//
 // AMDGPU DAG Nodes
 //
@@ -29,6 +37,12 @@ def AMDGPUdwordaddr : SDNode<"AMDGPUISD::DWORDADDR", SDTIntUnaryOp>;
 // out = a - floor(a)
 def AMDGPUfract : SDNode<"AMDGPUISD::FRACT", SDTFPUnaryOp>;
 
+// out = 1.0 / a
+def AMDGPUrcp : SDNode<"AMDGPUISD::RCP", SDTFPUnaryOp>;
+
+// out = 1.0 / sqrt(a)
+def AMDGPUrsq : SDNode<"AMDGPUISD::RSQ", SDTFPUnaryOp>;
+
 // out = max(a, b) a and b are floats
 def AMDGPUfmax : SDNode<"AMDGPUISD::FMAX", SDTFPBinOp,
   [SDNPCommutative, SDNPAssociative]
@@ -78,6 +92,21 @@ def AMDGPUcvt_f32_ubyte3 : SDNode<"AMDGPUISD::CVT_F32_UBYTE3",
 // e is rounding error
 def AMDGPUurecip : SDNode<"AMDGPUISD::URECIP", SDTIntUnaryOp>;
 
+// Special case divide preop and flags.
+def AMDGPUdiv_scale : SDNode<"AMDGPUISD::DIV_SCALE", AMDGPUDivScaleOp>;
+
+//  Special case divide FMA with scale and flags (src0 = Quotient,
+//  src1 = Denominator, src2 = Numerator).
+def AMDGPUdiv_fmas : SDNode<"AMDGPUISD::DIV_FMAS", SDTFPTernaryOp>;
+
+// Single or double precision division fixup.
+// Special case divide fixup and flags(src0 = Quotient, src1 =
+// Denominator, src2 = Numerator).
+def AMDGPUdiv_fixup : SDNode<"AMDGPUISD::DIV_FIXUP", SDTFPTernaryOp>;
+
+// Look Up 2.0 / pi src0 with segment select src1[4:0]
+def AMDGPUtrig_preop : SDNode<"AMDGPUISD::TRIG_PREOP", AMDGPUTrigPreOp>;
+
 def AMDGPUregister_load : SDNode<"AMDGPUISD::REGISTER_LOAD",
                           SDTypeProfile<1, 2, [SDTCisPtrTy<1>, SDTCisInt<2>]>,
                           [SDNPHasChain, SDNPMayLoad]>;
diff --git a/lib/Target/R600/AMDGPUInstructions.td b/lib/Target/R600/AMDGPUInstructions.td
index 8bfc11cd46..14bfd8cc18 100644
--- a/lib/Target/R600/AMDGPUInstructions.td
+++ b/lib/Target/R600/AMDGPUInstructions.td
@@ -519,6 +519,16 @@ multiclass Expand24UBitOps<Instruction MulInst, Instruction AddInst> {
   >;
 }
 
+class RcpPat<Instruction RcpInst, ValueType vt> : Pat <
+  (fdiv FP_ONE, vt:$src),
+  (RcpInst $src)
+>;
+
+class RsqPat<Instruction RsqInst, ValueType vt> : Pat <
+  (AMDGPUrcp (fsqrt vt:$src)),
+  (RsqInst $src)
+>;
+
 include "R600Instructions.td"
 include "R700Instructions.td"
 include "EvergreenInstructions.td"
diff --git a/lib/Target/R600/AMDGPUIntrinsics.td b/lib/Target/R600/AMDGPUIntrinsics.td
index 6dc7612d46..538b4cd8af 100644
--- a/lib/Target/R600/AMDGPUIntrinsics.td
+++ b/lib/Target/R600/AMDGPUIntrinsics.td
@@ -30,8 +30,6 @@ let TargetPrefix = "AMDGPU", isTarget = 1 in {
   def int_AMDGPU_lrp : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
   def int_AMDGPU_mul : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
   def int_AMDGPU_pow : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
-  def int_AMDGPU_rcp : Intrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>;
-  def int_AMDGPU_rsq : Intrinsic<[llvm_float_ty], [llvm_float_ty], [IntrNoMem]>;
   def int_AMDGPU_seq : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
   def int_AMDGPU_sgt : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
   def int_AMDGPU_sge : Intrinsic<[llvm_float_ty], [llvm_float_ty, llvm_float_ty], [IntrNoMem]>;
diff --git a/lib/Target/R600/R600Instructions.td b/lib/Target/R600/R600Instructions.td
index 58c704d8ec..47b7da0955 100644
--- a/lib/Target/R600/R600Instructions.td
+++ b/lib/Target/R600/R600Instructions.td
@@ -1083,7 +1083,7 @@ class RECIP_UINT_Common <bits<11> inst> : R600_1OP_Helper <
 }
 
 class RECIPSQRT_CLAMPED_Common <bits<11> inst> : R600_1OP_Helper <
-  inst, "RECIPSQRT_CLAMPED", int_AMDGPU_rsq
+  inst, "RECIPSQRT_CLAMPED", AMDGPUrsq
 > {
   let Itinerary = TransALU;
 }
diff --git a/lib/Target/R600/SIInsertWaits.cpp b/lib/Target/R600/SIInsertWaits.cpp
index a17fed7e7e..173332674f 100644
--- a/lib/Target/R600/SIInsertWaits.cpp
+++ b/lib/Target/R600/SIInsertWaits.cpp
@@ -341,6 +341,8 @@ Counters SIInsertWaits::handleOperands(MachineInstr &MI) {
   return Result;
 }
 
+// FIXME: Insert waits listed in Table 4.2 "Required User-Inserted Wait States"
+// around other non-memory instructions.
 bool SIInsertWaits::runOnMachineFunction(MachineFunction &MF) {
   bool Changes = false;
 
diff --git a/lib/Target/R600/SIInstructions.td b/lib/Target/R600/SIInstructions.td
index 60eb8f9781..26024dc4f2 100644
--- a/lib/Target/R600/SIInstructions.td
+++ b/lib/Target/R600/SIInstructions.td
@@ -1116,22 +1116,23 @@ defm V_LOG_CLAMP_F32 : VOP1_32 <0x00000026, "V_LOG_CLAMP_F32", []>;
 defm V_LOG_F32 : VOP1_32 <0x00000027, "V_LOG_F32",
   [(set f32:$dst, (flog2 f32:$src0))]
 >;
+
 defm V_RCP_CLAMP_F32 : VOP1_32 <0x00000028, "V_RCP_CLAMP_F32", []>;
 defm V_RCP_LEGACY_F32 : VOP1_32 <0x00000029, "V_RCP_LEGACY_F32", []>;
 defm V_RCP_F32 : VOP1_32 <0x0000002a, "V_RCP_F32",
-  [(set f32:$dst, (fdiv FP_ONE, f32:$src0))]
+  [(set f32:$dst, (AMDGPUrcp f32:$src0))]
 >;
 defm V_RCP_IFLAG_F32 : VOP1_32 <0x0000002b, "V_RCP_IFLAG_F32", []>;
 defm V_RSQ_CLAMP_F32 : VOP1_32 <0x0000002c, "V_RSQ_CLAMP_F32", []>;
 defm V_RSQ_LEGACY_F32 : VOP1_32 <
   0x0000002d, "V_RSQ_LEGACY_F32",
-  [(set f32:$dst, (int_AMDGPU_rsq f32:$src0))]
+  [(set f32:$dst, (AMDGPUrsq f32:$src0))]
 >;
 defm V_RSQ_F32 : VOP1_32 <0x0000002e, "V_RSQ_F32",
   [(set f32:$dst, (fdiv FP_ONE, (fsqrt f32:$src0)))]
 >;
 defm V_RCP_F64 : VOP1_64 <0x0000002f, "V_RCP_F64",
-  [(set f64:$dst, (fdiv FP_ONE, f64:$src0))]
+  [(set f64:$dst, (AMDGPUrcp f64:$src0))]
 >;
 defm V_RCP_CLAMP_F64 : VOP1_64 <0x00000030, "V_RCP_CLAMP_F64", []>;
 defm V_RSQ_F64 : VOP1_64 <0x00000031, "V_RSQ_F64",
@@ -1417,8 +1418,12 @@ defm V_MULLIT_F32 : VOP3_32 <0x00000150, "V_MULLIT_F32", []>;
 //def V_SAD_U16 : VOP3_U16 <0x0000015c, "V_SAD_U16", []>;
 defm V_SAD_U32 : VOP3_32 <0x0000015d, "V_SAD_U32", []>;
 ////def V_CVT_PK_U8_F32 : VOP3_U8 <0x0000015e, "V_CVT_PK_U8_F32", []>;
-defm V_DIV_FIXUP_F32 : VOP3_32 <0x0000015f, "V_DIV_FIXUP_F32", []>;
-def V_DIV_FIXUP_F64 : VOP3_64 <0x00000160, "V_DIV_FIXUP_F64", []>;
+defm V_DIV_FIXUP_F32 : VOP3_32 <0x0000015f, "V_DIV_FIXUP_F32",
+  [(set f32:$dst, (AMDGPUdiv_fixup f32:$src0, f32:$src1, f32:$src2))]
+>;
+def V_DIV_FIXUP_F64 : VOP3_64 <0x00000160, "V_DIV_FIXUP_F64",
+  [(set f64:$dst, (AMDGPUdiv_fixup f64:$src0, f64:$src1, f64:$src2))]
+>;
 
 def V_LSHL_B64 : VOP3_64_32 <0x00000161, "V_LSHL_B64",
   [(set i64:$dst, (shl i64:$src0, i32:$src1))]
@@ -1452,12 +1457,19 @@ defm V_MUL_HI_I32 : VOP3_32 <0x0000016c, "V_MUL_HI_I32", []>;
 
 defm V_DIV_SCALE_F32 : VOP3_32 <0x0000016d, "V_DIV_SCALE_F32", []>;
 def V_DIV_SCALE_F64 : VOP3_64 <0x0000016e, "V_DIV_SCALE_F64", []>;
-defm V_DIV_FMAS_F32 : VOP3_32 <0x0000016f, "V_DIV_FMAS_F32", []>;
-def V_DIV_FMAS_F64 : VOP3_64 <0x00000170, "V_DIV_FMAS_F64", []>;
+
+defm V_DIV_FMAS_F32 : VOP3_32 <0x0000016f, "V_DIV_FMAS_F32",
+  [(set f32:$dst, (AMDGPUdiv_fmas f32:$src0, f32:$src1, f32:$src2))]
+>;
+def V_DIV_FMAS_F64 : VOP3_64 <0x00000170, "V_DIV_FMAS_F64",
+  [(set f64:$dst, (AMDGPUdiv_fmas f64:$src0, f64:$src1, f64:$src2))]
+>;
 //def V_MSAD_U8 : VOP3_U8 <0x00000171, "V_MSAD_U8", []>;
 //def V_QSAD_U8 : VOP3_U8 <0x00000172, "V_QSAD_U8", []>;
 //def V_MQSAD_U8 : VOP3_U8 <0x00000173, "V_MQSAD_U8", []>;
-def V_TRIG_PREOP_F64 : VOP3_64 <0x00000174, "V_TRIG_PREOP_F64", []>;
+def V_TRIG_PREOP_F64 : VOP3_64_32 <0x00000174, "V_TRIG_PREOP_F64",
+  [(set f64:$dst, (AMDGPUtrig_preop f64:$src0, i32:$src1))]
+>;
 
 //===----------------------------------------------------------------------===//
 // Pseudo Instructions
@@ -1749,6 +1761,15 @@ def : Pat <
 >;
 
 //===----------------------------------------------------------------------===//
+// VOP1 Patterns
+//===----------------------------------------------------------------------===//
+
+def : RcpPat<V_RCP_F32_e32, f32>;
+def : RcpPat<V_RCP_F64_e32, f64>;
+def : RsqPat<V_RSQ_F32_e32, f32>;
+def : RsqPat<V_RSQ_F64_e32, f64>;
+
+//===----------------------------------------------------------------------===//
 // VOP2 Patterns
 //===----------------------------------------------------------------------===//
 
diff --git a/lib/Transforms/InstCombine/InstCombineCalls.cpp b/lib/Transforms/InstCombine/InstCombineCalls.cpp
index d4bdd75fa8..ff74564158 100644
--- a/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -922,6 +922,20 @@ Instruction *InstCombiner::visitCallInst(CallInst &CI) {
     break;
   }
 
+  case Intrinsic::AMDGPU_rcp: {
+    if (const ConstantFP *C = dyn_cast<ConstantFP>(II->getArgOperand(0))) {
+      const APFloat &ArgVal = C->getValueAPF();
+      APFloat Val(ArgVal.getSemantics(), 1.0);
+      APFloat::opStatus Status = Val.divide(ArgVal,
+                                            APFloat::rmNearestTiesToEven);
+      // Only do this if it was exact and therefore not dependent on the
+      // rounding mode.
+      if (Status == APFloat::opOK)
+        return ReplaceInstUsesWith(CI, ConstantFP::get(II->getContext(), Val));
+    }
+
+    break;
+  }
   case Intrinsic::stackrestore: {
     // If the save is right next to the restore, remove the restore.  This can
     // happen when variable allocas are DCE'd.
diff --git a/test/CodeGen/R600/big_alu.ll b/test/CodeGen/R600/big_alu.ll
index 6b683769fe..511e8ef629 100644
--- a/test/CodeGen/R600/big_alu.ll
+++ b/test/CodeGen/R600/big_alu.ll
@@ -101,7 +101,7 @@ IF137:                                            ; preds = %main_body
   %88 = insertelement <4 x float> %87, float %32, i32 2
   %89 = insertelement <4 x float> %88, float 0.000000e+00, i32 3
   %90 = call float @llvm.AMDGPU.dp4(<4 x float> %85, <4 x float> %89)
-  %91 = call float @llvm.AMDGPU.rsq(float %90)
+  %91 = call float @llvm.AMDGPU.rsq.f32(float %90)
   %92 = fmul float %30, %91
   %93 = fmul float %31, %91
   %94 = fmul float %32, %91
@@ -344,7 +344,7 @@ ENDIF136:                                         ; preds = %main_body, %ENDIF15
   %325 = insertelement <4 x float> %324, float %318, i32 2
   %326 = insertelement <4 x float> %325, float 0.000000e+00, i32 3
   %327 = call float @llvm.AMDGPU.dp4(<4 x float> %322, <4 x float> %326)
-  %328 = call float @llvm.AMDGPU.rsq(float %327)
+  %328 = call float @llvm.AMDGPU.rsq.f32(float %327)
   %329 = fmul float %314, %328
   %330 = fmul float %316, %328
   %331 = fmul float %318, %328
@@ -377,7 +377,7 @@ ENDIF136:                                         ; preds = %main_body, %ENDIF15
   %358 = insertelement <4 x float> %357, float %45, i32 2
   %359 = insertelement <4 x float> %358, float 0.000000e+00, i32 3
   %360 = call float @llvm.AMDGPU.dp4(<4 x float> %355, <4 x float> %359)
-  %361 = call float @llvm.AMDGPU.rsq(float %360)
+  %361 = call float @llvm.AMDGPU.rsq.f32(float %360)
   %362 = fmul float %45, %361
   %363 = call float @fabs(float %362)
   %364 = fmul float %176, 0x3FECCCCCC0000000
@@ -403,7 +403,7 @@ ENDIF136:                                         ; preds = %main_body, %ENDIF15
   %384 = insertelement <4 x float> %383, float %45, i32 2
   %385 = insertelement <4 x float> %384, float 0.000000e+00, i32 3
   %386 = call float @llvm.AMDGPU.dp4(<4 x float> %381, <4 x float> %385)
-  %387 = call float @llvm.AMDGPU.rsq(float %386)
+  %387 = call float @llvm.AMDGPU.rsq.f32(float %386)
   %388 = fmul float %45, %387
   %389 = call float @fabs(float %388)
   %390 = fmul float %176, 0x3FF51EB860000000
@@ -1041,7 +1041,7 @@ IF179:                                            ; preds = %ENDIF175
   %896 = insertelement <4 x float> %895, float %45, i32 2
   %897 = insertelement <4 x float> %896, float 0.000000e+00, i32 3
   %898 = call float @llvm.AMDGPU.dp4(<4 x float> %893, <4 x float> %897)
-  %899 = call float @llvm.AMDGPU.rsq(float %898)
+  %899 = call float @llvm.AMDGPU.rsq.f32(float %898)
   %900 = fmul float %45, %899
   %901 = call float @fabs(float %900)
   %902 = fmul float %176, 0x3FECCCCCC0000000
@@ -1150,7 +1150,7 @@ ENDIF178:                                         ; preds = %ENDIF175, %IF179
 declare float @llvm.AMDGPU.dp4(<4 x float>, <4 x float>) #1
 
 ; Function Attrs: readnone
-declare float @llvm.AMDGPU.rsq(float) #1
+declare float @llvm.AMDGPU.rsq.f32(float) #1
 
 ; Function Attrs: readnone
 declare <4 x float> @llvm.AMDGPU.tex(<4 x float>, i32, i32, i32) #1
diff --git a/test/CodeGen/R600/llvm.AMDGPU.div_fixup.ll b/test/CodeGen/R600/llvm.AMDGPU.div_fixup.ll
new file mode 100644
index 0000000000..c8c73573e0
--- /dev/null
+++ b/test/CodeGen/R600/llvm.AMDGPU.div_fixup.ll
@@ -0,0 +1,27 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
+
+declare float @llvm.AMDGPU.div.fixup.f32(float, float, float) nounwind readnone
+declare double @llvm.AMDGPU.div.fixup.f64(double, double, double) nounwind readnone
+
+; SI-LABEL: @test_div_fixup_f32:
+; SI-DAG: S_LOAD_DWORD [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI-DAG: S_LOAD_DWORD [[SC:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd
+; SI-DAG: V_MOV_B32_e32 [[VC:v[0-9]+]], [[SC]]
+; SI-DAG: S_LOAD_DWORD [[SB:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xc
+; SI: V_MOV_B32_e32 [[VB:v[0-9]+]], [[SB]]
+; SI: V_DIV_FIXUP_F32 [[RESULT:v[0-9]+]], [[SA]], [[VB]], [[VC]]
+; SI: BUFFER_STORE_DWORD [[RESULT]],
+; SI: S_ENDPGM
+define void @test_div_fixup_f32(float addrspace(1)* %out, float %a, float %b, float %c) nounwind {
+  %result = call float @llvm.AMDGPU.div.fixup.f32(float %a, float %b, float %c) nounwind readnone
+  store float %result, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @test_div_fixup_f64:
+; SI: V_DIV_FIXUP_F64
+define void @test_div_fixup_f64(double addrspace(1)* %out, double %a, double %b, double %c) nounwind {
+  %result = call double @llvm.AMDGPU.div.fixup.f64(double %a, double %b, double %c) nounwind readnone
+  store double %result, double addrspace(1)* %out, align 8
+  ret void
+}
diff --git a/test/CodeGen/R600/llvm.AMDGPU.div_fmas.ll b/test/CodeGen/R600/llvm.AMDGPU.div_fmas.ll
new file mode 100644
index 0000000000..4f1e827c2c
--- /dev/null
+++ b/test/CodeGen/R600/llvm.AMDGPU.div_fmas.ll
@@ -0,0 +1,27 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
+
+declare float @llvm.AMDGPU.div.fmas.f32(float, float, float) nounwind readnone
+declare double @llvm.AMDGPU.div.fmas.f64(double, double, double) nounwind readnone
+
+; SI-LABEL: @test_div_fmas_f32:
+; SI-DAG: S_LOAD_DWORD [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI-DAG: S_LOAD_DWORD [[SC:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd
+; SI-DAG: V_MOV_B32_e32 [[VC:v[0-9]+]], [[SC]]
+; SI-DAG: S_LOAD_DWORD [[SB:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xc
+; SI: V_MOV_B32_e32 [[VB:v[0-9]+]], [[SB]]
+; SI: V_DIV_FMAS_F32 [[RESULT:v[0-9]+]], [[SA]], [[VB]], [[VC]]
+; SI: BUFFER_STORE_DWORD [[RESULT]],
+; SI: S_ENDPGM
+define void @test_div_fmas_f32(float addrspace(1)* %out, float %a, float %b, float %c) nounwind {
+  %result = call float @llvm.AMDGPU.div.fmas.f32(float %a, float %b, float %c) nounwind readnone
+  store float %result, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: @test_div_fmas_f64:
+; SI: V_DIV_FMAS_F64
+define void @test_div_fmas_f64(double addrspace(1)* %out, double %a, double %b, double %c) nounwind {
+  %result = call double @llvm.AMDGPU.div.fmas.f64(double %a, double %b, double %c) nounwind readnone
+  store double %result, double addrspace(1)* %out, align 8
+  ret void
+}
diff --git a/test/CodeGen/R600/llvm.AMDGPU.div_scale.ll b/test/CodeGen/R600/llvm.AMDGPU.div_scale.ll
new file mode 100644
index 0000000000..1bcbe2f859
--- /dev/null
+++ b/test/CodeGen/R600/llvm.AMDGPU.div_scale.ll
@@ -0,0 +1,23 @@
+; XFAIL: *
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
+
+declare float @llvm.AMDGPU.div.scale.f32(float, float) nounwind readnone
+declare double @llvm.AMDGPU.div.scale.f64(double, double) nounwind readnone
+
+; SI-LABEL @test_div_scale_f32:
+define void @test_div_scale_f32(float addrspace(1)* %out, float addrspace(1)* %aptr, float addrspace(1)* %bptr) nounwind {
+  %a = load float addrspace(1)* %aptr, align 4
+  %b = load float addrspace(1)* %bptr, align 4
+  %result = call float @llvm.AMDGPU.div.scale.f32(float %a, float %b) nounwind readnone
+  store float %result, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL @test_div_scale_f64:
+define void @test_div_scale_f64(double addrspace(1)* %out, double addrspace(1)* %aptr, double addrspace(1)* %bptr) nounwind {
+  %a = load double addrspace(1)* %aptr, align 8
+  %b = load double addrspace(1)* %bptr, align 8
+  %result = call double @llvm.AMDGPU.div.scale.f64(double %a, double %b) nounwind readnone
+  store double %result, double addrspace(1)* %out, align 8
+  ret void
+}
diff --git a/test/CodeGen/R600/llvm.AMDGPU.rcp.ll b/test/CodeGen/R600/llvm.AMDGPU.rcp.ll
new file mode 100644
index 0000000000..ca5260dc5b
--- /dev/null
+++ b/test/CodeGen/R600/llvm.AMDGPU.rcp.ll
@@ -0,0 +1,58 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI -check-prefix=FUNC %s
+
+declare float @llvm.AMDGPU.rcp.f32(float) nounwind readnone
+declare double @llvm.AMDGPU.rcp.f64(double) nounwind readnone
+
+
+declare float @llvm.sqrt.f32(float) nounwind readnone
+declare double @llvm.sqrt.f64(double) nounwind readnone
+
+; FUNC-LABEL: @rcp_f32
+; SI: V_RCP_F32_e32
+define void @rcp_f32(float addrspace(1)* %out, float %src) nounwind {
+  %rcp = call float @llvm.AMDGPU.rcp.f32(float %src) nounwind readnone
+  store float %rcp, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @rcp_f64
+; SI: V_RCP_F64_e32
+define void @rcp_f64(double addrspace(1)* %out, double %src) nounwind {
+  %rcp = call double @llvm.AMDGPU.rcp.f64(double %src) nounwind readnone
+  store double %rcp, double addrspace(1)* %out, align 8
+  ret void
+}
+
+; FUNC-LABEL: @rcp_pat_f32
+; SI: V_RCP_F32_e32
+define void @rcp_pat_f32(float addrspace(1)* %out, float %src) nounwind {
+  %rcp = fdiv float 1.0, %src
+  store float %rcp, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @rcp_pat_f64
+; SI: V_RCP_F64_e32
+define void @rcp_pat_f64(double addrspace(1)* %out, double %src) nounwind {
+  %rcp = fdiv double 1.0, %src
+  store double %rcp, double addrspace(1)* %out, align 8
+  ret void
+}
+
+; FUNC-LABEL: @rsq_rcp_pat_f32
+; SI: V_RSQ_F32_e32
+define void @rsq_rcp_pat_f32(float addrspace(1)* %out, float %src) nounwind {
+  %sqrt = call float @llvm.sqrt.f32(float %src) nounwind readnone
+  %rcp = call float @llvm.AMDGPU.rcp.f32(float %sqrt) nounwind readnone
+  store float %rcp, float addrspace(1)* %out, align 4
+  ret void
+}
+
+; FUNC-LABEL: @rsq_rcp_pat_f64
+; SI: V_RSQ_F64_e32
+define void @rsq_rcp_pat_f64(double addrspace(1)* %out, double %src) nounwind {
+  %sqrt = call double @llvm.sqrt.f64(double %src) nounwind readnone
+  %rcp = call double @llvm.AMDGPU.rcp.f64(double %sqrt) nounwind readnone
+  store double %rcp, double addrspace(1)* %out, align 8
+  ret void
+}
diff --git a/test/CodeGen/R600/llvm.AMDGPU.trig_preop.ll b/test/CodeGen/R600/llvm.AMDGPU.trig_preop.ll
new file mode 100644
index 0000000000..1c736d447e
--- /dev/null
+++ b/test/CodeGen/R600/llvm.AMDGPU.trig_preop.ll
@@ -0,0 +1,29 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
+
+declare double @llvm.AMDGPU.trig.preop.f64(double, i32) nounwind readnone
+
+; SI-LABEL: @test_trig_preop_f64:
+; SI-DAG: BUFFER_LOAD_DWORD [[SEG:v[0-9]+]]
+; SI-DAG: BUFFER_LOAD_DWORDX2 [[SRC:v\[[0-9]+:[0-9]+\]]],
+; SI: V_TRIG_PREOP_F64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[SRC]], [[SEG]]
+; SI: BUFFER_STORE_DWORDX2 [[RESULT]],
+; SI: S_ENDPGM
+define void @test_trig_preop_f64(double addrspace(1)* %out, double addrspace(1)* %aptr, i32 addrspace(1)* %bptr) nounwind {
+  %a = load double addrspace(1)* %aptr, align 8
+  %b = load i32 addrspace(1)* %bptr, align 4
+  %result = call double @llvm.AMDGPU.trig.preop.f64(double %a, i32 %b) nounwind readnone
+  store double %result, double addrspace(1)* %out, align 8
+  ret void
+}
+
+; SI-LABEL: @test_trig_preop_f64_imm_segment:
+; SI: BUFFER_LOAD_DWORDX2 [[SRC:v\[[0-9]+:[0-9]+\]]],
+; SI: V_TRIG_PREOP_F64 [[RESULT:v\[[0-9]+:[0-9]+\]]], [[SRC]], 7
+; SI: BUFFER_STORE_DWORDX2 [[RESULT]],
+; SI: S_ENDPGM
+define void @test_trig_preop_f64_imm_segment(double addrspace(1)* %out, double addrspace(1)* %aptr) nounwind {
+  %a = load double addrspace(1)* %aptr, align 8
+  %result = call double @llvm.AMDGPU.trig.preop.f64(double %a, i32 7) nounwind readnone
+  store double %result, double addrspace(1)* %out, align 8
+  ret void
+}
diff --git a/test/CodeGen/R600/pv.ll b/test/CodeGen/R600/pv.ll
index f322bc71c6..55eb56d3fb 100644
--- a/test/CodeGen/R600/pv.ll
+++ b/test/CodeGen/R600/pv.ll
@@ -103,7 +103,7 @@ main_body:
   %95 = insertelement <4 x float> %94, float 0.000000e+00, i32 3
   %96 = call float @llvm.AMDGPU.dp4(<4 x float> %91, <4 x float> %95)
   %97 = call float @fabs(float %96)
-  %98 = call float @llvm.AMDGPU.rsq(float %97)
+  %98 = call float @llvm.AMDGPU.rsq.f32(float %97)
   %99 = fmul float %4, %98
   %100 = fmul float %5, %98
   %101 = fmul float %6, %98
@@ -225,7 +225,7 @@ declare float @llvm.AMDGPU.dp4(<4 x float>, <4 x float>) #1
 declare float @fabs(float) #2
 
 ; Function Attrs: readnone
-declare float @llvm.AMDGPU.rsq(float) #1
+declare float @llvm.AMDGPU.rsq.f32(float) #1
 
 ; Function Attrs: readnone
 declare float @llvm.AMDIL.clamp.(float, float, float) #1
diff --git a/test/CodeGen/R600/sgpr-copy.ll b/test/CodeGen/R600/sgpr-copy.ll
index c581d86b99..c7d5bf9064 100644
--- a/test/CodeGen/R600/sgpr-copy.ll
+++ b/test/CodeGen/R600/sgpr-copy.ll
@@ -70,7 +70,7 @@ main_body:
   %55 = fadd float %54, %53
   %56 = fmul float %45, %45
   %57 = fadd float %55, %56
-  %58 = call float @llvm.AMDGPU.rsq(float %57)
+  %58 = call float @llvm.AMDGPU.rsq.f32(float %57)
   %59 = fmul float %43, %58
   %60 = fmul float %44, %58
   %61 = fmul float %45, %58
@@ -212,7 +212,7 @@ declare float @llvm.SI.fs.interp(i32, i32, i32, <2 x i32>) #1
 declare <4 x float> @llvm.SI.sample.v2i32(<2 x i32>, <32 x i8>, <16 x i8>, i32) #1
 
 ; Function Attrs: readnone
-declare float @llvm.AMDGPU.rsq(float) #3
+declare float @llvm.AMDGPU.rsq.f32(float) #3
 
 ; Function Attrs: readnone
 declare float @llvm.AMDIL.exp.(float) #3
diff --git a/test/CodeGen/R600/si-sgpr-spill.ll b/test/CodeGen/R600/si-sgpr-spill.ll
index b34a757d9b..53a096513b 100644
--- a/test/CodeGen/R600/si-sgpr-spill.ll
+++ b/test/CodeGen/R600/si-sgpr-spill.ll
@@ -203,7 +203,7 @@ main_body:
   %198 = fadd float %197, %196
   %199 = fmul float %97, %97
   %200 = fadd float %198, %199
-  %201 = call float @llvm.AMDGPU.rsq(float %200)
+  %201 = call float @llvm.AMDGPU.rsq.f32(float %200)
   %202 = fmul float %95, %201
   %203 = fmul float %96, %201
   %204 = fmul float %202, %29
@@ -384,7 +384,7 @@ IF67:                                             ; preds = %LOOP65
   %355 = fadd float %354, %353
   %356 = fmul float %352, %352
   %357 = fadd float %355, %356
-  %358 = call float @llvm.AMDGPU.rsq(float %357)
+  %358 = call float @llvm.AMDGPU.rsq.f32(float %357)
   %359 = fmul float %350, %358
   %360 = fmul float %351, %358
   %361 = fmul float %352, %358
@@ -512,7 +512,7 @@ IF67:                                             ; preds = %LOOP65
   %483 = fadd float %482, %481
   %484 = fmul float %109, %109
   %485 = fadd float %483, %484
-  %486 = call float @llvm.AMDGPU.rsq(float %485)
+  %486 = call float @llvm.AMDGPU.rsq.f32(float %485)
   %487 = fmul float %107, %486
   %488 = fmul float %108, %486
   %489 = fmul float %109, %486
@@ -541,7 +541,7 @@ IF67:                                             ; preds = %LOOP65
   %512 = fadd float %511, %510
   %513 = fmul float %97, %97
   %514 = fadd float %512, %513
-  %515 = call float @llvm.AMDGPU.rsq(float %514)
+  %515 = call float @llvm.AMDGPU.rsq.f32(float %514)
   %516 = fmul float %95, %515
   %517 = fmul float %96, %515
   %518 = fmul float %97, %515
@@ -658,7 +658,7 @@ declare i32 @llvm.SI.tid() #2
 declare float @ceil(float) #3
 
 ; Function Attrs: readnone
-declare float @llvm.AMDGPU.rsq(float) #2
+declare float @llvm.AMDGPU.rsq.f32(float) #2
 
 ; Function Attrs: nounwind readnone
 declare <4 x float> @llvm.SI.sampled.v8i32(<8 x i32>, <32 x i8>, <16 x i8>, i32) #1
@@ -887,7 +887,7 @@ main_body:
   %212 = fadd float %211, %210
   %213 = fmul float %209, %209
   %214 = fadd float %212, %213
-  %215 = call float @llvm.AMDGPU.rsq(float %214)
+  %215 = call float @llvm.AMDGPU.rsq.f32(float %214)
   %216 = fmul float %205, %215
   %217 = fmul float %207, %215
   %218 = fmul float %209, %215
@@ -1123,7 +1123,7 @@ IF189:                                            ; preds = %LOOP
   %434 = fsub float -0.000000e+00, %433
   %435 = fadd float 0x3FF00068E0000000, %434
   %436 = call float @llvm.AMDIL.clamp.(float %435, float 0.000000e+00, float 1.000000e+00)
-  %437 = call float @llvm.AMDGPU.rsq(float %436)
+  %437 = call float @llvm.AMDGPU.rsq.f32(float %436)
   %438 = fmul float %437, %436
   %439 = fsub float -0.000000e+00, %436
   %440 = call float @llvm.AMDGPU.cndlt(float %439, float %438, float 0.000000e+00)
@@ -1147,7 +1147,7 @@ IF189:                                            ; preds = %LOOP
   %458 = fadd float %457, %456
   %459 = fmul float %455, %455
   %460 = fadd float %458, %459
-  %461 = call float @llvm.AMDGPU.rsq(float %460)
+  %461 = call float @llvm.AMDGPU.rsq.f32(float %460)
   %462 = fmul float %451, %461
   %463 = fmul float %453, %461
   %464 = fmul float %455, %461
@@ -1257,7 +1257,7 @@ ENDIF197:                                         ; preds = %IF189, %IF198
   %559 = fadd float %558, %557
   %560 = fmul float %556, %556
   %561 = fadd float %559, %560
-  %562 = call float @llvm.AMDGPU.rsq(float %561)
+  %562 = call float @llvm.AMDGPU.rsq.f32(float %561)
   %563 = fmul float %562, %561
   %564 = fsub float -0.000000e+00, %561
   %565 = call float @llvm.AMDGPU.cndlt(float %564, float %563, float 0.000000e+00)
diff --git a/test/Transforms/InstCombine/r600-intrinsics.ll b/test/Transforms/InstCombine/r600-intrinsics.ll
new file mode 100644
index 0000000000..1db6b0d28b
--- /dev/null
+++ b/test/Transforms/InstCombine/r600-intrinsics.ll
@@ -0,0 +1,47 @@
+; RUN: opt -instcombine -S < %s | FileCheck %s
+
+declare float @llvm.AMDGPU.rcp.f32(float) nounwind readnone
+declare double @llvm.AMDGPU.rcp.f64(double) nounwind readnone
+
+; CHECK-LABEL: @test_constant_fold_rcp_f32_1
+; CHECK-NEXT: ret float 1.000000e+00
+define float @test_constant_fold_rcp_f32_1() nounwind {
+  %val = call float @llvm.AMDGPU.rcp.f32(float 1.0) nounwind readnone
+  ret float %val
+}
+
+; CHECK-LABEL: @test_constant_fold_rcp_f64_1
+; CHECK-NEXT:  ret double 1.000000e+00
+define double @test_constant_fold_rcp_f64_1() nounwind {
+  %val = call double @llvm.AMDGPU.rcp.f64(double 1.0) nounwind readnone
+  ret double %val
+}
+
+; CHECK-LABEL: @test_constant_fold_rcp_f32_half
+; CHECK-NEXT: ret float 2.000000e+00
+define float @test_constant_fold_rcp_f32_half() nounwind {
+  %val = call float @llvm.AMDGPU.rcp.f32(float 0.5) nounwind readnone
+  ret float %val
+}
+
+; CHECK-LABEL: @test_constant_fold_rcp_f64_half
+; CHECK-NEXT:  ret double 2.000000e+00
+define double @test_constant_fold_rcp_f64_half() nounwind {
+  %val = call double @llvm.AMDGPU.rcp.f64(double 0.5) nounwind readnone
+  ret double %val
+}
+
+; CHECK-LABEL: @test_constant_fold_rcp_f32_43
+; CHECK-NEXT: call float @llvm.AMDGPU.rcp.f32(float 4.300000e+01)
+define float @test_constant_fold_rcp_f32_43() nounwind {
+ %val = call float @llvm.AMDGPU.rcp.f32(float 4.300000e+01) nounwind readnone
+ ret float %val
+}
+
+; CHECK-LABEL: @test_constant_fold_rcp_f64_43
+; CHECK-NEXT: call double @llvm.AMDGPU.rcp.f64(double 4.300000e+01)
+define double @test_constant_fold_rcp_f64_43() nounwind {
+  %val = call double @llvm.AMDGPU.rcp.f64(double 4.300000e+01) nounwind readnone
+  ret double %val
+}
+