AMDGPU: Improve accuracy of instruction rates for VOPC

[arsenm]

Summary

These were all using the default 32-bit VALU write class,
but the i64/f64 compares are half rate.

I'm not sure this is really correct, because they are still using
the write to VALU write class, even though they really write
to the SALU.

[arsenm]

ping

[• tstellarAMD]

LGTM.

[arsenm]

r248582

lib/Target/AMDGPU/SIInstrInfo.td

def _si : VOP3b_Real_si <op.SI3, outs, ins, asm, opName>,
             VOP3DisableFields<1, 1, HasMods>;
 
   def _vi : VOP3b_Real_vi <op.VI3, outs, ins, asm, opName>,
             VOP3DisableFields<1, 1, HasMods>;
 }
 
 multiclass VOP3_C_m <vop op, dag outs, dag ins, string asm,
                      list<dag> pattern, string opName,
-                     bit HasMods, bit defExec, string revOp> {
+                     bit HasMods, bit defExec,
+                     string revOp, list<SchedReadWrite> sched> {
 
   def "" : VOP3_Pseudo <outs, ins, pattern, opName>,
-           VOP2_REV<revOp#"_e64", !eq(revOp, opName)>;
+           VOP2_REV<revOp#"_e64", !eq(revOp, opName)> {
+    let SchedRW = sched;
+  }
 
   def _si : VOP3_Real_si <op.SI3, outs, ins, asm, opName>,
             VOP3DisableFields<1, 0, HasMods> {
     let Defs = !if(defExec, [EXEC], []);
+    let SchedRW = sched;
   }
 
   def _vi : VOP3_Real_vi <op.VI3, outs, ins, asm, opName>,
             VOP3DisableFields<1, 0, HasMods> {
     let Defs = !if(defExec, [EXEC], []);
+    let SchedRW = sched;
   }
 }
 
 // An instruction that is VOP2 on SI and VOP3 on VI, no modifiers.
 multiclass VOP2SI_3VI_m <vop3 op, string opName, dag outs, dag ins,
                          string asm, list<dag> pattern = []> {
   let isPseudo = 1, isCodeGenOnly = 1 in {
     def "" : VOPAnyCommon <outs, ins, "", pattern>,
class VOPC_Pseudo <dag ins, list<dag> pattern, string opName> :
   VOP <opName>,
   SIMCInstr<opName#"_e32", SISubtarget.NONE> {
   let isPseudo = 1;
   let isCodeGenOnly = 1;
 }
 
 multiclass VOPC_m <vopc op, dag ins, string op_asm, list<dag> pattern,
                    string opName, bit DefExec, VOPProfile p,
+                   list<SchedReadWrite> sched,
                    string revOpName = "", string asm = opName#"_e32 "#op_asm,
                    string alias_asm = opName#" "#op_asm> {
-  def "" : VOPC_Pseudo <ins, pattern, opName>;
+  def "" : VOPC_Pseudo <ins, pattern, opName> {
+    let SchedRW = sched;
+  }
 
   let AssemblerPredicates = [isSICI] in {
-
     def _si : VOPC<op.SI, ins, asm, []>,
               SIMCInstr <opName#"_e32", SISubtarget.SI> {
       let Defs = !if(DefExec, [VCC, EXEC], [VCC]);
       let hasSideEffects = DefExec;
+      let SchedRW = sched;
     }
 
     def : SIInstAlias <
       alias_asm,
       (!cast<Instruction>(NAME#"_e32_si") p.Src0RC32:$src0, p.Src1RC32:$src1)
     >;
 
   } // End AssemblerPredicates = [isSICI]
 
-
   let AssemblerPredicates = [isVI] in {
-
     def _vi : VOPC<op.VI, ins, asm, []>,
               SIMCInstr <opName#"_e32", SISubtarget.VI> {
       let Defs = !if(DefExec, [VCC, EXEC], [VCC]);
       let hasSideEffects = DefExec;
+      let SchedRW = sched;
     }
 
     def : SIInstAlias <
       alias_asm,
       (!cast<Instruction>(NAME#"_e32_vi") p.Src0RC32:$src0, p.Src1RC32:$src1)
     >;
-
   } // End AssemblerPredicates = [isVI]
 }
 
 multiclass VOPC_Helper <vopc op, string opName,
                         dag ins32, string asm32, list<dag> pat32,
                         dag out64, dag ins64, string asm64, list<dag> pat64,
                         bit HasMods, bit DefExec, string revOp,
-                        VOPProfile p> {
-  defm _e32 : VOPC_m <op, ins32, asm32, pat32, opName, DefExec, p>;
+                        VOPProfile p,
+                        list<SchedReadWrite> sched> {
+  defm _e32 : VOPC_m <op, ins32, asm32, pat32, opName, DefExec, p, sched>;
 
   defm _e64 : VOP3_C_m <op, out64, ins64, opName#asm64, pat64,
-                        opName, HasMods, DefExec, revOp>;
+                        opName, HasMods, DefExec, revOp,
+                        sched>;
 }
 
 // Special case for class instructions which only have modifiers on
 // the 1st source operand.
 multiclass VOPC_Class_Helper <vopc op, string opName,
                              dag ins32, string asm32, list<dag> pat32,
                              dag out64, dag ins64, string asm64, list<dag> pat64,
                              bit HasMods, bit DefExec, string revOp,
-                             VOPProfile p> {
-  defm _e32 : VOPC_m <op, ins32, asm32, pat32, opName, DefExec, p>;
+                             VOPProfile p,
+                             list<SchedReadWrite> sched> {
+  defm _e32 : VOPC_m <op, ins32, asm32, pat32, opName, DefExec, p, sched>;
 
   defm _e64 : VOP3_C_m <op, out64, ins64, opName#asm64, pat64,
-                        opName, HasMods, DefExec, revOp>,
+                        opName, HasMods, DefExec, revOp, sched>,
                         VOP3DisableModFields<1, 0, 0>;
 }
 
 multiclass VOPCInst <vopc op, string opName,
                      VOPProfile P, PatLeaf cond = COND_NULL,
                      string revOp = opName,
-                     bit DefExec = 0> : VOPC_Helper <
+                     bit DefExec = 0,
+                     list<SchedReadWrite> sched = [Write32Bit]> :
+                     VOPC_Helper <
   op, opName,
   P.Ins32, P.Asm32, [],
   (outs VOPDstS64:$dst), P.Ins64, P.Asm64,
   !if(P.HasModifiers,
       [(set i1:$dst,
           (setcc (P.Src0VT (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers,
                                       i1:$clamp, i32:$omod)),
                  (P.Src1VT (VOP3Mods P.Src1VT:$src1, i32:$src1_modifiers)),
                  cond))],
       [(set i1:$dst, (setcc P.Src0VT:$src0, P.Src1VT:$src1, cond))]),
-  P.HasModifiers, DefExec, revOp, P
+  P.HasModifiers, DefExec, revOp, P, sched
 >;
 
 multiclass VOPCClassInst <vopc op, string opName, VOPProfile P,
-                     bit DefExec = 0> : VOPC_Class_Helper <
+                     bit DefExec = 0,
+                     list<SchedReadWrite> sched> : VOPC_Class_Helper <
   op, opName,
   P.Ins32, P.Asm32, [],
   (outs VOPDstS64:$dst), P.Ins64, P.Asm64,
   !if(P.HasModifiers,
       [(set i1:$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, opName, P
+  P.HasModifiers, DefExec, opName, P, sched
 >;
 
 
 multiclass VOPC_F32 <vopc op, string opName, PatLeaf cond = COND_NULL, string revOp = opName> :
   VOPCInst <op, opName, VOPC_I1_F32_F32, cond, revOp>;
 
 multiclass VOPC_F64 <vopc op, string opName, PatLeaf cond = COND_NULL, string revOp = opName> :
-  VOPCInst <op, opName, VOPC_I1_F64_F64, cond, revOp>;
+  VOPCInst <op, opName, VOPC_I1_F64_F64, cond, revOp, 0, [WriteDoubleAdd]>;
 
 multiclass VOPC_I32 <vopc op, string opName, PatLeaf cond = COND_NULL, string revOp = opName> :
   VOPCInst <op, opName, VOPC_I1_I32_I32, cond, revOp>;
 
 multiclass VOPC_I64 <vopc op, string opName, PatLeaf cond = COND_NULL, string revOp = opName> :
-  VOPCInst <op, opName, VOPC_I1_I64_I64, cond, revOp>;
+  VOPCInst <op, opName, VOPC_I1_I64_I64, cond, revOp, 0, [Write64Bit]>;
 
 
 multiclass VOPCX <vopc op, string opName, VOPProfile P,
                   PatLeaf cond = COND_NULL,
+                  list<SchedReadWrite> sched,
                   string revOp = "">
-  : VOPCInst <op, opName, P, cond, revOp, 1>;
+  : VOPCInst <op, opName, P, cond, revOp, 1, sched>;
 
 multiclass VOPCX_F32 <vopc op, string opName, string revOp = opName> :
-  VOPCX <op, opName, VOPC_I1_F32_F32, COND_NULL, revOp>;
+  VOPCX <op, opName, VOPC_I1_F32_F32, COND_NULL, [Write32Bit], revOp>;
 
 multiclass VOPCX_F64 <vopc op, string opName, string revOp = opName> :
-  VOPCX <op, opName, VOPC_I1_F64_F64, COND_NULL, revOp>;
+  VOPCX <op, opName, VOPC_I1_F64_F64, COND_NULL, [WriteDoubleAdd], revOp>;
 
 multiclass VOPCX_I32 <vopc op, string opName, string revOp = opName> :
-  VOPCX <op, opName, VOPC_I1_I32_I32, COND_NULL, revOp>;
+  VOPCX <op, opName, VOPC_I1_I32_I32, COND_NULL, [Write32Bit], revOp>;
 
 multiclass VOPCX_I64 <vopc op, string opName, string revOp = opName> :
-  VOPCX <op, opName, VOPC_I1_I64_I64, COND_NULL, revOp>;
+  VOPCX <op, opName, VOPC_I1_I64_I64, COND_NULL, [Write64Bit], revOp>;
 
 multiclass VOP3_Helper <vop3 op, string opName, dag outs, dag ins, string asm,
                         list<dag> pat, int NumSrcArgs, bit HasMods> : VOP3_m <
     op, outs, ins, opName#" "#asm, pat, opName, NumSrcArgs, HasMods
 >;
 
 multiclass VOPC_CLASS_F32 <vopc op, string opName> :
-  VOPCClassInst <op, opName, VOPC_I1_F32_I32, 0>;
+  VOPCClassInst <op, opName, VOPC_I1_F32_I32, 0, [Write32Bit]>;
 
 multiclass VOPCX_CLASS_F32 <vopc op, string opName> :
-  VOPCClassInst <op, opName, VOPC_I1_F32_I32, 1>;
+  VOPCClassInst <op, opName, VOPC_I1_F32_I32, 1, [Write32Bit]>;
 
 multiclass VOPC_CLASS_F64 <vopc op, string opName> :
-  VOPCClassInst <op, opName, VOPC_I1_F64_I32, 0>;
+  VOPCClassInst <op, opName, VOPC_I1_F64_I32, 0, [WriteDoubleAdd]>;
 
 multiclass VOPCX_CLASS_F64 <vopc op, string opName> :
-  VOPCClassInst <op, opName, VOPC_I1_F64_I32, 1>;
+  VOPCClassInst <op, opName, VOPC_I1_F64_I32, 1, [WriteDoubleAdd]>;
 
 multiclass VOP3Inst <vop3 op, string opName, VOPProfile P,
                      SDPatternOperator node = null_frag> : VOP3_Helper <
   op, opName, (outs P.DstRC.RegClass:$dst), P.Ins64, P.Asm64,
   !if(!eq(P.NumSrcArgs, 3),
     !if(P.HasModifiers,
         [(set P.DstVT:$dst,
             (node (P.Src0VT (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers,

lib/Target/AMDGPU/SISchedule.td

 def WriteLDS    : SchedWrite;
 def WriteSALU   : SchedWrite;
 def WriteSMEM   : SchedWrite;
 def WriteVMEM   : SchedWrite;
 
 // Vector ALU instructions
 def Write32Bit         : SchedWrite;
 def WriteQuarterRate32 : SchedWrite;
+def WriteFullOrQuarterRate32 : SchedWrite;
 
 def WriteFloatFMA   : SchedWrite;
 
-def WriteDouble     : SchedWrite;
+// Slow quarter rate f64 instruction.
+def WriteDouble : SchedWrite;
+
+// half rate f64 instruction (same as v_add_f64)
 def WriteDoubleAdd  : SchedWrite;
 
+// Half rate 64-bit instructions.
+def Write64Bit : SchedWrite;
+
+// FIXME: Should there be a class for instructions which are VALU
+// instructions and have VALU rates, but write to the SALU (i.e. VOPC
+// instructions)
+
 def SIFullSpeedModel : SchedMachineModel;
 def SIQuarterSpeedModel : SchedMachineModel;
 
 // BufferSize = 0 means the processors are in-order.
 let BufferSize = 0 in {
 
 // XXX: Are the resource counts correct?
 def HWBranch : ProcResource<1>;
class HWWriteRes<SchedWrite write, list<ProcResourceKind> resources,
   let Latency = latency;
 }
 
 class HWVALUWriteRes<SchedWrite write, int latency> :
   HWWriteRes<write, [HWVALU], latency>;
 
 
 // The latency numbers are taken from AMD Accelerated Parallel Processing
-// guide.  They may not be acurate.
+// guide. They may not be accurate.
 
 // The latency values are 1 / (operations / cycle) / 4.
 multiclass SICommonWriteRes {
 
   def : HWWriteRes<WriteBranch,  [HWBranch], 100>; // XXX: Guessed ???
   def : HWWriteRes<WriteExport,  [HWExport], 100>; // XXX: Guessed ???
   def : HWWriteRes<WriteLDS,     [HWLGKM],    32>; // 2 - 64
   def : HWWriteRes<WriteSALU,    [HWSALU],     1>;
   def : HWWriteRes<WriteSMEM,    [HWLGKM],    10>; // XXX: Guessed ???
   def : HWWriteRes<WriteVMEM,    [HWVMEM],   450>; // 300 - 600
 
   def : HWVALUWriteRes<Write32Bit,         1>;
+  def : HWVALUWriteRes<Write64Bit,         2>;
   def : HWVALUWriteRes<WriteQuarterRate32, 4>;
 }
 
 
 let SchedModel = SIFullSpeedModel in {
 
 defm : SICommonWriteRes;
 

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

define void @test_class_64_f64(i32 addrspace(1)* %out, double %a) #0 {
   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-NOT: vcc
+; SI: 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-NOT: vcc
+; SI: 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, double addrspace(1)* %in, i32 %tid
   %gep.out = getelementptr i32, i32 addrspace(1)* %out, i32 %tid
   %a = load double, double addrspace(1)* %in
 

test/CodeGen/AMDGPU/valu-i1.ll

 
 ; Clear exec bits for workitems that load -1s
 ; SI: BB3_3:
 ; SI: buffer_load_dword [[B:v[0-9]+]]
 ; SI: buffer_load_dword [[A:v[0-9]+]]
 ; SI-DAG: v_cmp_ne_i32_e64 [[NEG1_CHECK_0:s\[[0-9]+:[0-9]+\]]], -1, [[A]]
 ; SI-DAG: v_cmp_ne_i32_e32 [[NEG1_CHECK_1:vcc]], -1, [[B]]
 ; SI: s_and_b64 [[ORNEG1:s\[[0-9]+:[0-9]+\]]], [[NEG1_CHECK_1]], [[NEG1_CHECK_0]]
-; SI: s_and_saveexec_b64 [[ORNEG1]], [[ORNEG1]]
-; SI: s_xor_b64 [[ORNEG1]], exec, [[ORNEG1]]
+; SI: s_and_saveexec_b64 [[ORNEG2:s\[[0-9]+:[0-9]+\]]], [[ORNEG1]]
+; SI: s_xor_b64 [[ORNEG2]], exec, [[ORNEG2]]
 ; SI: s_cbranch_execz BB3_5
 
 ; SI: BB#4:
 ; SI: buffer_store_dword
-; SI: v_cmp_ge_i64_e32 vcc
-; SI: s_or_b64 [[COND_STATE]], vcc, [[COND_STATE]]
+; SI: v_cmp_ge_i64_e64 [[CMP:s\[[0-9]+:[0-9]+\]]]
+; SI: s_or_b64 [[COND_STATE]], [[CMP]], [[COND_STATE]]
 
 ; SI: BB3_5:
-; SI: s_or_b64 exec, exec, [[ORNEG1]]
-; SI: s_or_b64 [[COND_STATE]], [[ORNEG1]], [[COND_STATE]]
+; SI: s_or_b64 exec, exec, [[ORNEG2]]
+; SI: s_or_b64 [[COND_STATE]], [[ORNEG2]], [[COND_STATE]]
 ; SI: s_andn2_b64 exec, exec, [[COND_STATE]]
 ; SI: s_cbranch_execnz BB3_3
 
 ; SI: BB#6
 ; SI: s_or_b64 exec, exec, [[COND_STATE]]
 
 ; SI: BB3_2:
 ; SI-NOT: [[COND_STATE]]