Reapply "[AMDGPU][MC] Allow op_sel in v_alignbit_b32 etc in GFX9 and …

llvm/lib/Target/AMDGPU/SIInstructions.td

@@ -2473,6 +2473,7 @@ def : AMDGPUPat <
 >;
 
 let True16Predicate = NotHasTrue16BitInsts in {
+let SubtargetPredicate = isNotGFX9Plus in {
 def : ROTRPattern <V_ALIGNBIT_B32_e64>;
 
 def : GCNPat<(i32 (trunc (srl i64:$src0, (and i32:$src1, (i32 31))))),
@@ -2482,6 +2483,35 @@ def : GCNPat<(i32 (trunc (srl i64:$src0, (and i32:$src1, (i32 31))))),
 def : GCNPat<(i32 (trunc (srl i64:$src0, (i32 ShiftAmt32Imm:$src1)))),
           (V_ALIGNBIT_B32_e64 (i32 (EXTRACT_SUBREG (i64 $src0), sub1)),
                           (i32 (EXTRACT_SUBREG (i64 $src0), sub0)), $src1)>;
+} // isNotGFX9Plus
+
+let SubtargetPredicate = isGFX9GFX10 in {
+def : GCNPat <
+        (rotr i32:$src0, i32:$src1),
+        (V_ALIGNBIT_B32_opsel_e64 /* src0_modifiers */ 0, $src0,
+                                  /* src1_modifiers */ 0, $src0,
+                                  /* src2_modifiers */ 0,
+                                  $src1, /* clamp */ 0, /* op_sel */ 0)
+>;
+
+foreach pat = [(i32 (trunc (srl i64:$src0, (and i32:$src1, (i32 31))))),
+               (i32 (trunc (srl i64:$src0, (i32 ShiftAmt32Imm:$src1))))] in
+def : GCNPat<pat,
+        (V_ALIGNBIT_B32_opsel_e64 0, /* src0_modifiers */
+                                  (i32 (EXTRACT_SUBREG (i64 $src0), sub1)),
+                                  0, /* src1_modifiers */
+                                  (i32 (EXTRACT_SUBREG (i64 $src0), sub0)),
+                                  0, /* src2_modifiers */
+                                  $src1, /* clamp */ 0, /* op_sel */ 0)
+>;
+
+def : GCNPat<(fshr i32:$src0, i32:$src1, i32:$src2),
+        (V_ALIGNBIT_B32_opsel_e64 /* src0_modifiers */ 0, $src0,
+                                  /* src1_modifiers */ 0, $src1,
+                                  /* src2_modifiers */ 0,
+                                  $src2, /* clamp */ 0, /* op_sel */ 0)
+>;
+} // isGFX9GFX10
 } // end True16Predicate = NotHasTrue16BitInsts
 
 let True16Predicate = UseRealTrue16Insts in {
@@ -3082,6 +3112,8 @@ def : GCNPat <
                     (i32 (EXTRACT_SUBREG $a, sub0))), (i32 1))
 >;
 
+// This pattern for bswap is used for pre-GFX8. For GFX8+, bswap is mapped
+// to V_PERM_B32.
 let True16Predicate = NotHasTrue16BitInsts in
 def : GCNPat <
   (i32 (bswap i32:$a)),
@@ -3557,15 +3589,20 @@ def : GCNPat <
 
 // Take the upper 16 bits from V[0] and the lower 16 bits from V[1]
 // Special case, can use V_ALIGNBIT (always uses encoded literal)
-let True16Predicate = NotHasTrue16BitInsts in
-def : GCNPat <
+let True16Predicate = NotHasTrue16BitInsts in {
+defvar BuildVectorToAlignBitPat =
   (vecTy (DivergentBinFrag<build_vector>
     (Ty !if(!eq(Ty, i16),
       (Ty (trunc (srl VGPR_32:$a, (i32 16)))),
       (Ty (bitconvert (i16 (trunc (srl VGPR_32:$a, (i32 16)))))))),
-    (Ty VGPR_32:$b))),
-    (V_ALIGNBIT_B32_e64 VGPR_32:$b, VGPR_32:$a, (i32 16))
->;
+    (Ty VGPR_32:$b)));
+
+let SubtargetPredicate = isNotGFX9Plus in
+def : GCNPat<BuildVectorToAlignBitPat, (V_ALIGNBIT_B32_e64 VGPR_32:$b, VGPR_32:$a, (i32 16))>;
+
+let SubtargetPredicate = isGFX9GFX10 in
+def : GCNPat<BuildVectorToAlignBitPat, (V_ALIGNBIT_B32_opsel_e64 0, VGPR_32:$b, 0, VGPR_32:$a, 0, (i32 16), 0, 0)>;
+} //True16Predicate = NotHasTrue16BitInsts
 
 let True16Predicate = UseFakeTrue16Insts in
 def : GCNPat <

llvm/lib/Target/AMDGPU/VOP3Instructions.td

@@ -224,6 +224,12 @@ defm V_ALIGNBIT_B32 : VOP3Inst_t16_with_profiles <"v_alignbit_b32",
                                                    fshr, null_frag>;
 
 defm V_ALIGNBYTE_B32 : VOP3Inst <"v_alignbyte_b32", VOP3_Profile<VOP_I32_I32_I32_I32>, int_amdgcn_alignbyte>;
+
+// In gfx9 and 10, opsel is allowed for V_ALIGNBIT_B32 and V_ALIGNBYTE_B32.
+// Hardware uses opsel[1:0] to byte-select src2. Other opsel bits are ignored.
+defm V_ALIGNBIT_B32_opsel : VOP3Inst <"v_alignbit_b32_opsel", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_OPSEL>>;
+defm V_ALIGNBYTE_B32_opsel : VOP3Inst <"v_alignbyte_b32_opsel", VOP3_Profile<VOP_I32_I32_I32_I32, VOP3_OPSEL>>;
+
 let True16Predicate = UseRealTrue16Insts in
 defm V_ALIGNBYTE_B32_t16 : VOP3Inst <"v_alignbyte_b32_t16", VOP3_Profile_True16<VOP_I32_I32_I32_I16, VOP3_OPSEL>>;
 let True16Predicate = UseFakeTrue16Insts in
@@ -265,6 +271,16 @@ let SchedRW = [WriteDoubleAdd], FPDPRounding = 1 in {
 } // End SchedRW = [WriteDoubleAdd], FPDPRounding = 1
 } // End isReMaterializable = 1
 
+let SubtargetPredicate = isGFX9GFX10 in
+def : GCNPat <
+(i32 (int_amdgcn_alignbyte (i32 (VOP3OpSelMods i32:$src0, i32:$src0_modifiers)),
+                           (i32 (VOP3OpSelMods i32:$src1, i32:$src1_modifiers)),
+                           (i32 (VOP3OpSelMods i32:$src2, i32:$src2_modifiers)))),
+(V_ALIGNBYTE_B32_opsel_e64 i32:$src0_modifiers, VSrc_b32:$src0,
+                            i32:$src1_modifiers, VSrc_b32:$src1,
+                            i32:$src2_modifiers, VGPR_32:$src2)
+>;
+
 let True16Predicate = UseFakeTrue16Insts in
 def : GCNPat <
 (i32 (int_amdgcn_alignbyte (i32 (VOP3OpSelMods i32:$src0, i32:$src0_modifiers)),
@@ -1954,6 +1970,9 @@ let AssemblerPredicate = isGFX10Only, DecoderNamespace = "GFX10" in {
   }
 } // End AssemblerPredicate = isGFX10Only, DecoderNamespace = "GFX10"
 
+defm V_ALIGNBIT_B32_opsel  : VOP3OpSel_Real_gfx10_with_name<0x14e, "V_ALIGNBIT_B32_opsel", "v_alignbit_b32">;
+defm V_ALIGNBYTE_B32_opsel  : VOP3OpSel_Real_gfx10_with_name<0x14f, "V_ALIGNBYTE_B32_opsel", "v_alignbyte_b32">;
+
 defm V_READLANE_B32  : VOP3_Real_No_Suffix_gfx10<0x360>;
 
 let InOperandList = (ins SSrcOrLds_b32:$src0, SCSrc_b32:$src1, VGPR_32:$vdst_in) in {
@@ -2104,8 +2123,8 @@ defm V_BFI_B32         : VOP3_Real_gfx6_gfx7_gfx10<0x14a>;
 defm V_FMA_F32         : VOP3_Real_gfx6_gfx7_gfx10<0x14b>;
 defm V_FMA_F64         : VOP3_Real_gfx6_gfx7_gfx10<0x14c>;
 defm V_LERP_U8         : VOP3_Real_gfx6_gfx7_gfx10<0x14d>;
-defm V_ALIGNBIT_B32    : VOP3_Real_gfx6_gfx7_gfx10<0x14e>;
-defm V_ALIGNBYTE_B32   : VOP3_Real_gfx6_gfx7_gfx10<0x14f>;
+defm V_ALIGNBIT_B32    : VOP3_Real_gfx6_gfx7<0x14e>;
+defm V_ALIGNBYTE_B32   : VOP3_Real_gfx6_gfx7<0x14f>;
 defm V_MULLIT_F32      : VOP3_Real_gfx6_gfx7_gfx10<0x150>;
 defm V_MIN3_F32        : VOP3_Real_gfx6_gfx7_gfx10<0x151>;
 defm V_MIN3_I32        : VOP3_Real_gfx6_gfx7_gfx10<0x152>;
@@ -2248,6 +2267,17 @@ multiclass VOP3_Real_BITOP3_gfx9<bits<10> op, string AsmName, bit isSingle = 0>
   }
 }
 
+// Instructions such as v_alignbyte_b32 allows op_sel in gfx9, but not in vi.
+// The following is created to support that.
+multiclass VOP3OpSel_Real_gfx9_with_name<bits<10> op, string opName, string AsmName> {
+  defvar psName = opName#"_e64";
+  def _gfx9 : VOP3_Real<!cast<VOP3_Pseudo>(psName), SIEncodingFamily.VI>, // note: encoding family is VI
+            VOP3OpSel_gfx9 <op, !cast<VOP3_Pseudo>(psName).Pfl> {
+              VOP3_Pseudo ps = !cast<VOP3_Pseudo>(psName);
+              let AsmString = AsmName # ps.AsmOperands;
+            }
+}
+
 } // End AssemblerPredicate = isGFX9Only, DecoderNamespace = "GFX9"
 
 defm V_MAD_U64_U32      : VOP3be_Real_vi <0x1E8>;
@@ -2267,8 +2297,10 @@ defm V_BFI_B32          : VOP3_Real_vi <0x1ca>;
 defm V_FMA_F32          : VOP3_Real_vi <0x1cb>;
 defm V_FMA_F64          : VOP3_Real_vi <0x1cc>;
 defm V_LERP_U8          : VOP3_Real_vi <0x1cd>;
+let SubtargetPredicate = isGFX8Only in {
 defm V_ALIGNBIT_B32     : VOP3_Real_vi <0x1ce>;
 defm V_ALIGNBYTE_B32    : VOP3_Real_vi <0x1cf>;
+}
 defm V_MIN3_F32         : VOP3_Real_vi <0x1d0>;
 defm V_MIN3_I32         : VOP3_Real_vi <0x1d1>;
 defm V_MIN3_U32         : VOP3_Real_vi <0x1d2>;
@@ -2313,6 +2345,9 @@ defm V_INTERP_P2_LEGACY_F16 : VOP3Interp_F16_Real_gfx9 <0x276, "V_INTERP_P2_F16"
 defm V_MAD_LEGACY_U16       : VOP3_F16_Real_gfx9 <0x1eb, "V_MAD_U16",       "v_mad_legacy_u16">;
 defm V_MAD_LEGACY_I16       : VOP3_F16_Real_gfx9 <0x1ec, "V_MAD_I16",       "v_mad_legacy_i16">;
 
+defm V_ALIGNBIT_B32_opsel   : VOP3OpSel_Real_gfx9_with_name <0x1ce, "V_ALIGNBIT_B32_opsel", "v_alignbit_b32">;
+defm V_ALIGNBYTE_B32_opsel  : VOP3OpSel_Real_gfx9_with_name <0x1cf, "V_ALIGNBYTE_B32_opsel", "v_alignbyte_b32">;
+
 defm V_MAD_F16_gfx9         : VOP3OpSel_F16_Real_gfx9 <0x203, "v_mad_f16">;
 defm V_MAD_U16_gfx9         : VOP3OpSel_F16_Real_gfx9 <0x204, "v_mad_u16">;
 defm V_MAD_I16_gfx9         : VOP3OpSel_F16_Real_gfx9 <0x205, "v_mad_i16">;

llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-bswap.mir

@@ -1,6 +1,8 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
 # RUN: llc -mtriple=amdgcn -mcpu=hawaii -run-pass=instruction-select -verify-machineinstrs -o - %s  | FileCheck -check-prefix=GFX7 %s
 # RUN: llc -mtriple=amdgcn -mcpu=fiji -run-pass=instruction-select -verify-machineinstrs -o - %s  | FileCheck -check-prefix=GFX8 %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=instruction-select -verify-machineinstrs -o - %s  | FileCheck -check-prefix=GFX9 %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -run-pass=instruction-select -verify-machineinstrs -o - %s  | FileCheck -check-prefix=GFX10 %s
 
 ---
 name: bswap_i32_vv
@@ -19,13 +21,30 @@ body: |
     ; GFX7-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 16711935
     ; GFX7-NEXT: [[V_BFI_B32_e64_:%[0-9]+]]:vgpr_32 = V_BFI_B32_e64 [[S_MOV_B32_]], [[V_ALIGNBIT_B32_e64_1]], [[V_ALIGNBIT_B32_e64_]], implicit $exec
     ; GFX7-NEXT: S_ENDPGM 0, implicit [[V_BFI_B32_e64_]]
+    ;
     ; GFX8-LABEL: name: bswap_i32_vv
     ; GFX8: liveins: $vgpr0
     ; GFX8-NEXT: {{  $}}
     ; GFX8-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
     ; GFX8-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 66051
     ; GFX8-NEXT: [[V_PERM_B32_e64_:%[0-9]+]]:vgpr_32 = V_PERM_B32_e64 0, [[COPY]], [[S_MOV_B32_]], implicit $exec
     ; GFX8-NEXT: S_ENDPGM 0, implicit [[V_PERM_B32_e64_]]
+    ;
+    ; GFX9-LABEL: name: bswap_i32_vv
+    ; GFX9: liveins: $vgpr0
+    ; GFX9-NEXT: {{  $}}
+    ; GFX9-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; GFX9-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 66051
+    ; GFX9-NEXT: [[V_PERM_B32_e64_:%[0-9]+]]:vgpr_32 = V_PERM_B32_e64 0, [[COPY]], [[S_MOV_B32_]], implicit $exec
+    ; GFX9-NEXT: S_ENDPGM 0, implicit [[V_PERM_B32_e64_]]
+    ;
+    ; GFX10-LABEL: name: bswap_i32_vv
+    ; GFX10: liveins: $vgpr0
+    ; GFX10-NEXT: {{  $}}
+    ; GFX10-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; GFX10-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 66051
+    ; GFX10-NEXT: [[V_PERM_B32_e64_:%[0-9]+]]:vgpr_32 = V_PERM_B32_e64 0, [[COPY]], [[S_MOV_B32_]], implicit $exec
+    ; GFX10-NEXT: S_ENDPGM 0, implicit [[V_PERM_B32_e64_]]
     %0:vgpr(s32) = COPY $vgpr0
     %1:vgpr(s32) = G_BSWAP %0
     S_ENDPGM 0, implicit %1

llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-fshr.mir

@@ -1,8 +1,8 @@
 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
 # RUN: llc -mtriple=amdgcn -mcpu=tahiti -run-pass=instruction-select -verify-machineinstrs -o - %s | FileCheck -check-prefix=GCN %s
 # RUN: llc -mtriple=amdgcn -mcpu=fiji -run-pass=instruction-select -verify-machineinstrs -o - %s | FileCheck -check-prefix=GCN %s
-# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=instruction-select -verify-machineinstrs -o - %s | FileCheck -check-prefix=GCN %s
-# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -run-pass=instruction-select -verify-machineinstrs -o - %s | FileCheck -check-prefix=GCN %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx900 -run-pass=instruction-select -verify-machineinstrs -o - %s | FileCheck -check-prefix=GFX9 %s
+# RUN: llc -mtriple=amdgcn -mcpu=gfx1010 -run-pass=instruction-select -verify-machineinstrs -o - %s | FileCheck -check-prefix=GFX10 %s
 # RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -run-pass=instruction-select -verify-machineinstrs -o - %s | FileCheck -check-prefixes=GFX11 %s
 
 ---
@@ -24,6 +24,24 @@ body: |
     ; GCN-NEXT: [[V_ALIGNBIT_B32_e64_:%[0-9]+]]:vgpr_32 = V_ALIGNBIT_B32_e64 [[COPY]], [[COPY1]], [[COPY2]], implicit $exec
     ; GCN-NEXT: S_ENDPGM 0, implicit [[V_ALIGNBIT_B32_e64_]]
     ;
+    ; GFX9-LABEL: name: fshr_s32
+    ; GFX9: liveins: $vgpr0, $vgpr1, $vgpr2
+    ; GFX9-NEXT: {{  $}}
+    ; GFX9-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; GFX9-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+    ; GFX9-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+    ; GFX9-NEXT: [[V_ALIGNBIT_B32_opsel_e64_:%[0-9]+]]:vgpr_32 = V_ALIGNBIT_B32_opsel_e64 0, [[COPY]], 0, [[COPY1]], 0, [[COPY2]], 0, 0, implicit $exec
+    ; GFX9-NEXT: S_ENDPGM 0, implicit [[V_ALIGNBIT_B32_opsel_e64_]]
+    ;
+    ; GFX10-LABEL: name: fshr_s32
+    ; GFX10: liveins: $vgpr0, $vgpr1, $vgpr2
+    ; GFX10-NEXT: {{  $}}
+    ; GFX10-NEXT: [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+    ; GFX10-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+    ; GFX10-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+    ; GFX10-NEXT: [[V_ALIGNBIT_B32_opsel_e64_:%[0-9]+]]:vgpr_32 = V_ALIGNBIT_B32_opsel_e64 0, [[COPY]], 0, [[COPY1]], 0, [[COPY2]], 0, 0, implicit $exec
+    ; GFX10-NEXT: S_ENDPGM 0, implicit [[V_ALIGNBIT_B32_opsel_e64_]]
+    ;
     ; GFX11-LABEL: name: fshr_s32
     ; GFX11: liveins: $vgpr0, $vgpr1, $vgpr2
     ; GFX11-NEXT: {{  $}}

llvm/test/CodeGen/AMDGPU/branch-folding-implicit-def-subreg.ll

@@ -766,10 +766,10 @@ define amdgpu_kernel void @f1(ptr addrspace(1) %arg, ptr addrspace(1) %arg1, i64
   ; GFX90A-NEXT:   renamable $vgpr10 = COPY renamable $sgpr22, implicit $exec
   ; GFX90A-NEXT:   renamable $vgpr12_vgpr13 = DS_READ_B64_gfx9 killed renamable $vgpr10, 0, 0, implicit $exec :: (load (s64) from %ir.8, addrspace 3)
   ; GFX90A-NEXT:   renamable $vgpr10 = COPY renamable $sgpr46, implicit $exec
-  ; GFX90A-NEXT:   renamable $vgpr11 = V_ALIGNBIT_B32_e64 killed $sgpr47, killed $vgpr10, 1, implicit $exec
-  ; GFX90A-NEXT:   renamable $vgpr52 = V_ALIGNBIT_B32_e64 $vgpr17, $vgpr16, 1, implicit $exec
+  ; GFX90A-NEXT:   renamable $vgpr11 = V_ALIGNBIT_B32_opsel_e64 0, killed $sgpr47, 0, killed $vgpr10, 0, 1, 0, 0, implicit $exec
+  ; GFX90A-NEXT:   renamable $vgpr52 = V_ALIGNBIT_B32_opsel_e64 0, $vgpr17, 0, $vgpr16, 0, 1, 0, 0, implicit $exec
   ; GFX90A-NEXT:   renamable $vgpr17 = V_CNDMASK_B32_e64 0, 0, 0, 1, $sgpr12_sgpr13, implicit $exec
-  ; GFX90A-NEXT:   renamable $vgpr15 = V_ALIGNBIT_B32_e64 $vgpr15, $vgpr14, 1, implicit $exec
+  ; GFX90A-NEXT:   renamable $vgpr15 = V_ALIGNBIT_B32_opsel_e64 0, $vgpr15, 0, $vgpr14, 0, 1, 0, 0, implicit $exec
   ; GFX90A-NEXT:   renamable $sgpr52_sgpr53 = S_XOR_B64 $exec, -1, implicit-def dead $scc
   ; GFX90A-NEXT:   renamable $sgpr62_sgpr63 = S_OR_B64 renamable $sgpr36_sgpr37, $exec, implicit-def dead $scc
   ; GFX90A-NEXT:   renamable $vgpr10 = COPY renamable $vgpr14, implicit $exec

llvm/test/CodeGen/AMDGPU/llvm.amdgcn.alignbyte.ll

@@ -1,5 +1,7 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
 ; RUN: llc -mtriple=amdgcn -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX9 %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1030 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX10 %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=+real-true16 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-TRUE16 %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -mattr=-real-true16 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-FAKE16 %s
 
@@ -19,6 +21,30 @@ define amdgpu_kernel void @v_alignbyte_b32(ptr addrspace(1) %out, i32 %src1, i32
 ; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
 ;
+; GFX9-LABEL: v_alignbyte_b32:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x24
+; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v1, s1
+; GFX9-NEXT:    v_mov_b32_e32 v2, s2
+; GFX9-NEXT:    v_alignbyte_b32 v1, s0, v1, v2
+; GFX9-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: v_alignbyte_b32:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_clause 0x1
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x2c
+; GFX10-NEXT:    s_load_dwordx2 s[4:5], s[4:5], 0x24
+; GFX10-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-NEXT:    v_alignbyte_b32 v0, s0, s1, v0
+; GFX10-NEXT:    global_store_dword v1, v0, s[4:5]
+; GFX10-NEXT:    s_endpgm
+;
 ; GFX11-TRUE16-LABEL: v_alignbyte_b32:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_clause 0x1
@@ -73,6 +99,41 @@ define amdgpu_kernel void @v_alignbyte_b32_2(ptr addrspace(1) %out, ptr addrspac
 ; GCN-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; GCN-NEXT:    s_endpgm
 ;
+; GFX9-LABEL: v_alignbyte_b32_2:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX9-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
+; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    global_load_dword v1, v0, s[2:3] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    global_load_dword v2, v0, s[6:7] glc
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_load_dword s2, s[4:5], 0x3c
+; GFX9-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NEXT:    v_alignbyte_b32 v1, v1, v2, s2
+; GFX9-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NEXT:    s_endpgm
+;
+; GFX10-LABEL: v_alignbyte_b32_2:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_clause 0x1
+; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
+; GFX10-NEXT:    s_load_dwordx2 s[6:7], s[4:5], 0x34
+; GFX10-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    global_load_dword v1, v0, s[2:3] glc dlc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    global_load_dword v0, v0, s[6:7] glc dlc
+; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    s_load_dword s2, s[4:5], 0x3c
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    v_alignbyte_b32 v0, v1, v0, s2
+; GFX10-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-NEXT:    s_endpgm
+;
 ; GFX11-TRUE16-LABEL: v_alignbyte_b32_2:
 ; GFX11-TRUE16:       ; %bb.0:
 ; GFX11-TRUE16-NEXT:    s_clause 0x1

llvm/test/MC/AMDGPU/gfx10_asm_vop3.s

@@ -3628,6 +3628,18 @@ v_alignbit_b32 v5, v1, v2, exec_lo
 v_alignbit_b32 v5, v1, v2, exec_hi
 // GFX10: encoding: [0x05,0x00,0x4e,0xd5,0x01,0x05,0xfe,0x01]
 
+v_alignbit_b32 v5, v1, v2, v3 op_sel:[1]
+// GFX10: v_alignbit_b32 v5, v1, v2, v3 op_sel:[1,0,0,0] ; encoding: [0x05,0x08,0x4e,0xd5,0x01,0x05,0x0e,0x04]
+
+v_alignbit_b32 v5, v1, v2, v3 op_sel:[1,1]
+// GFX10: v_alignbit_b32 v5, v1, v2, v3 op_sel:[1,1,0,0] ; encoding: [0x05,0x18,0x4e,0xd5,0x01,0x05,0x0e,0x04]
+
+v_alignbit_b32 v5, v1, v2, v3 op_sel:[1,1,1]
+// GFX10: v_alignbit_b32 v5, v1, v2, v3 op_sel:[1,1,1,0] ; encoding: [0x05,0x38,0x4e,0xd5,0x01,0x05,0x0e,0x04]
+
+v_alignbit_b32 v5, v1, v2, v3 op_sel:[1,1,1,1]
+// GFX10: v_alignbit_b32 v5, v1, v2, v3 op_sel:[1,1,1,1] ; encoding: [0x05,0x78,0x4e,0xd5,0x01,0x05,0x0e,0x04]
+
 v_alignbyte_b32 v5, v1, v2, v3
 // GFX10: encoding: [0x05,0x00,0x4f,0xd5,0x01,0x05,0x0e,0x04]
 
@@ -3715,6 +3727,18 @@ v_alignbyte_b32 v5, v1, v2, exec_lo
 v_alignbyte_b32 v5, v1, v2, exec_hi
 // GFX10: encoding: [0x05,0x00,0x4f,0xd5,0x01,0x05,0xfe,0x01]
 
+v_alignbyte_b32 v5, v1, v2, v3 op_sel:[1]
+// GFX10: v_alignbyte_b32 v5, v1, v2, v3 op_sel:[1,0,0,0] ; encoding: [0x05,0x08,0x4f,0xd5,0x01,0x05,0x0e,0x04]
+
+v_alignbyte_b32 v5, v1, v2, v3 op_sel:[1,1]
+// GFX10: v_alignbyte_b32 v5, v1, v2, v3 op_sel:[1,1,0,0] ; encoding: [0x05,0x18,0x4f,0xd5,0x01,0x05,0x0e,0x04]
+
+v_alignbyte_b32 v5, v1, v2, v3 op_sel:[1,1,1]
+// GFX10: v_alignbyte_b32 v5, v1, v2, v3 op_sel:[1,1,1,0] ; encoding: [0x05,0x38,0x4f,0xd5,0x01,0x05,0x0e,0x04]
+
+v_alignbyte_b32 v5, v1, v2, v3 op_sel:[1,1,1,1]
+// GFX10: v_alignbyte_b32 v5, v1, v2, v3 op_sel:[1,1,1,1] ; encoding: [0x05,0x78,0x4f,0xd5,0x01,0x05,0x0e,0x04]
+
 v_mullit_f32 v5, v1, v2, v3
 // GFX10: encoding: [0x05,0x00,0x50,0xd5,0x01,0x05,0x0e,0x04]