[AMDGPU] Add commute for some VOP3 inst (#121326)

add commute for some VOP3 inst, allow commute for both inline constant
operand, adjust tests

Fixes #111205

Author: Shoreshen

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

@@ -2749,6 +2749,63 @@ static MachineInstr *swapRegAndNonRegOperand(MachineInstr &MI,
   return &MI;
 }
 
+static MachineInstr *swapImmOperands(MachineInstr &MI,
+                                     MachineOperand &NonRegOp1,
+                                     MachineOperand &NonRegOp2) {
+  unsigned TargetFlags = NonRegOp1.getTargetFlags();
+  int64_t NonRegVal = NonRegOp1.getImm();
+
+  NonRegOp1.setImm(NonRegOp2.getImm());
+  NonRegOp2.setImm(NonRegVal);
+  NonRegOp1.setTargetFlags(NonRegOp2.getTargetFlags());
+  NonRegOp2.setTargetFlags(TargetFlags);
+  return &MI;
+}
+
+bool SIInstrInfo::isLegalToSwap(const MachineInstr &MI, unsigned OpIdx0,
+                                const MachineOperand *MO0, unsigned OpIdx1,
+                                const MachineOperand *MO1) const {
+  const MCInstrDesc &InstDesc = MI.getDesc();
+  const MCOperandInfo &OpInfo0 = InstDesc.operands()[OpIdx0];
+  const MCOperandInfo &OpInfo1 = InstDesc.operands()[OpIdx1];
+  const TargetRegisterClass *DefinedRC1 =
+      OpInfo1.RegClass != -1 ? RI.getRegClass(OpInfo1.RegClass) : nullptr;
+  const TargetRegisterClass *DefinedRC0 =
+      OpInfo1.RegClass != -1 ? RI.getRegClass(OpInfo0.RegClass) : nullptr;
+
+  unsigned Opc = MI.getOpcode();
+  int Src0Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0);
+
+  // Swap doesn't breach constant bus or literal limits
+  // It may move literal to position other than src0, this is not allowed
+  // pre-gfx10 However, most test cases need literals in Src0 for VOP
+  // FIXME: After gfx9, literal can be in place other than Src0
+  if (isVALU(MI)) {
+    if ((int)OpIdx0 == Src0Idx && !MO0->isReg() &&
+        !isInlineConstant(*MO0, OpInfo1))
+      return false;
+    if ((int)OpIdx1 == Src0Idx && !MO1->isReg() &&
+        !isInlineConstant(*MO1, OpInfo0))
+      return false;
+  }
+
+  if (OpIdx1 != Src0Idx && MO0->isReg()) {
+    if (!DefinedRC1)
+      return OpInfo1.OperandType == MCOI::OPERAND_UNKNOWN;
+    return isLegalRegOperand(MI, OpIdx1, *MO0);
+  }
+  if (OpIdx0 != Src0Idx && MO1->isReg()) {
+    if (!DefinedRC0)
+      return OpInfo0.OperandType == MCOI::OPERAND_UNKNOWN;
+    return isLegalRegOperand(MI, OpIdx0, *MO1);
+  }
+
+  // No need to check 64-bit literals since swapping does not bring new
+  // 64-bit literals into current instruction to fold to 32-bit
+
+  return isImmOperandLegal(MI, OpIdx1, *MO0);
+}
+
 MachineInstr *SIInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
                                                   unsigned Src0Idx,
                                                   unsigned Src1Idx) const {
@@ -2770,21 +2827,20 @@ MachineInstr *SIInstrInfo::commuteInstructionImpl(MachineInstr &MI, bool NewMI,
 
   MachineOperand &Src0 = MI.getOperand(Src0Idx);
   MachineOperand &Src1 = MI.getOperand(Src1Idx);
-
+  if (!isLegalToSwap(MI, Src0Idx, &Src0, Src1Idx, &Src1)) {
+    return nullptr;
+  }
   MachineInstr *CommutedMI = nullptr;
   if (Src0.isReg() && Src1.isReg()) {
-    if (isOperandLegal(MI, Src1Idx, &Src0)) {
-      // Be sure to copy the source modifiers to the right place.
-      CommutedMI
-        = TargetInstrInfo::commuteInstructionImpl(MI, NewMI, Src0Idx, Src1Idx);
-    }
-
+    // Be sure to copy the source modifiers to the right place.
+    CommutedMI =
+        TargetInstrInfo::commuteInstructionImpl(MI, NewMI, Src0Idx, Src1Idx);
   } else if (Src0.isReg() && !Src1.isReg()) {
-    if (isOperandLegal(MI, Src1Idx, &Src0))
-      CommutedMI = swapRegAndNonRegOperand(MI, Src0, Src1);
+    CommutedMI = swapRegAndNonRegOperand(MI, Src0, Src1);
   } else if (!Src0.isReg() && Src1.isReg()) {
-    if (isOperandLegal(MI, Src1Idx, &Src0))
-      CommutedMI = swapRegAndNonRegOperand(MI, Src1, Src0);
+    CommutedMI = swapRegAndNonRegOperand(MI, Src1, Src0);
+  } else if (Src0.isImm() && Src1.isImm()) {
+    CommutedMI = swapImmOperands(MI, Src0, Src1);
   } else {
     // FIXME: Found two non registers to commute. This does happen.
     return nullptr;
@@ -5817,6 +5873,49 @@ bool SIInstrInfo::isLegalRegOperand(const MachineRegisterInfo &MRI,
   return RC->hasSuperClassEq(DRC);
 }
 
+bool SIInstrInfo::isLegalRegOperand(const MachineInstr &MI, unsigned OpIdx,
+                                    const MachineOperand &MO) const {
+  const MachineRegisterInfo &MRI = MI.getParent()->getParent()->getRegInfo();
+  const MCOperandInfo OpInfo = MI.getDesc().operands()[OpIdx];
+  unsigned Opc = MI.getOpcode();
+
+  if (!isLegalRegOperand(MRI, OpInfo, MO))
+    return false;
+
+  // check Accumulate GPR operand
+  bool IsAGPR = RI.isAGPR(MRI, MO.getReg());
+  if (IsAGPR && !ST.hasMAIInsts())
+    return false;
+  if (IsAGPR && (!ST.hasGFX90AInsts() || !MRI.reservedRegsFrozen()) &&
+      (MI.mayLoad() || MI.mayStore() || isDS(Opc) || isMIMG(Opc)))
+    return false;
+  // Atomics should have both vdst and vdata either vgpr or agpr.
+  const int VDstIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
+  const int DataIdx = AMDGPU::getNamedOperandIdx(
+      Opc, isDS(Opc) ? AMDGPU::OpName::data0 : AMDGPU::OpName::vdata);
+  if ((int)OpIdx == VDstIdx && DataIdx != -1 &&
+      MI.getOperand(DataIdx).isReg() &&
+      RI.isAGPR(MRI, MI.getOperand(DataIdx).getReg()) != IsAGPR)
+    return false;
+  if ((int)OpIdx == DataIdx) {
+    if (VDstIdx != -1 &&
+        RI.isAGPR(MRI, MI.getOperand(VDstIdx).getReg()) != IsAGPR)
+      return false;
+    // DS instructions with 2 src operands also must have tied RC.
+    const int Data1Idx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::data1);
+    if (Data1Idx != -1 && MI.getOperand(Data1Idx).isReg() &&
+        RI.isAGPR(MRI, MI.getOperand(Data1Idx).getReg()) != IsAGPR)
+      return false;
+  }
+
+  // Check V_ACCVGPR_WRITE_B32_e64
+  if (Opc == AMDGPU::V_ACCVGPR_WRITE_B32_e64 && !ST.hasGFX90AInsts() &&
+      (int)OpIdx == AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0) &&
+      RI.isSGPRReg(MRI, MO.getReg()))
+    return false;
+  return true;
+}
+
 bool SIInstrInfo::isLegalVSrcOperand(const MachineRegisterInfo &MRI,
                                      const MCOperandInfo &OpInfo,
                                      const MachineOperand &MO) const {
@@ -5879,40 +5978,7 @@ bool SIInstrInfo::isOperandLegal(const MachineInstr &MI, unsigned OpIdx,
   if (MO->isReg()) {
     if (!DefinedRC)
       return OpInfo.OperandType == MCOI::OPERAND_UNKNOWN;
-    if (!isLegalRegOperand(MRI, OpInfo, *MO))
-      return false;
-    bool IsAGPR = RI.isAGPR(MRI, MO->getReg());
-    if (IsAGPR && !ST.hasMAIInsts())
-      return false;
-    unsigned Opc = MI.getOpcode();
-    if (IsAGPR &&
-        (!ST.hasGFX90AInsts() || !MRI.reservedRegsFrozen()) &&
-        (MI.mayLoad() || MI.mayStore() || isDS(Opc) || isMIMG(Opc)))
-      return false;
-    // Atomics should have both vdst and vdata either vgpr or agpr.
-    const int VDstIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdst);
-    const int DataIdx = AMDGPU::getNamedOperandIdx(Opc,
-        isDS(Opc) ? AMDGPU::OpName::data0 : AMDGPU::OpName::vdata);
-    if ((int)OpIdx == VDstIdx && DataIdx != -1 &&
-        MI.getOperand(DataIdx).isReg() &&
-        RI.isAGPR(MRI, MI.getOperand(DataIdx).getReg()) != IsAGPR)
-      return false;
-    if ((int)OpIdx == DataIdx) {
-      if (VDstIdx != -1 &&
-          RI.isAGPR(MRI, MI.getOperand(VDstIdx).getReg()) != IsAGPR)
-        return false;
-      // DS instructions with 2 src operands also must have tied RC.
-      const int Data1Idx = AMDGPU::getNamedOperandIdx(Opc,
-                                                      AMDGPU::OpName::data1);
-      if (Data1Idx != -1 && MI.getOperand(Data1Idx).isReg() &&
-          RI.isAGPR(MRI, MI.getOperand(Data1Idx).getReg()) != IsAGPR)
-        return false;
-    }
-    if (Opc == AMDGPU::V_ACCVGPR_WRITE_B32_e64 && !ST.hasGFX90AInsts() &&
-        (int)OpIdx == AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src0) &&
-        RI.isSGPRReg(MRI, MO->getReg()))
-      return false;
-    return true;
+    return isLegalRegOperand(MI, OpIdx, *MO);
   }
 
   if (MO->isImm()) {

llvm/lib/Target/AMDGPU/SIInstrInfo.h

@@ -193,7 +193,9 @@ class SIInstrInfo final : public AMDGPUGenInstrInfo {
   bool swapSourceModifiers(MachineInstr &MI,
                            MachineOperand &Src0, unsigned Src0OpName,
                            MachineOperand &Src1, unsigned Src1OpName) const;
-
+  bool isLegalToSwap(const MachineInstr &MI, unsigned fromIdx,
+                     const MachineOperand *fromMO, unsigned toIdx,
+                     const MachineOperand *toMO) const;
   MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI,
                                        unsigned OpIdx0,
                                        unsigned OpIdx1) const override;
@@ -1218,11 +1220,13 @@ class SIInstrInfo final : public AMDGPUGenInstrInfo {
                           const MachineOperand &MO) const;
 
   /// Check if \p MO (a register operand) is a legal register for the
-  /// given operand description.
+  /// given operand description or operand index.
+  /// The operand index version provide more legality checks
   bool isLegalRegOperand(const MachineRegisterInfo &MRI,
                          const MCOperandInfo &OpInfo,
                          const MachineOperand &MO) const;
-
+  bool isLegalRegOperand(const MachineInstr &MI, unsigned OpIdx,
+                         const MachineOperand &MO) const;
   /// Legalize operands in \p MI by either commuting it or inserting a
   /// copy of src1.
   void legalizeOperandsVOP2(MachineRegisterInfo &MRI, MachineInstr &MI) const;

llvm/lib/Target/AMDGPU/VOP3Instructions.td

@@ -335,7 +335,9 @@ let isCommutable = 1, SchedRW = [WriteIntMul, WriteSALU] in {
 let FPDPRounding = 1 in {
   let Predicates = [Has16BitInsts, isGFX8Only] in {
     defm V_DIV_FIXUP_F16 : VOP3Inst <"v_div_fixup_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, AMDGPUdiv_fixup>;
-    defm V_FMA_F16 : VOP3Inst <"v_fma_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, any_fma>;
+    let isCommutable = 1 in {
+      defm V_FMA_F16 : VOP3Inst <"v_fma_f16", VOP3_Profile<VOP_F16_F16_F16_F16>, any_fma>;
+    } // End isCommutable = 1
   } // End Predicates = [Has16BitInsts, isGFX8Only]
 
   let SubtargetPredicate = isGFX9Plus in {
@@ -639,8 +641,10 @@ let SubtargetPredicate = HasMinimum3Maximum3F16, ReadsModeReg = 0 in {
 defm V_ADD_I16 : VOP3Inst_t16 <"v_add_i16", VOP_I16_I16_I16>;
 defm V_SUB_I16 : VOP3Inst_t16 <"v_sub_i16", VOP_I16_I16_I16>;
 
-defm V_MAD_U32_U16 : VOP3Inst <"v_mad_u32_u16", VOP3_Profile<VOP_I32_I16_I16_I32, VOP3_OPSEL>>;
-defm V_MAD_I32_I16 : VOP3Inst <"v_mad_i32_i16", VOP3_Profile<VOP_I32_I16_I16_I32, VOP3_OPSEL>>;
+let isCommutable = 1 in {
+  defm V_MAD_U32_U16 : VOP3Inst <"v_mad_u32_u16", VOP3_Profile<VOP_I32_I16_I16_I32, VOP3_OPSEL>>;
+  defm V_MAD_I32_I16 : VOP3Inst <"v_mad_i32_i16", VOP3_Profile<VOP_I32_I16_I16_I32, VOP3_OPSEL>>;
+} // End isCommutable = 1
 
 defm V_CVT_PKNORM_I16_F16 : VOP3Inst_t16 <"v_cvt_pknorm_i16_f16", VOP_B32_F16_F16>;
 defm V_CVT_PKNORM_U16_F16 : VOP3Inst_t16 <"v_cvt_pknorm_u16_f16", VOP_B32_F16_F16>;
@@ -1254,8 +1258,9 @@ let SubtargetPredicate = isGFX10Plus in {
     def : PermlanePat<int_amdgcn_permlane16, V_PERMLANE16_B32_e64, vt>;
     def : PermlanePat<int_amdgcn_permlanex16, V_PERMLANEX16_B32_e64, vt>;
   }
-
-  defm V_ADD_NC_U16 : VOP3Inst_t16 <"v_add_nc_u16", VOP_I16_I16_I16, add>;
+  let isCommutable = 1 in {
+    defm V_ADD_NC_U16 : VOP3Inst_t16 <"v_add_nc_u16", VOP_I16_I16_I16, add>;
+  } // End isCommutable = 1
   defm V_SUB_NC_U16 : VOP3Inst_t16 <"v_sub_nc_u16", VOP_I16_I16_I16, sub>;
 
 } // End SubtargetPredicate = isGFX10Plus

llvm/test/CodeGen/AMDGPU/carryout-selection.ll

@@ -355,7 +355,7 @@ define amdgpu_kernel void @vadd64ri(ptr addrspace(1) %out) {
 ; GFX1010-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 ; GFX1010-NEXT:    v_add_co_u32 v0, s2, 0x56789876, v0
 ; GFX1010-NEXT:    v_mov_b32_e32 v2, 0
-; GFX1010-NEXT:    v_add_co_ci_u32_e64 v1, s2, 0, 0x1234, s2
+; GFX1010-NEXT:    v_add_co_ci_u32_e64 v1, s2, 0x1234, 0, s2
 ; GFX1010-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1010-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
 ; GFX1010-NEXT:    s_endpgm
@@ -365,7 +365,7 @@ define amdgpu_kernel void @vadd64ri(ptr addrspace(1) %out) {
 ; GFX1030W32-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 ; GFX1030W32-NEXT:    v_add_co_u32 v0, s2, 0x56789876, v0
 ; GFX1030W32-NEXT:    v_mov_b32_e32 v2, 0
-; GFX1030W32-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, 0x1234, s2
+; GFX1030W32-NEXT:    v_add_co_ci_u32_e64 v1, null, 0x1234, 0, s2
 ; GFX1030W32-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1030W32-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
 ; GFX1030W32-NEXT:    s_endpgm
@@ -375,7 +375,7 @@ define amdgpu_kernel void @vadd64ri(ptr addrspace(1) %out) {
 ; GFX1030W64-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x24
 ; GFX1030W64-NEXT:    v_add_co_u32 v0, s[2:3], 0x56789876, v0
 ; GFX1030W64-NEXT:    v_mov_b32_e32 v2, 0
-; GFX1030W64-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, 0x1234, s[2:3]
+; GFX1030W64-NEXT:    v_add_co_ci_u32_e64 v1, null, 0x1234, 0, s[2:3]
 ; GFX1030W64-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX1030W64-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
 ; GFX1030W64-NEXT:    s_endpgm
@@ -387,7 +387,7 @@ define amdgpu_kernel void @vadd64ri(ptr addrspace(1) %out) {
 ; GFX11-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-NEXT:    v_add_co_u32 v0, s2, 0x56789876, v0
-; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0, 0x1234, s2
+; GFX11-NEXT:    v_add_co_ci_u32_e64 v1, null, 0x1234, 0, s2
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_store_b64 v2, v[0:1], s[0:1]
 ; GFX11-NEXT:    s_endpgm

llvm/test/CodeGen/AMDGPU/cmp_shrink.mir

@@ -7,6 +7,6 @@ name:             not_shrink_icmp
 body:             |
   bb.0:
     ; GCN-LABEL: name: not_shrink_icmp
-    ; GCN: S_CMP_GT_I32 1, 65, implicit-def $scc
+    ; GCN: S_CMP_LT_I32 65, 1, implicit-def $scc
     S_CMP_GT_I32 1, 65, implicit-def $scc
 ...