[AMDGPU] Assembler: change v_madmk operands to have same order as mad.

[nhaustov]

Summary

The constant is now at source operand 1 (previously at 2).
This is also how it is in legacy AMD sp3 assembler.
Update tests.

[• tstellarAMD]

LGTM.

lib/Target/AMDGPU/SIInstrInfo.cpp

if (Opc == AMDGPU::V_MAD_F32 || Opc == AMDGPU::V_MAC_F32_e64) {
     // We should only expect these to be on src0 due to canonicalizations.
     if (Src0->isReg() && Src0->getReg() == Reg) {
       if (!Src1->isReg() || RI.isSGPRClass(MRI->getRegClass(Src1->getReg())))
         return false;
 
       if (!Src2->isReg() || RI.isSGPRClass(MRI->getRegClass(Src2->getReg())))
         return false;
 
-      // We need to do some weird looking operand shuffling since the madmk
-      // operands are out of the normal expected order with the multiplied
-      // constant as the last operand.
-      //
-      // v_mad_f32 src0, src1, src2 -> v_madmk_f32 src0 * src2K + src1
-      // src0 -> src2 K
-      // src1 -> src0
-      // src2 -> src1
+      // We need to swap operands 0 and 1 since madmk constant is at operand 1.
 
       const int64_t Imm = DefMI->getOperand(1).getImm();
 
       // FIXME: This would be a lot easier if we could return a new instruction
       // instead of having to modify in place.
 
       // Remove these first since they are at the end.
       UseMI->RemoveOperand(AMDGPU::getNamedOperandIdx(Opc,
                                                       AMDGPU::OpName::omod));
       UseMI->RemoveOperand(AMDGPU::getNamedOperandIdx(Opc,
                                                       AMDGPU::OpName::clamp));
 
       unsigned Src1Reg = Src1->getReg();
       unsigned Src1SubReg = Src1->getSubReg();
-      unsigned Src2Reg = Src2->getReg();
-      unsigned Src2SubReg = Src2->getSubReg();
       Src0->setReg(Src1Reg);
       Src0->setSubReg(Src1SubReg);
       Src0->setIsKill(Src1->isKill());
 
-      Src1->setReg(Src2Reg);
-      Src1->setSubReg(Src2SubReg);
-      Src1->setIsKill(Src2->isKill());
-
       if (Opc == AMDGPU::V_MAC_F32_e64) {
         UseMI->untieRegOperand(
           AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::src2));
       }
 
-      Src2->ChangeToImmediate(Imm);
+      Src1->ChangeToImmediate(Imm);
 
       removeModOperands(*UseMI);
       UseMI->setDesc(get(AMDGPU::V_MADMK_F32));
 
       bool DeleteDef = MRI->hasOneNonDBGUse(Reg);
       if (DeleteDef)
         DefMI->eraseFromParent();
 

lib/Target/AMDGPU/SIInstrInfo.td

 def VOP_I64_I64_I64 : VOPProfile <[i64, i64, i64, untyped]>;
 def VOP_CNDMASK : VOPProfile <[i32, i32, i32, untyped]> {
   let Ins32 = (ins Src0RC32:$src0, Src1RC32:$src1);
   let Ins64 = (ins Src0RC64:$src0, Src1RC64:$src1, SSrc_64:$src2);
   let Asm64 = "$vdst, $src0, $src1, $src2";
 }
 
 def VOP_F32_F32_F32_F32 : VOPProfile <[f32, f32, f32, f32]>;
-def VOP_MADK : VOPProfile <[f32, f32, f32, f32]> {
-  field dag Ins = (ins VCSrc_32:$src0, VGPR_32:$vsrc1, u32imm:$src2);
-  field string Asm = "$vdst, $src0, $vsrc1, $src2";
+def VOP_MADAK : VOPProfile <[f32, f32, f32, f32]> {
+  field dag Ins32 = (ins VCSrc_32:$src0, VGPR_32:$vsrc1, u32imm:$k);
+  field string Asm32 = "$vdst, $src0, $vsrc1, $k";
+}
+def VOP_MADMK : VOPProfile <[f32, f32, f32, f32]> {
+  field dag Ins32 = (ins VCSrc_32:$src0, u32imm:$k, VGPR_32:$vsrc1);
+  field string Asm32 = "$vdst, $src0, $k, $vsrc1";
 }
 def VOP_MAC : VOPProfile <[f32, f32, f32, f32]> {
   let Ins32 = (ins Src0RC32:$src0, Src1RC32:$src1, VGPR_32:$src2);
   let Ins64 = getIns64<Src0RC64, Src1RC64, RegisterOperand<VGPR_32>, 3,
                              HasModifiers>.ret;
   let Asm32 = getAsm32<1, 2, f32>.ret;
   let Asm64 = getAsm64<1, 2, HasModifiers, f32>.ret;
 }
!if(P.HasModifiers,
       [(set P.DstVT:$vdst,
            (node (P.Src0VT (VOP3Mods0 P.Src0VT:$src0, i32:$src0_modifiers,
                                       i1:$clamp, i32:$omod)),
                  (P.Src1VT (VOP3Mods P.Src1VT:$src1, i32:$src1_modifiers))))],
       [(set P.DstVT:$vdst, (node P.Src0VT:$src0, P.Src1VT:$src1))]),
   revOp
 >;
 
-multiclass VOP2MADK <vop2 op, string opName, list<dag> pattern = []> {
+multiclass VOP2MADK <vop2 op, string opName, VOPProfile P, list<dag> pattern = []> {
 
-  def "" : VOP2_Pseudo <VOP_MADK.Outs, VOP_MADK.Ins, pattern, opName>;
+  def "" : VOP2_Pseudo <P.Outs, P.Ins32, pattern, opName>;
 
 let isCodeGenOnly = 0 in {
-  def _si : VOP2Common <VOP_MADK.Outs, VOP_MADK.Ins,
-                        !strconcat(opName, VOP_MADK.Asm), []>,
+  def _si : VOP2Common <P.Outs, P.Ins32,
+                        !strconcat(opName, P.Asm32), []>,
             SIMCInstr <opName#"_e32", SISubtarget.SI>,
             VOP2_MADKe <op.SI> {
             let AssemblerPredicates = [isSICI];
             let DecoderNamespace = "SICI";
             let DisableDecoder = DisableSIDecoder;
             }
 
-  def _vi : VOP2Common <VOP_MADK.Outs, VOP_MADK.Ins,
-                        !strconcat(opName, VOP_MADK.Asm), []>,
+  def _vi : VOP2Common <P.Outs, P.Ins32,
+                        !strconcat(opName, P.Asm32), []>,
             SIMCInstr <opName#"_e32", SISubtarget.VI>,
             VOP2_MADKe <op.VI> {
             let AssemblerPredicates = [isVI];
             let DecoderNamespace = "VI";
             let DisableDecoder = DisableVIDecoder;
             }
 } // End isCodeGenOnly = 0
 }

lib/Target/AMDGPU/SIInstructions.td

 defm V_XOR_B32 : VOP2Inst <vop2<0x1d, 0x15>, "v_xor_b32", VOP_I32_I32_I32>;
 
 let Constraints = "$vdst = $src2", DisableEncoding="$src2",
     isConvertibleToThreeAddress = 1 in {
 defm V_MAC_F32 : VOP2Inst <vop2<0x1f, 0x16>, "v_mac_f32", VOP_MAC>;
 }
 } // End isCommutable = 1
 
-defm V_MADMK_F32 : VOP2MADK <vop2<0x20, 0x17>, "v_madmk_f32">;
+defm V_MADMK_F32 : VOP2MADK <vop2<0x20, 0x17>, "v_madmk_f32", VOP_MADMK>;
 
 let isCommutable = 1 in {
-defm V_MADAK_F32 : VOP2MADK <vop2<0x21, 0x18>, "v_madak_f32">;
+defm V_MADAK_F32 : VOP2MADK <vop2<0x21, 0x18>, "v_madak_f32", VOP_MADAK>;
 } // End isCommutable = 1
 
 let isCommutable = 1 in {
 // No patterns so that the scalar instructions are always selected.
 // The scalar versions will be replaced with vector when needed later.
 
 // V_ADD_I32, V_SUB_I32, and V_SUBREV_I32 where renamed to *_U32 in VI,
 // but the VI instructions behave the same as the SI versions.

lib/Target/AMDGPU/VIInstructions.td

 defm V_ADD_F16 : VOP2Inst <vop2<0, 0x1f>, "v_add_f16", VOP_F16_F16_F16>;
 defm V_SUB_F16 : VOP2Inst <vop2<0, 0x20>, "v_sub_f16", VOP_F16_F16_F16>;
 defm V_SUBREV_F16 : VOP2Inst <vop2<0, 0x21>, "v_subrev_f16", VOP_F16_F16_F16,
   null_frag, "v_sub_f16"
 >;
 defm V_MUL_F16 : VOP2Inst <vop2<0, 0x22>, "v_mul_f16", VOP_F16_F16_F16>;
 defm V_MAC_F16 : VOP2Inst <vop2<0, 0x23>, "v_mac_f16", VOP_F16_F16_F16>;
 } // End isCommutable = 1
-defm V_MADMK_F16 : VOP2MADK <vop2<0,0x24>, "v_madmk_f16">;
+defm V_MADMK_F16 : VOP2MADK <vop2<0,0x24>, "v_madmk_f16", VOP_MADMK>;
 let isCommutable = 1 in {
-defm V_MADAK_F16 : VOP2MADK <vop2<0,0x25>, "v_madak_f16">;
+defm V_MADAK_F16 : VOP2MADK <vop2<0,0x25>, "v_madak_f16", VOP_MADAK>;
 defm V_ADD_U16 : VOP2Inst <vop2<0,0x26>, "v_add_u16", VOP_I16_I16_I16>;
 defm V_SUB_U16 : VOP2Inst <vop2<0,0x27>, "v_sub_u16" , VOP_I16_I16_I16>;
 defm V_SUBREV_U16 : VOP2Inst <vop2<0,0x28>, "v_subrev_u16", VOP_I16_I16_I16>;
 defm V_MUL_LO_U16 : VOP2Inst <vop2<0,0x29>, "v_mul_lo_u16", VOP_I16_I16_I16>;
 } // End isCommutable = 1
 defm V_LSHLREV_B16 : VOP2Inst <vop2<0,0x2a>, "v_lshlrev_b16", VOP_I16_I16_I16>;
 defm V_LSHRREV_B16 : VOP2Inst <vop2<0,0x2b>, "v_lshrrev_b16", VOP_I16_I16_I16>;
 defm V_ASHRREV_B16 : VOP2Inst <vop2<0,0x2c>, "v_ashrrev_b16", VOP_I16_I16_I16>;

test/CodeGen/AMDGPU/madmk.ll

 ; RUN: llc -march=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=SI %s
 ; XUN: llc -march=amdgcn -mcpu=tonga -verify-machineinstrs < %s | FileCheck -check-prefix=GCN -check-prefix=VI %s
 
 declare i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
 declare float @llvm.fabs.f32(float) nounwind readnone
 
 ; GCN-LABEL: {{^}}madmk_f32:
 ; GCN-DAG: buffer_load_dword [[VA:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64{{$}}
 ; GCN-DAG: buffer_load_dword [[VB:v[0-9]+]], {{v\[[0-9]+:[0-9]+\]}}, {{s\[[0-9]+:[0-9]+\]}}, 0 addr64 offset:4
-; GCN: v_madmk_f32_e32 {{v[0-9]+}}, [[VA]], [[VB]], 0x41200000
+; GCN: v_madmk_f32_e32 {{v[0-9]+}}, [[VA]], 0x41200000, [[VB]]
 define void @madmk_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
   %tid = tail call i32 @llvm.amdgcn.workitem.id.x() nounwind readnone
   %gep.0 = getelementptr float, float addrspace(1)* %in, i32 %tid
   %gep.1 = getelementptr float, float addrspace(1)* %gep.0, i32 1
   %out.gep = getelementptr float, float addrspace(1)* %out, i32 %tid
 
   %a = load float, float addrspace(1)* %gep.0, align 4
   %b = load float, float addrspace(1)* %gep.1, align 4
define void @madmk_add_inline_imm_f32(float addrspace(1)* noalias %out, float addrspace(1)* noalias %in) nounwind {
   %mul = fmul float %a, 10.0
   %madmk = fadd float %mul, 2.0
   store float %madmk, float addrspace(1)* %out.gep, align 4
   ret void
 }
 
 ; SI-LABEL: {{^}}kill_madmk_verifier_error:
 ; SI: s_xor_b64
-; SI: v_madmk_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, {{v[0-9]+}}, 0x472aee8c
+; SI: v_madmk_f32_e32 {{v[0-9]+}}, {{v[0-9]+}}, 0x472aee8c, {{v[0-9]+}}
 ; SI: s_or_b64
 define void @kill_madmk_verifier_error() nounwind {
 bb:
   br label %bb2
 
 bb1:                                              ; preds = %bb2
   ret void
 

test/MC/AMDGPU/vop2.s

 // SICI: v_bfm_b32_e32 v1, v2, v3 ; encoding: [0x02,0x07,0x02,0x3c]
 // VI:   v_bfm_b32_e64 v1, v2, v3 ; encoding: [0x01,0x00,0x93,0xd2,0x02,0x07,0x02,0x00]
 v_bfm_b32 v1, v2, v3
 
 // SICI: v_mac_f32_e32 v1, v2, v3 ; encoding: [0x02,0x07,0x02,0x3e]
 // VI:   v_mac_f32_e32 v1, v2, v3 ; encoding: [0x02,0x07,0x02,0x2c]
 v_mac_f32 v1, v2, v3
 
-// SICI: v_madmk_f32_e32 v1, v2, v3, 0x42800000 ; encoding: [0x02,0x07,0x02,0x40,0x00,0x00,0x80,0x42]
-// VI:   v_madmk_f32_e32 v1, v2, v3, 0x42800000 ; encoding: [0x02,0x07,0x02,0x2e,0x00,0x00,0x80,0x42]
-v_madmk_f32 v1, v2, v3, 64.0
+// SICI: v_madmk_f32_e32 v1, v2, 0x42800000, v3 ; encoding: [0x02,0x07,0x02,0x40,0x00,0x00,0x80,0x42]
+// VI:   v_madmk_f32_e32 v1, v2, 0x42800000, v3 ; encoding: [0x02,0x07,0x02,0x2e,0x00,0x00,0x80,0x42]
+v_madmk_f32 v1, v2, 64.0, v3
 
 // SICI: v_madak_f32_e32 v1, v2, v3, 0x42800000 ; encoding: [0x02,0x07,0x02,0x42,0x00,0x00,0x80,0x42]
-// VI:  v_madak_f32_e32 v1, v2, v3, 0x42800000 ; encoding: [0x02,0x07,0x02,0x30,0x00,0x00,0x80,0x42]
+// VI:   v_madak_f32_e32 v1, v2, v3, 0x42800000 ; encoding: [0x02,0x07,0x02,0x30,0x00,0x00,0x80,0x42]
 v_madak_f32 v1, v2, v3, 64.0
 
 // SICI: v_bcnt_u32_b32_e32 v1, v2, v3 ; encoding: [0x02,0x07,0x02,0x44]
 // VI:   v_bcnt_u32_b32_e64 v1, v2, v3 ; encoding: [0x01,0x00,0x8b,0xd2,0x02,0x07,0x02,0x00]
 v_bcnt_u32_b32 v1, v2, v3
 
 // SICI: v_mbcnt_lo_u32_b32_e32 v1, v2, v3 ; encoding: [0x02,0x07,0x02,0x46]
 // VI:   v_mbcnt_lo_u32_b32_e64 v1, v2, v3 ; encoding: [0x01,0x00,0x8c,0xd2,0x02,0x07,0x02,0x00]

 v_mul_f16 v1, v2, v3
 
 // NOSICI: error: instruction not supported on this GPU
 // NOSICI: v_mac_f16 v1, v2, v3
 // VI:     v_mac_f16_e32 v1, v2, v3 ; encoding: [0x02,0x07,0x02,0x46]
 v_mac_f16 v1, v2, v3
 
 // NOSICI: error: instruction not supported on this GPU
-// NOSICI: v_madmk_f16 v1, v2, v3, 64.0
-// VI:     v_madmk_f16_e32 v1, v2, v3, 0x42800000 ; encoding: [0x02,0x07,0x02,0x48,0x00,0x00,0x80,0x42]
-v_madmk_f16 v1, v2, v3, 64.0
+// NOSICI: v_madmk_f16 v1, v2, 64.0, v3
+// VI:     v_madmk_f16_e32 v1, v2, 0x42800000, v3 ; encoding: [0x02,0x07,0x02,0x48,0x00,0x00,0x80,0x42]
+v_madmk_f16 v1, v2, 64.0, v3
 
 // NOSICI: error: instruction not supported on this GPU
 // NOSICI: v_madak_f16 v1, v2, v3, 64.0
 // VI:     v_madak_f16_e32 v1, v2, v3, 0x42800000 ; encoding: [0x02,0x07,0x02,0x4a,0x00,0x00,0x80,0x42]
 v_madak_f16 v1, v2, v3, 64.0
 
 // NOSICI: error: instruction not supported on this GPU
 // NOSICI: v_add_u16 v1, v2, v3