R600/SI: Add class intrinsic

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@225305 91177308-0d34-0410-b5e6-96231b3b80d8

include/llvm/IR/IntrinsicsR600.td

@@ -76,6 +76,9 @@ def int_AMDGPU_rsq_clamped : GCCBuiltin<"__builtin_amdgpu_rsq_clamped">,
 def int_AMDGPU_ldexp : GCCBuiltin<"__builtin_amdgpu_ldexp">,
   Intrinsic<[llvm_anyfloat_ty], [LLVMMatchType<0>, llvm_i32_ty], [IntrNoMem]>;
 
+def int_AMDGPU_class : GCCBuiltin<"__builtin_amdgpu_class">,
+  Intrinsic<[llvm_i1_ty], [llvm_anyfloat_ty, llvm_i32_ty], [IntrNoMem]>;
+
 def int_AMDGPU_read_workdim : AMDGPUReadPreloadRegisterIntrinsic <
                                        "__builtin_amdgpu_read_workdim">;
 

lib/Target/R600/AMDGPUISelDAGToDAG.cpp

@@ -113,6 +113,9 @@ private:
 
   bool SelectVOP3Mods0Clamp(SDValue In, SDValue &Src, SDValue &SrcMods,
                             SDValue &Omod) const;
+  bool SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src, SDValue &SrcMods,
+                                 SDValue &Clamp,
+                                 SDValue &Omod) const;
 
   SDNode *SelectADD_SUB_I64(SDNode *N);
   SDNode *SelectDIV_SCALE(SDNode *N);
@@ -1143,6 +1146,14 @@ bool AMDGPUDAGToDAGISel::SelectVOP3Mods0Clamp(SDValue In, SDValue &Src,
   return SelectVOP3Mods(In, Src, SrcMods);
 }
 
+bool AMDGPUDAGToDAGISel::SelectVOP3Mods0Clamp0OMod(SDValue In, SDValue &Src,
+                                                   SDValue &SrcMods,
+                                                   SDValue &Clamp,
+                                                   SDValue &Omod) const {
+  Clamp = Omod = CurDAG->getTargetConstant(0, MVT::i32);
+  return SelectVOP3Mods(In, Src, SrcMods);
+}
+
 void AMDGPUDAGToDAGISel::PostprocessISelDAG() {
   const AMDGPUTargetLowering& Lowering =
     *static_cast<const AMDGPUTargetLowering*>(getTargetLowering());

lib/Target/R600/AMDGPUISelLowering.cpp

@@ -1000,6 +1000,10 @@ SDValue AMDGPUTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
     case AMDGPUIntrinsic::AMDGPU_brev:
       return DAG.getNode(AMDGPUISD::BREV, DL, VT, Op.getOperand(1));
 
+  case Intrinsic::AMDGPU_class:
+    return DAG.getNode(AMDGPUISD::FP_CLASS, DL, VT,
+                       Op.getOperand(1), Op.getOperand(2));
+
     case AMDGPUIntrinsic::AMDIL_exp: // Legacy name.
       return DAG.getNode(ISD::FEXP2, DL, VT, Op.getOperand(1));
 
@@ -2516,6 +2520,7 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
   NODE_NAME_CASE(RSQ_LEGACY)
   NODE_NAME_CASE(RSQ_CLAMPED)
   NODE_NAME_CASE(LDEXP)
+  NODE_NAME_CASE(FP_CLASS)
   NODE_NAME_CASE(DOT4)
   NODE_NAME_CASE(BFE_U32)
   NODE_NAME_CASE(BFE_I32)

lib/Target/R600/AMDGPUISelLowering.h

@@ -234,6 +234,7 @@ enum {
   RSQ_LEGACY,
   RSQ_CLAMPED,
   LDEXP,
+  FP_CLASS,
   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.

lib/Target/R600/AMDGPUInstrInfo.td

@@ -27,6 +27,10 @@ def AMDGPULdExpOp : SDTypeProfile<1, 2,
   [SDTCisSameAs<0, 1>, SDTCisFP<0>, SDTCisInt<2>]
 >;
 
+def AMDGPUFPClassOp : SDTypeProfile<1, 2,
+  [SDTCisInt<0>, SDTCisFP<1>, SDTCisInt<2>]
+>;
+
 def AMDGPUDivScaleOp : SDTypeProfile<2, 3,
   [SDTCisFP<0>, SDTCisInt<1>, SDTCisSameAs<0, 2>, SDTCisSameAs<0, 3>, SDTCisSameAs<0, 4>]
 >;
@@ -58,6 +62,8 @@ def AMDGPUrsq_clamped : SDNode<"AMDGPUISD::RSQ_CLAMPED", SDTFPUnaryOp>;
 
 def AMDGPUldexp : SDNode<"AMDGPUISD::LDEXP", AMDGPULdExpOp>;
 
+def AMDGPUfp_class : SDNode<"AMDGPUISD::FP_CLASS", AMDGPUFPClassOp>;
+
 // out = max(a, b) a and b are floats, where a nan comparison fails.
 // This is not commutative because this gives the second operand:
 //   x < nan ? x : nan -> nan

lib/Target/R600/SIDefines.h

@@ -44,6 +44,21 @@ namespace SIInstrFlags {
     EXP_CNT = 1 << 1,
     LGKM_CNT = 1 << 2
   };
+
+  // v_cmp_class_* etc. use a 10-bit mask for what operation is checked.
+  // The result is true if any of these tests are true.
+  enum ClassFlags {
+    S_NAN = 1 << 0,        // Signaling NaN
+    Q_NAN = 1 << 1,        // Quiet NaN
+    N_INFINITY = 1 << 2,   // Negative infinity
+    N_NORMAL = 1 << 3,     // Negative normal
+    N_SUBNORMAL = 1 << 4,  // Negative subnormal
+    N_ZERO = 1 << 5,       // Negative zero
+    P_ZERO = 1 << 6,       // Positive zero
+    P_SUBNORMAL = 1 << 7,  // Positive subnormal
+    P_NORMAL = 1 << 8,     // Positive normal
+    P_INFINITY = 1 << 9    // Positive infinity
+  };
 }
 
 namespace SISrcMods {

lib/Target/R600/SIInstrInfo.td

@@ -283,6 +283,7 @@ def MUBUFOffsetAtomic : ComplexPattern<i64, 4, "SelectMUBUFOffset">;
 
 def VOP3Mods0 : ComplexPattern<untyped, 4, "SelectVOP3Mods0">;
 def VOP3Mods0Clamp : ComplexPattern<untyped, 3, "SelectVOP3Mods0Clamp">;
+def VOP3Mods0Clamp0OMod : ComplexPattern<untyped, 4, "SelectVOP3Mods0Clamp0OMod">;
 def VOP3Mods  : ComplexPattern<untyped, 2, "SelectVOP3Mods">;
 
 //===----------------------------------------------------------------------===//
@@ -611,7 +612,9 @@ class getNumSrcArgs<ValueType Src1, ValueType Src2> {
 // Returns the register class to use for the destination of VOP[123C]
 // instructions for the given VT.
 class getVALUDstForVT<ValueType VT> {
-  RegisterClass ret = !if(!eq(VT.Size, 32), VReg_32, VReg_64);
+  RegisterClass ret = !if(!eq(VT.Size, 32), VReg_32,
+                          !if(!eq(VT.Size, 64), VReg_64,
+                            SReg_64)); // else VT == i1
 }
 
 // Returns the register class to use for source 0 of VOP[12C]
@@ -775,6 +778,17 @@ def VOP_I32_I32_I32 : VOPProfile <[i32, i32, i32, untyped]>;
 def VOP_I32_I32_I32_VCC : VOPProfile <[i32, i32, i32, untyped]> {
   let Src0RC32 = VCSrc_32;
 }
+
+def VOP_I1_F32_I32 : VOPProfile <[i1, f32, i32, untyped]> {
+  let Ins64 = (ins InputModsNoDefault:$src0_modifiers, Src0RC64:$src0, Src1RC64:$src1);
+  let Asm64 = " $dst, $src0_modifiers, $src1";
+}
+
+def VOP_I1_F64_I32 : VOPProfile <[i1, f64, i32, untyped]> {
+  let Ins64 = (ins InputModsNoDefault:$src0_modifiers, Src0RC64:$src0, Src1RC64:$src1);
+  let Asm64 = " $dst, $src0_modifiers, $src1";
+}
+
 def VOP_I64_I64_I32 : VOPProfile <[i64, i64, i32, untyped]>;
 def VOP_I64_I64_I64 : VOPProfile <[i64, i64, i64, untyped]>;
 
@@ -1111,6 +1125,19 @@ multiclass VOPCInst <vopc op, string opName,
   P.HasModifiers, DefExec
 >;
 
+multiclass VOPCClassInst <vopc op, string opName, VOPProfile P,
+                     bit DefExec = 0> : VOPC_Helper <
+  op, opName,
+  P.Ins32, P.Asm32, [],
+  (outs SReg_64:$dst), P.Ins64, P.Asm64,
+  !if(P.HasModifiers,
+      [(set i1:$dst,
+          (AMDGPUfp_class (P.Src0VT (VOP3Mods0Clamp0OMod P.Src0VT:$src0, i32:$src0_modifiers)), P.Src1VT:$src1))],
+      [(set i1:$dst, (AMDGPUfp_class P.Src0VT:$src0, P.Src1VT:$src1))]),
+  P.HasModifiers, DefExec
+>;
+
+
 multiclass VOPC_F32 <vopc op, string opName, PatLeaf cond = COND_NULL> :
   VOPCInst <op, opName, VOP_F32_F32_F32, cond>;
 
@@ -1145,6 +1172,18 @@ multiclass VOP3_Helper <vop3 op, string opName, dag outs, dag ins, string asm,
     op, outs, ins, opName#asm, pat, opName, NumSrcArgs, HasMods
 >;
 
+multiclass VOPC_CLASS_F32 <vopc op, string opName> :
+  VOPCClassInst <op, opName, VOP_I1_F32_I32, 0>;
+
+multiclass VOPCX_CLASS_F32 <vopc op, string opName> :
+  VOPCClassInst <op, opName, VOP_I1_F32_I32, 1>;
+
+multiclass VOPC_CLASS_F64 <vopc op, string opName> :
+  VOPCClassInst <op, opName, VOP_I1_F64_I32, 0>;
+
+multiclass VOPCX_CLASS_F64 <vopc op, string opName> :
+  VOPCClassInst <op, opName, VOP_I1_F64_I32, 1>;
+
 multiclass VOP3Inst <vop3 op, string opName, VOPProfile P,
                      SDPatternOperator node = null_frag> : VOP3_Helper <
   op, opName, P.Outs, P.Ins64, P.Asm64,

lib/Target/R600/SIInstructions.td

@@ -745,16 +745,16 @@ defm V_CMPX_T_U64 : VOPCX_I64 <vopc<0xf7, 0xff>, "v_cmpx_t_u64">;
 
 } // End hasSideEffects = 1
 
-defm V_CMP_CLASS_F32 : VOPC_F32 <vopc<0x88, 0x10>, "v_cmp_class_f32">;
+defm V_CMP_CLASS_F32 : VOPC_CLASS_F32 <vopc<0x88, 0x10>, "v_cmp_class_f32">;
 
 let hasSideEffects = 1 in {
-defm V_CMPX_CLASS_F32 : VOPCX_F32 <vopc<0x98, 0x11>, "v_cmpx_class_f32">;
+defm V_CMPX_CLASS_F32 : VOPCX_CLASS_F32 <vopc<0x98, 0x11>, "v_cmpx_class_f32">;
 } // End hasSideEffects = 1
 
-defm V_CMP_CLASS_F64 : VOPC_F64 <vopc<0xa8, 0x12>, "v_cmp_class_f64">;
+defm V_CMP_CLASS_F64 : VOPC_CLASS_F64 <vopc<0xa8, 0x12>, "v_cmp_class_f64">;
 
 let hasSideEffects = 1 in {
-defm V_CMPX_CLASS_F64 : VOPCX_F64 <vopc<0xb8, 0x13>, "v_cmpx_class_f64">;
+defm V_CMPX_CLASS_F64 : VOPCX_CLASS_F64 <vopc<0xb8, 0x13>, "v_cmpx_class_f64">;
 } // End hasSideEffects = 1
 
 } // End isCompare = 1

test/CodeGen/R600/llvm.AMDGPU.class.ll

@@ -0,0 +1,335 @@
+; RUN: llc -march=r600 -mcpu=SI -verify-machineinstrs < %s | FileCheck -check-prefix=SI %s
+
+declare i1 @llvm.AMDGPU.class.f32(float, i32) #1
+declare i1 @llvm.AMDGPU.class.f64(double, i32) #1
+declare i32 @llvm.r600.read.tidig.x() #1
+declare float @llvm.fabs.f32(float) #1
+declare double @llvm.fabs.f64(double) #1
+
+; SI-LABEL: {{^}}test_class_f32:
+; SI-DAG: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; 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_cmp_class_f32_e32 vcc, [[SA]], [[VB]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, vcc
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_f32(i32 addrspace(1)* %out, float %a, i32 %b) #0 {
+  %result = call i1 @llvm.AMDGPU.class.f32(float %a, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_fabs_f32:
+; SI-DAG: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; 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_cmp_class_f32_e64 [[CMP:s\[[0-9]+:[0-9]+\]]], |[[SA]]|, [[VB]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, [[CMP]]
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_fabs_f32(i32 addrspace(1)* %out, float %a, i32 %b) #0 {
+  %a.fabs = call float @llvm.fabs.f32(float %a) #1
+  %result = call i1 @llvm.AMDGPU.class.f32(float %a.fabs, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_fneg_f32:
+; SI-DAG: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; 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_cmp_class_f32_e64 [[CMP:s\[[0-9]+:[0-9]+\]]], -[[SA]], [[VB]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, [[CMP]]
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_fneg_f32(i32 addrspace(1)* %out, float %a, i32 %b) #0 {
+  %a.fneg = fsub float -0.0, %a
+  %result = call i1 @llvm.AMDGPU.class.f32(float %a.fneg, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_fneg_fabs_f32:
+; SI-DAG: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; 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_cmp_class_f32_e64 [[CMP:s\[[0-9]+:[0-9]+\]]], -|[[SA]]|, [[VB]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, [[CMP]]
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_fneg_fabs_f32(i32 addrspace(1)* %out, float %a, i32 %b) #0 {
+  %a.fabs = call float @llvm.fabs.f32(float %a) #1
+  %a.fneg.fabs = fsub float -0.0, %a.fabs
+  %result = call i1 @llvm.AMDGPU.class.f32(float %a.fneg.fabs, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_1_f32:
+; SI: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI: v_cmp_class_f32_e64 [[COND:s\[[0-9]+:[0-9]+\]]], [[SA]], 1{{$}}
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, [[COND]]
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_1_f32(i32 addrspace(1)* %out, float %a) #0 {
+  %result = call i1 @llvm.AMDGPU.class.f32(float %a, i32 1) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_64_f32:
+; SI: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI: v_cmp_class_f32_e64 [[COND:s\[[0-9]+:[0-9]+\]]], [[SA]], 64{{$}}
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, [[COND]]
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_64_f32(i32 addrspace(1)* %out, float %a) #0 {
+  %result = call i1 @llvm.AMDGPU.class.f32(float %a, i32 64) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; Set all 10 bits of mask
+; SI-LABEL: {{^}}test_class_full_mask_f32:
+; SI: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI: v_mov_b32_e32 [[MASK:v[0-9]+]], 0x3ff{{$}}
+; SI: v_cmp_class_f32_e32 vcc, [[SA]], [[MASK]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, vcc
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_full_mask_f32(i32 addrspace(1)* %out, float %a) #0 {
+  %result = call i1 @llvm.AMDGPU.class.f32(float %a, i32 1023) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_9bit_mask_f32:
+; SI: s_load_dword [[SA:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI: v_mov_b32_e32 [[MASK:v[0-9]+]], 0x1ff{{$}}
+; SI: v_cmp_class_f32_e32 vcc, [[SA]], [[MASK]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, vcc
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_9bit_mask_f32(i32 addrspace(1)* %out, float %a) #0 {
+  %result = call i1 @llvm.AMDGPU.class.f32(float %a, i32 511) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}v_test_class_full_mask_f32:
+; SI-DAG: buffer_load_dword [[VA:v[0-9]+]]
+; SI-DAG: v_mov_b32_e32 [[MASK:v[0-9]+]], 0x1ff{{$}}
+; SI: v_cmp_class_f32_e32 vcc, [[VA]], [[MASK]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, vcc
+; SI: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @v_test_class_full_mask_f32(i32 addrspace(1)* %out, float addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.in = getelementptr float addrspace(1)* %in, i32 %tid
+  %gep.out = getelementptr i32 addrspace(1)* %out, i32 %tid
+  %a = load float addrspace(1)* %gep.in
+
+  %result = call i1 @llvm.AMDGPU.class.f32(float %a, i32 511) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %gep.out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_inline_imm_constant_dynamic_mask_f32:
+; SI-DAG: buffer_load_dword [[VB:v[0-9]+]]
+; SI: v_cmp_class_f32_e32 vcc, 1.0, [[VB]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, vcc
+; SI: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_inline_imm_constant_dynamic_mask_f32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.in = getelementptr i32 addrspace(1)* %in, i32 %tid
+  %gep.out = getelementptr i32 addrspace(1)* %out, i32 %tid
+  %b = load i32 addrspace(1)* %gep.in
+
+  %result = call i1 @llvm.AMDGPU.class.f32(float 1.0, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %gep.out, align 4
+  ret void
+}
+
+; FIXME: Why isn't this using a literal constant operand?
+; SI-LABEL: {{^}}test_class_lit_constant_dynamic_mask_f32:
+; SI-DAG: buffer_load_dword [[VB:v[0-9]+]]
+; SI-DAG: v_mov_b32_e32 [[VK:v[0-9]+]], 0x44800000
+; SI: v_cmp_class_f32_e32 vcc, [[VK]], [[VB]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, vcc
+; SI: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_lit_constant_dynamic_mask_f32(i32 addrspace(1)* %out, i32 addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.in = getelementptr i32 addrspace(1)* %in, i32 %tid
+  %gep.out = getelementptr i32 addrspace(1)* %out, i32 %tid
+  %b = load i32 addrspace(1)* %gep.in
+
+  %result = call i1 @llvm.AMDGPU.class.f32(float 1024.0, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %gep.out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_f64:
+; SI-DAG: s_load_dwordx2 [[SA:s\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI-DAG: s_load_dword [[SB:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd
+; SI-DAG: v_mov_b32_e32 [[VB:v[0-9]+]], [[SB]]
+; SI: v_cmp_class_f64_e32 vcc, [[SA]], [[VB]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, vcc
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_f64(i32 addrspace(1)* %out, double %a, i32 %b) #0 {
+  %result = call i1 @llvm.AMDGPU.class.f64(double %a, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_fabs_f64:
+; SI-DAG: s_load_dwordx2 [[SA:s\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI-DAG: s_load_dword [[SB:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd
+; SI-DAG: v_mov_b32_e32 [[VB:v[0-9]+]], [[SB]]
+; SI: v_cmp_class_f64_e64 [[CMP:s\[[0-9]+:[0-9]+\]]], |[[SA]]|, [[VB]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, [[CMP]]
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_fabs_f64(i32 addrspace(1)* %out, double %a, i32 %b) #0 {
+  %a.fabs = call double @llvm.fabs.f64(double %a) #1
+  %result = call i1 @llvm.AMDGPU.class.f64(double %a.fabs, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_fneg_f64:
+; SI-DAG: s_load_dwordx2 [[SA:s\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI-DAG: s_load_dword [[SB:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd
+; SI-DAG: v_mov_b32_e32 [[VB:v[0-9]+]], [[SB]]
+; SI: v_cmp_class_f64_e64 [[CMP:s\[[0-9]+:[0-9]+\]]], -[[SA]], [[VB]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, [[CMP]]
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_fneg_f64(i32 addrspace(1)* %out, double %a, i32 %b) #0 {
+  %a.fneg = fsub double -0.0, %a
+  %result = call i1 @llvm.AMDGPU.class.f64(double %a.fneg, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_fneg_fabs_f64:
+; SI-DAG: s_load_dwordx2 [[SA:s\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI-DAG: s_load_dword [[SB:s[0-9]+]], s{{\[[0-9]+:[0-9]+\]}}, 0xd
+; SI-DAG: v_mov_b32_e32 [[VB:v[0-9]+]], [[SB]]
+; SI: v_cmp_class_f64_e64 [[CMP:s\[[0-9]+:[0-9]+\]]], -|[[SA]]|, [[VB]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, [[CMP]]
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_fneg_fabs_f64(i32 addrspace(1)* %out, double %a, i32 %b) #0 {
+  %a.fabs = call double @llvm.fabs.f64(double %a) #1
+  %a.fneg.fabs = fsub double -0.0, %a.fabs
+  %result = call i1 @llvm.AMDGPU.class.f64(double %a.fneg.fabs, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_1_f64:
+; SI: v_cmp_class_f64_e64 {{s\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 1{{$}}
+; SI: s_endpgm
+define void @test_class_1_f64(i32 addrspace(1)* %out, double %a) #0 {
+  %result = call i1 @llvm.AMDGPU.class.f64(double %a, i32 1) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_64_f64:
+; SI: v_cmp_class_f64_e64 {{s\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 64{{$}}
+; SI: s_endpgm
+define void @test_class_64_f64(i32 addrspace(1)* %out, double %a) #0 {
+  %result = call i1 @llvm.AMDGPU.class.f64(double %a, i32 64) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; Set all 9 bits of mask
+; SI-LABEL: {{^}}test_class_full_mask_f64:
+; SI: s_load_dwordx2 [[SA:s\[[0-9]+:[0-9]+\]]], s{{\[[0-9]+:[0-9]+\]}}, 0xb
+; SI: v_mov_b32_e32 [[MASK:v[0-9]+]], 0x1ff{{$}}
+; SI: v_cmp_class_f64_e32 vcc, [[SA]], [[MASK]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, vcc
+; SI-NEXT: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @test_class_full_mask_f64(i32 addrspace(1)* %out, double %a) #0 {
+  %result = call i1 @llvm.AMDGPU.class.f64(double %a, i32 511) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}v_test_class_full_mask_f64:
+; SI-DAG: buffer_load_dwordx2 [[VA:v\[[0-9]+:[0-9]+\]]]
+; SI-DAG: v_mov_b32_e32 [[MASK:v[0-9]+]], 0x1ff{{$}}
+; SI: v_cmp_class_f64_e32 vcc, [[VA]], [[MASK]]
+; SI-NEXT: v_cndmask_b32_e64 [[RESULT:v[0-9]+]], 0, -1, vcc
+; SI: buffer_store_dword [[RESULT]]
+; SI: s_endpgm
+define void @v_test_class_full_mask_f64(i32 addrspace(1)* %out, double addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.in = getelementptr double addrspace(1)* %in, i32 %tid
+  %gep.out = getelementptr i32 addrspace(1)* %out, i32 %tid
+  %a = load double addrspace(1)* %in
+
+  %result = call i1 @llvm.AMDGPU.class.f64(double %a, i32 511) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %gep.out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_inline_imm_constant_dynamic_mask_f64:
+; XSI: v_cmp_class_f64_e32 vcc, 1.0,
+; SI: v_cmp_class_f64_e32 vcc,
+; SI: s_endpgm
+define void @test_class_inline_imm_constant_dynamic_mask_f64(i32 addrspace(1)* %out, i32 addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.in = getelementptr i32 addrspace(1)* %in, i32 %tid
+  %gep.out = getelementptr i32 addrspace(1)* %out, i32 %tid
+  %b = load i32 addrspace(1)* %gep.in
+
+  %result = call i1 @llvm.AMDGPU.class.f64(double 1.0, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %gep.out, align 4
+  ret void
+}
+
+; SI-LABEL: {{^}}test_class_lit_constant_dynamic_mask_f64:
+; SI: v_cmp_class_f64_e32 vcc, s{{\[[0-9]+:[0-9]+\]}}, v{{[0-9]+}}
+; SI: s_endpgm
+define void @test_class_lit_constant_dynamic_mask_f64(i32 addrspace(1)* %out, i32 addrspace(1)* %in) #0 {
+  %tid = call i32 @llvm.r600.read.tidig.x() #1
+  %gep.in = getelementptr i32 addrspace(1)* %in, i32 %tid
+  %gep.out = getelementptr i32 addrspace(1)* %out, i32 %tid
+  %b = load i32 addrspace(1)* %gep.in
+
+  %result = call i1 @llvm.AMDGPU.class.f64(double 1024.0, i32 %b) #1
+  %sext = sext i1 %result to i32
+  store i32 %sext, i32 addrspace(1)* %gep.out, align 4
+  ret void
+}
+
+attributes #0 = { nounwind }
+attributes #1 = { nounwind readnone }