[AMDGPU] Add feature predicate for v_fmac_f64 instruction

[mariusz-sikora-at-amd]

Summary

Introducing feature predicate VFmacF64Inst for targets which
supports v_fmac_f64 instructions.

[foad]

LGTM but please wait to hear from @rampitec too.

[rampitec]

LGTM

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPU.td

539

the current convention drops the v prefix from the name

[mariusz-sikora-at-amd]

llvm/lib/Target/AMDGPU/AMDGPU.td

539

Thanks, I will prepare new change.

llvm/lib/Target/AMDGPU/AMDGPU.td

 >;
 
 def FeatureDLInsts : SubtargetFeature<"dl-insts",
   "HasDLInsts",
   "true",
   "Has v_fmac_f32 and v_xnor_b32 instructions"
 >;
 
+def FeatureVFmacF64Inst : SubtargetFeature<"vfmacf64-inst",
+  "HasVFmacF64Inst",
+  "true",
+  "Has v_fmac_f64 instruction"
+>;
+
 def FeatureDot1Insts : SubtargetFeature<"dot1-insts",
   "HasDot1Insts",
   "true",
   "Has v_dot4_i32_i8 and v_dot8_i32_i4 instructions"
 >;
 
 def FeatureDot2Insts : SubtargetFeature<"dot2-insts",
   "HasDot2Insts",
def FeatureISAVersion9_0_9 : FeatureSet<
    FeatureImageGather4D16Bug]>;
 
 def FeatureISAVersion9_0_A : FeatureSet<
   [FeatureGFX9,
    FeatureGFX90AInsts,
    FeatureFmaMixInsts,
    FeatureLDSBankCount32,
    FeatureDLInsts,
+   FeatureVFmacF64Inst,
    FeatureDot1Insts,
    FeatureDot2Insts,
    FeatureDot3Insts,
    FeatureDot4Insts,
    FeatureDot5Insts,
    FeatureDot6Insts,
    FeatureDot7Insts,
    Feature64BitDPP,

 
 def FeatureISAVersion9_4_0 : FeatureSet<
   [FeatureGFX9,
    FeatureGFX90AInsts,
    FeatureGFX940Insts,
    FeatureFmaMixInsts,
    FeatureLDSBankCount32,
    FeatureDLInsts,
+   FeatureVFmacF64Inst,
    FeatureDot1Insts,
    FeatureDot2Insts,
    FeatureDot3Insts,
    FeatureDot4Insts,
    FeatureDot5Insts,
    FeatureDot6Insts,
    FeatureDot7Insts,
    Feature64BitDPP,
def HasMovrel : Predicate<"Subtarget->hasMovrel()">,
                 AssemblerPredicate<(all_of FeatureMovrel)>;
 
 def HasFmaMixInsts : Predicate<"Subtarget->hasFmaMixInsts()">,
   AssemblerPredicate<(all_of FeatureFmaMixInsts)>;
 
 def HasDLInsts : Predicate<"Subtarget->hasDLInsts()">,
   AssemblerPredicate<(all_of FeatureDLInsts)>;
 
+def HasVFmacF64Inst : Predicate<"Subtarget->hasVFmacF64Inst()">,
+  AssemblerPredicate<(all_of FeatureVFmacF64Inst)>;
+
 def HasDot1Insts : Predicate<"Subtarget->hasDot1Insts()">,
   AssemblerPredicate<(all_of FeatureDot1Insts)>;
 
 def HasDot2Insts : Predicate<"Subtarget->hasDot2Insts()">,
   AssemblerPredicate<(all_of FeatureDot2Insts)>;
 
 def HasDot3Insts : Predicate<"Subtarget->hasDot3Insts()">,
   AssemblerPredicate<(all_of FeatureDot3Insts)>;

llvm/lib/Target/AMDGPU/GCNSubtarget.h

protected:
   bool HasR128A16 = false;
   bool HasA16 = false;
   bool HasG16 = false;
   bool HasNSAEncoding = false;
   unsigned NSAMaxSize = 0;
   bool GFX10_AEncoding = false;
   bool GFX10_BEncoding = false;
   bool HasDLInsts = false;
+  bool HasVFmacF64Inst = false;
   bool HasDot1Insts = false;
   bool HasDot2Insts = false;
   bool HasDot3Insts = false;
   bool HasDot4Insts = false;
   bool HasDot5Insts = false;
   bool HasDot6Insts = false;
   bool HasDot7Insts = false;
   bool HasDot8Insts = false;
public:
   bool hasSDWAOutModsVOPC() const {
     return HasSDWAOutModsVOPC;
   }
 
   bool hasDLInsts() const {
     return HasDLInsts;
   }
 
+  bool hasVFmacF64Inst() const { return HasVFmacF64Inst; }
+
   bool hasDot1Insts() const {
     return HasDot1Insts;
   }
 
   bool hasDot2Insts() const {
     return HasDot2Insts;
   }
 

llvm/lib/Target/AMDGPU/SIInstructions.td

def : GCNPat <
   (fma (f16 (VOP3NoMods f32:$src0)),
        (f16 (VOP3NoMods f32:$src1)),
        (f16 (VOP3NoMods f32:$src2))),
   (V_FMAC_F16_t16_e64 SRCMODS.NONE, $src0, SRCMODS.NONE, $src1,
                   SRCMODS.NONE, $src2)
 >;
 }
 
-let SubtargetPredicate = isGFX90APlus in
+let OtherPredicates = [HasVFmacF64Inst] in
 // Don't allow source modifiers. If there are any source modifiers then it's
 // better to select fma instead of fmac.
 def : GCNPat <
   (fma (f64 (VOP3NoMods f64:$src0)),
        (f64 (VOP3NoMods f64:$src1)),
        (f64 (VOP3NoMods f64:$src2))),
   (V_FMAC_F64_e64 SRCMODS.NONE, $src0, SRCMODS.NONE, $src1,
                   SRCMODS.NONE, $src2)

llvm/lib/Target/AMDGPU/VOP2Instructions.td

 let Constraints = "$vdst = $src2",
     DisableEncoding = "$src2",
     isConvertibleToThreeAddress = 1,
     isCommutable = 1 in
 defm V_FMAC_LEGACY_F32 : VOP2Inst <"v_fmac_legacy_f32", VOP_MAC_LEGACY_F32>;
 
 } // End SubtargetPredicate = HasFmaLegacy32
 
-let SubtargetPredicate = isGFX90APlus,
+let SubtargetPredicate = HasVFmacF64Inst,
     Constraints = "$vdst = $src2",
     DisableEncoding="$src2",
     isConvertibleToThreeAddress = 1,
     isCommutable = 1,
     SchedRW = [WriteDoubleAdd] in
 defm V_FMAC_F64 : VOP2Inst <"v_fmac_f64", VOP_MAC_F64>;
 
 let Constraints = "$vdst = $src2",
foreach _ = BoolToList<!cast<VOP2_Pseudo>(NAME#"_e32").Pfl.HasExtDPP>.ret in
       def _dpp_gfx90a :
         VOP_DPP_Real<!cast<VOP2_DPP_Pseudo>(NAME#"_dpp"), SIEncodingFamily.GFX90A>,
         VOP2_DPPe<op, !cast<VOP2_DPP_Pseudo>(NAME#"_dpp")> {
           let DecoderNamespace = "SDWA9";
         }
   }
 } // End AssemblerPredicate = isGFX90APlus, DecoderNamespace = "GFX90A"
 
-let SubtargetPredicate = isGFX90APlus in {
+let SubtargetPredicate = HasVFmacF64Inst in {
   defm V_FMAC_F64       : VOP2_Real_e32e64_gfx90a <0x4>;
-  let IsSingle = 1 in {
-    defm V_MUL_LEGACY_F32 : VOP2_Real_e64_gfx90a <0x2a1>;
-  }
-} // End SubtargetPredicate = isGFX90APlus
+} // End SubtargetPredicate = HasVFmacF64Inst
+
+let SubtargetPredicate = isGFX90APlus, IsSingle = 1 in {
 
 let SubtargetPredicate = HasFmaakFmamkF32Insts in {
 defm V_FMAMK_F32        : VOP2_Real_MADK_gfx940 <0x17>;
 defm V_FMAAK_F32        : VOP2_Real_MADK_gfx940 <0x18>;
 }
 
 multiclass VOP2_Real_DOT_ACC_gfx9<bits<6> op> : Base_VOP2_Real_e32e64_vi<op> {
   def _dpp_vi : VOP2_DPP<op, !cast<VOP2_DPP_Pseudo>(NAME#"_dpp")>;