[AMDGPU] Remove extract_subvector patterns

[foad]

Summary

Removing them seems to slightly increase code quality as well as
simplifying both the tablegen and C++ parts of the code.

[Joe_Nash]

LGTM

[arsenm]

llvm/lib/Target/AMDGPU/SIInstructions.td

1399–1407

Most likely caused https://github.com/llvm/llvm-project/issues/64142

llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp

SDValue AMDGPUTargetLowering::LowerEXTRACT_SUBVECTOR(SDValue Op,
   EVT VT = Op.getValueType();
   EVT SrcVT = Op.getOperand(0).getValueType();
 
   if (VT.getScalarSizeInBits() == 16 && Start % 2 == 0) {
     unsigned NumElt = VT.getVectorNumElements();
     unsigned NumSrcElt = SrcVT.getVectorNumElements();
     assert(NumElt % 2 == 0 && NumSrcElt % 2 == 0 && "expect legal types");
 
-    // We have some TableGen patterns for when the extracted vector is exactly
-    // the low or high half of the operand.
-    if ((NumSrcElt == 2 * NumElt) && (Start == 0 || Start == NumElt))
-      return Op;
-
     // Extract 32-bit registers at a time.
     EVT NewSrcVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumSrcElt / 2);
     EVT NewVT = NumElt == 2
                     ? MVT::i32
                     : EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElt / 2);
     SDValue Tmp = DAG.getNode(ISD::BITCAST, SL, NewSrcVT, Op.getOperand(0));
 
     DAG.ExtractVectorElements(Tmp, Args, Start / 2, NumElt / 2);

llvm/lib/Target/AMDGPU/SIInstructions.td

def Extract_Element_v16f32_#Index : Extract_Element <
     f32, v16f32, Index, !cast<SubRegIndex>(sub#Index)
   >;
   def Insert_Element_v16f32_#Index : Insert_Element <
     f32, v16f32, Index, !cast<SubRegIndex>(sub#Index)
   >;
 }
 
 
-def : Pat <
-  (extract_subvector v4i16:$vec, (i32 0)),
-  (v2i16 (EXTRACT_SUBREG v4i16:$vec, sub0))
->;
-
-def : Pat <
-  (extract_subvector v4i16:$vec, (i32 2)),
-  (v2i16 (EXTRACT_SUBREG v4i16:$vec, sub1))
->;
-
-def : Pat <
-  (extract_subvector v4f16:$vec, (i32 0)),
-  (v2f16 (EXTRACT_SUBREG v4f16:$vec, sub0))
->;
-
-def : Pat <
-  (extract_subvector v4f16:$vec, (i32 2)),
-  (v2f16 (EXTRACT_SUBREG v4f16:$vec, sub1))
->;
-
-def : Pat <
-  (extract_subvector v8i16:$vec, (i32 0)),
-  (v4i16 (EXTRACT_SUBREG v8i16:$vec, sub0_sub1))
->;
-
-def : Pat <
-  (extract_subvector v8i16:$vec, (i32 4)),
-  (v4i16 (EXTRACT_SUBREG v8i16:$vec, sub2_sub3))
->;
-
-def : Pat <
-  (extract_subvector v8f16:$vec, (i32 0)),
-  (v4f16 (EXTRACT_SUBREG v8f16:$vec, sub0_sub1))
->;
-
-def : Pat <
-  (extract_subvector v8f16:$vec, (i32 4)),
-  (v4f16 (EXTRACT_SUBREG v8f16:$vec, sub2_sub3))
->;
-
-def : Pat <
-  (extract_subvector v16i16:$vec, (i32 0)),
-  (v8i16 (EXTRACT_SUBREG v16i16:$vec, sub0_sub1_sub2_sub3))
->;
-
-def : Pat <
-  (extract_subvector v16i16:$vec, (i32 8)),
-  (v8i16 (EXTRACT_SUBREG v16i16:$vec, sub4_sub5_sub6_sub7))
->;
-
-def : Pat <
-  (extract_subvector v16f16:$vec, (i32 0)),
-  (v8f16 (EXTRACT_SUBREG v16f16:$vec, sub0_sub1_sub2_sub3))
->;
-
-def : Pat <
-  (extract_subvector v16f16:$vec, (i32 8)),
-  (v8f16 (EXTRACT_SUBREG v16f16:$vec, sub4_sub5_sub6_sub7))
->;
-
 foreach Index = 0-31 in {
   def Extract_Element_v32i32_#Index : Extract_Element <
     i32, v32i32, Index, !cast<SubRegIndex>(sub#Index)
   >;
 
   def Insert_Element_v32i32_#Index : Insert_Element <
     i32, v32i32, Index, !cast<SubRegIndex>(sub#Index)
   >;

llvm/test/CodeGen/AMDGPU/extract-subvector-16bit.ll

 ; GFX9-NEXT:    s_cbranch_execz .LBB2_3
 ; GFX9-NEXT:    s_branch .LBB2_4
 ; GFX9-NEXT:  .LBB2_2:
 ; GFX9-NEXT:    ; implicit-def: $vgpr2_vgpr3_vgpr4_vgpr5
 ; GFX9-NEXT:  .LBB2_3: ; %T
 ; GFX9-NEXT:    global_load_dwordx4 v[2:5], v[0:1], off glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:  .LBB2_4: ; %exit
-; GFX9-NEXT:    s_mov_b32 s4, 0x5040100
+; GFX9-NEXT:    v_mov_b32_e32 v0, 0x3900
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0x3d00
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_perm_b32 v0, v3, v3, s4
-; GFX9-NEXT:    v_mov_b32_e32 v1, 0x3800
-; GFX9-NEXT:    v_mov_b32_e32 v3, 0x3900
-; GFX9-NEXT:    v_mov_b32_e32 v4, 0x3d00
-; GFX9-NEXT:    v_cmp_ge_f16_e32 vcc, 0.5, v0
-; GFX9-NEXT:    v_cndmask_b32_e32 v5, v3, v4, vcc
-; GFX9-NEXT:    v_cmp_nle_f16_sdwa vcc, v0, v1 src0_sel:WORD_1 src1_sel:DWORD
-; GFX9-NEXT:    v_cndmask_b32_e32 v6, v4, v3, vcc
 ; GFX9-NEXT:    v_cmp_ge_f16_e32 vcc, 0.5, v2
-; GFX9-NEXT:    v_cndmask_b32_e32 v0, v3, v4, vcc
-; GFX9-NEXT:    v_cmp_le_f16_sdwa vcc, v2, v1 src0_sel:WORD_1 src1_sel:DWORD
-; GFX9-NEXT:    v_cndmask_b32_e32 v1, v3, v4, vcc
-; GFX9-NEXT:    v_pack_b32_f16 v0, v0, v1
-; GFX9-NEXT:    v_pack_b32_f16 v1, v5, v6
+; GFX9-NEXT:    v_mov_b32_e32 v5, 0x3800
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v0, v1, vcc
+; GFX9-NEXT:    v_cmp_le_f16_sdwa vcc, v2, v5 src0_sel:WORD_1 src1_sel:DWORD
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v0, v1, vcc
+; GFX9-NEXT:    v_cmp_nge_f16_e32 vcc, 0.5, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v5, v1, v0, vcc
+; GFX9-NEXT:    v_cmp_ge_f16_e32 vcc, 0.5, v3
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
+; GFX9-NEXT:    v_pack_b32_f16 v1, v0, v5
+; GFX9-NEXT:    v_pack_b32_f16 v0, v4, v2
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
   br i1 undef, label %T, label %F
 
 T:
   %t = load volatile <8 x half>, ptr addrspace(1) %p0
   br label %exit
 
 F:

 ; GFX9-NEXT:    ; implicit-def: $vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11
 ; GFX9-NEXT:  .LBB5_3: ; %T
 ; GFX9-NEXT:    global_load_dwordx4 v[2:5], v[0:1], off offset:16 glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    global_load_dwordx4 v[4:7], v[0:1], off glc
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-NEXT:    ; kill: killed $vgpr0 killed $vgpr1
 ; GFX9-NEXT:  .LBB5_4: ; %exit
-; GFX9-NEXT:    s_mov_b32 s4, 0x5040100
+; GFX9-NEXT:    v_mov_b32_e32 v0, 0x3900
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0x3d00
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_perm_b32 v0, v5, v5, s4
-; GFX9-NEXT:    v_mov_b32_e32 v1, 0x3800
-; GFX9-NEXT:    v_mov_b32_e32 v2, 0x3900
-; GFX9-NEXT:    v_mov_b32_e32 v3, 0x3d00
-; GFX9-NEXT:    v_cmp_ge_f16_e32 vcc, 0.5, v0
-; GFX9-NEXT:    v_cndmask_b32_e32 v5, v2, v3, vcc
-; GFX9-NEXT:    v_cmp_nle_f16_sdwa vcc, v0, v1 src0_sel:WORD_1 src1_sel:DWORD
-; GFX9-NEXT:    v_cndmask_b32_e32 v6, v3, v2, vcc
 ; GFX9-NEXT:    v_cmp_ge_f16_e32 vcc, 0.5, v4
-; GFX9-NEXT:    v_cndmask_b32_e32 v0, v2, v3, vcc
-; GFX9-NEXT:    v_cmp_le_f16_sdwa vcc, v4, v1 src0_sel:WORD_1 src1_sel:DWORD
-; GFX9-NEXT:    v_cndmask_b32_e32 v1, v2, v3, vcc
-; GFX9-NEXT:    v_pack_b32_f16 v0, v0, v1
-; GFX9-NEXT:    v_pack_b32_f16 v1, v5, v6
+; GFX9-NEXT:    v_mov_b32_e32 v3, 0x3800
+; GFX9-NEXT:    v_cndmask_b32_e32 v2, v0, v1, vcc
+; GFX9-NEXT:    v_cmp_le_f16_sdwa vcc, v4, v3 src0_sel:WORD_1 src1_sel:DWORD
+; GFX9-NEXT:    v_cndmask_b32_e32 v3, v0, v1, vcc
+; GFX9-NEXT:    v_cmp_nge_f16_e32 vcc, 0.5, v5
+; GFX9-NEXT:    v_cndmask_b32_e32 v4, v1, v0, vcc
+; GFX9-NEXT:    v_cmp_ge_f16_e32 vcc, 0.5, v5
+; GFX9-NEXT:    v_cndmask_b32_e32 v0, v0, v1, vcc
+; GFX9-NEXT:    v_pack_b32_f16 v1, v0, v4
+; GFX9-NEXT:    v_pack_b32_f16 v0, v2, v3
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
   br i1 undef, label %T, label %F
 
 T:
   %t = load volatile <16 x half>, ptr addrspace(1) %p0
   br label %exit
 
 F:

llvm/test/CodeGen/AMDGPU/greedy-reverse-local-assignment.ll

 ; FORWARDXNACK-NEXT:    v_mov_b32_e32 v0, v6
 ; FORWARDXNACK-NEXT:    s_waitcnt vmcnt(0)
 ; FORWARDXNACK-NEXT:    v_mov_b32_e32 v1, v4
 ; FORWARDXNACK-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; REVERSEXNACK-LABEL: shuffle_v4f16_234u:
 ; REVERSEXNACK:       ; %bb.0:
 ; REVERSEXNACK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
-; REVERSEXNACK-NEXT:    v_mov_b32_e32 v6, v1
-; REVERSEXNACK-NEXT:    v_mov_b32_e32 v5, v0
-; REVERSEXNACK-NEXT:    v_mov_b32_e32 v4, v3
-; REVERSEXNACK-NEXT:    v_mov_b32_e32 v3, v2
-; REVERSEXNACK-NEXT:    global_load_dword v0, v[5:6], off offset:4
-; REVERSEXNACK-NEXT:    global_load_dwordx2 v[1:2], v[3:4], off
+; REVERSEXNACK-NEXT:    v_mov_b32_e32 v6, v3
+; REVERSEXNACK-NEXT:    v_mov_b32_e32 v4, v1
+; REVERSEXNACK-NEXT:    v_mov_b32_e32 v3, v0
+; REVERSEXNACK-NEXT:    v_mov_b32_e32 v5, v2
+; REVERSEXNACK-NEXT:    global_load_dword v0, v[3:4], off offset:4
+; REVERSEXNACK-NEXT:    global_load_dwordx2 v[1:2], v[5:6], off
 ; REVERSEXNACK-NEXT:    s_waitcnt vmcnt(0)
 ; REVERSEXNACK-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; NOXNACK-LABEL: shuffle_v4f16_234u:
 ; NOXNACK:       ; %bb.0:
 ; NOXNACK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; NOXNACK-NEXT:    global_load_dword v0, v[0:1], off offset:4
 ; NOXNACK-NEXT:    global_load_dwordx2 v[1:2], v[2:3], off
 ; NOXNACK-NEXT:    s_waitcnt vmcnt(0)
 ; NOXNACK-NEXT:    s_setpc_b64 s[30:31]
   %val0 = load <4 x half>, ptr addrspace(1) %arg0
   %val1 = load <4 x half>, ptr addrspace(1) %arg1
   %shuffle = shufflevector <4 x half> %val0, <4 x half> %val1, <4 x i32> <i32 2, i32 3, i32 4, i32 undef>
   ret <4 x half> %shuffle
 }

llvm/test/CodeGen/AMDGPU/sra.ll

 ; VI-NEXT:    s_mov_b32 s11, s7
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_mov_b32 s8, s2
 ; VI-NEXT:    s_mov_b32 s9, s3
 ; VI-NEXT:    buffer_load_dwordx2 v[0:1], off, s[8:11], 0
 ; VI-NEXT:    s_mov_b32 s4, s0
 ; VI-NEXT:    s_mov_b32 s5, s1
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_readfirstlane_b32 s0, v0
-; VI-NEXT:    v_readfirstlane_b32 s1, v1
-; VI-NEXT:    s_ashr_i32 s2, s0, 16
-; VI-NEXT:    s_sext_i32_i16 s0, s0
-; VI-NEXT:    s_ashr_i32 s3, s1, 16
+; VI-NEXT:    v_readfirstlane_b32 s0, v1
+; VI-NEXT:    v_readfirstlane_b32 s1, v0
+; VI-NEXT:    s_ashr_i32 s2, s1, 16
 ; VI-NEXT:    s_sext_i32_i16 s1, s1
-; VI-NEXT:    s_ashr_i32 s0, s0, s1
+; VI-NEXT:    s_ashr_i32 s3, s0, 16
+; VI-NEXT:    s_sext_i32_i16 s0, s0
+; VI-NEXT:    s_ashr_i32 s0, s1, s0
 ; VI-NEXT:    s_ashr_i32 s1, s2, s3
 ; VI-NEXT:    s_lshl_b32 s1, s1, 16
 ; VI-NEXT:    s_and_b32 s0, s0, 0xffff
 ; VI-NEXT:    s_or_b32 s0, s0, s1
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;