[AMDGPU][GFX10] Support base+soffset+offset SMEM stores.

[kosarev]

Summary

Also makes another step towards resolving
https://github.com/llvm/llvm-project/issues/38652

[kosarev]

A couple notes:

• It seems I missed tests for scratch loads in D125117. Will add them later.
• Looking at how test/MC/AMDGPU/smem.s is organised, it's not perfectly clear to me if we want offset-validation tests for the stores where we already have them for loads.

[foad]

LGTM, thanks!

[dp]

LGTM

Looking at how test/MC/AMDGPU/smem.s is organised, it's not perfectly clear to me if we want offset-validation tests for the stores where we already have them for loads.

AFAIK, all features should be covered by lit tests, but there are no specific requirements on test coverage. If it is clear that existing tests cover new functionality, you do not have to add new tests.

llvm/lib/Target/AMDGPU/SMInstructions.td

class SM_Load_Pseudo <string opName, dag outs, dag ins, string asmOps, list<dag> pattern=[]>
   : SM_Pseudo<opName, outs, ins, asmOps, pattern> {
   RegisterClass BaseClass;
   let mayLoad = 1;
   let mayStore = 0;
   let has_glc = 1;
   let has_dlc = 1;
 }
 
-class SM_Store_Pseudo <string opName, dag ins, string asmOps, list<dag> pattern = []>
-  : SM_Pseudo<opName, (outs), ins, asmOps, pattern> {
-  RegisterClass BaseClass;
-  RegisterClass SrcClass;
+class SM_Store_Pseudo <string opName, RegisterClass baseClass,
+                       RegisterClass srcClass, dag ins, string asmOps>
+  : SM_Pseudo<opName, (outs), ins, asmOps, []> {
+  RegisterClass BaseClass = baseClass;
+  RegisterClass SrcClass = srcClass;
   let mayLoad = 0;
   let mayStore = 1;
   let has_glc = 1;
   let has_dlc = 1;
   let ScalarStore = 1;
 }
 
 class SM_Discard_Pseudo <string opName, dag ins, bit isImm>
def _SGPR_IMM  : SM_Load_Pseudo <opName,
     let has_glc = 1;
     let has_dlc = 1;
   }
 }
 
 multiclass SM_Pseudo_Stores<string opName,
                            RegisterClass baseClass,
                            RegisterClass srcClass> {
-  def _IMM  : SM_Store_Pseudo <opName,
+  def _IMM : SM_Store_Pseudo <opName, baseClass, srcClass,
     (ins srcClass:$sdata, baseClass:$sbase, i32imm:$offset, CPol:$cpol),
-    " $sdata, $sbase, $offset$cpol", []> {
+    " $sdata, $sbase, $offset$cpol"> {
     let has_offset = 1;
-    let BaseClass = baseClass;
-    let SrcClass = srcClass;
     let PseudoInstr = opName # "_IMM";
   }
 
-  def _SGPR  : SM_Store_Pseudo <opName,
+  def _SGPR : SM_Store_Pseudo <opName, baseClass, srcClass,
     (ins srcClass:$sdata, baseClass:$sbase, SReg_32:$soffset, CPol:$cpol),
-    " $sdata, $sbase, $soffset$cpol", []> {
+    " $sdata, $sbase, $soffset$cpol"> {
     let has_soffset = 1;
-    let BaseClass = baseClass;
-    let SrcClass = srcClass;
     let PseudoInstr = opName # "_SGPR";
   }
+
+  def _SGPR_IMM : SM_Store_Pseudo <opName, baseClass, srcClass,
+    (ins srcClass:$sdata, baseClass:$sbase, SReg_32:$soffset, i32imm:$offset,
+     CPol:$cpol),
+    " $sdata, $sbase, $soffset$offset$cpol"> {
+    let has_offset = 1;
+    let has_soffset = 1;
+    let PseudoInstr = opName # "_SGPR_IMM";
+  }
 }
 
 multiclass SM_Pseudo_Discards<string opName> {
   def _IMM  : SM_Discard_Pseudo <opName, (ins SReg_64:$sbase, smem_offset:$offset), 1>;
   def _SGPR : SM_Discard_Pseudo <opName, (ins SReg_64:$sbase, SReg_32:$soffset), 0>;
 }
 
 class SM_Time_Pseudo<string opName, SDPatternOperator node = null_frag> : SM_Pseudo<
multiclass SM_Real_Stores_gfx10<bits<8> op, string ps,
                                 SM_Store_Pseudo sgprPs = !cast<SM_Store_Pseudo>(ps#_SGPR)> {
   def _IMM_gfx10 : SMEM_Real_Store_gfx10 <op, immPs> {
     let InOperandList = (ins immPs.SrcClass:$sdata, immPs.BaseClass:$sbase, smem_offset:$offset, CPol:$cpol);
   }
 
   def _SGPR_gfx10 : SMEM_Real_Store_gfx10 <op, sgprPs> {
     let InOperandList = (ins sgprPs.SrcClass:$sdata, sgprPs.BaseClass:$sbase, SReg_32:$soffset, CPol:$cpol);
   }
+
+  def _SGPR_IMM_gfx10 : SMEM_Real_Store_gfx10 <op, !cast<SM_Store_Pseudo>(ps#_SGPR_IMM)> {
+    let InOperandList = (ins sgprPs.SrcClass:$sdata, sgprPs.BaseClass:$sbase,
+                             SReg_32:$soffset, smem_offset_mod:$offset, CPol:$cpol);
+  }
 }
 
 defm S_LOAD_DWORD            : SM_Real_Loads_gfx10<0x000, "S_LOAD_DWORD">;
 defm S_LOAD_DWORDX2          : SM_Real_Loads_gfx10<0x001, "S_LOAD_DWORDX2">;
 defm S_LOAD_DWORDX4          : SM_Real_Loads_gfx10<0x002, "S_LOAD_DWORDX4">;
 defm S_LOAD_DWORDX8          : SM_Real_Loads_gfx10<0x003, "S_LOAD_DWORDX8">;
 defm S_LOAD_DWORDX16         : SM_Real_Loads_gfx10<0x004, "S_LOAD_DWORDX16">;
 

llvm/test/MC/AMDGPU/gfx10_asm_smem.s

 // GFX10: encoding: [0x42,0x00,0x40,0xf4,0x00,0x00,0x00,0xf8]
 
 s_store_dword s1, s[4:5], null
 // GFX10: encoding: [0x42,0x00,0x40,0xf4,0x00,0x00,0x00,0xfa]
 
 s_store_dword s1, s[4:5], 0x0
 // GFX10: encoding: [0x42,0x00,0x40,0xf4,0x00,0x00,0x00,0xfa]
 
+s_store_dword s1, s[4:5], s0 offset:0x12345
+// GFX10: encoding: [0x42,0x00,0x40,0xf4,0x45,0x23,0x01,0x00]
+
 s_store_dword s1, s[4:5], s0 glc
 // GFX10: encoding: [0x42,0x00,0x41,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dword s1, s[4:5], s0 dlc
 // GFX10: encoding: [0x42,0x40,0x40,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dword s1, s[4:5], s0 glc dlc
 // GFX10: encoding: [0x42,0x40,0x41,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dword s1, s[4:5], 0x1234 glc dlc
 // GFX10: encoding: [0x42,0x40,0x41,0xf4,0x34,0x12,0x00,0xfa]
 
+s_store_dword s1, s[4:5], s0 offset:0x12345 glc dlc
+// GFX10: encoding: [0x42,0x40,0x41,0xf4,0x45,0x23,0x01,0x00]
+
 s_store_dwordx2 s[2:3], s[4:5], s0
 // GFX10: encoding: [0x82,0x00,0x44,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dwordx2 s[4:5], s[4:5], s0
 // GFX10: encoding: [0x02,0x01,0x44,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dwordx2 s[100:101], s[4:5], s0
 // GFX10: encoding: [0x02,0x19,0x44,0xf4,0x00,0x00,0x00,0x00]

 // GFX10: encoding: [0x82,0x00,0x44,0xf4,0x00,0x00,0x00,0xf8]
 
 s_store_dwordx2 s[2:3], s[4:5], null
 // GFX10: encoding: [0x82,0x00,0x44,0xf4,0x00,0x00,0x00,0xfa]
 
 s_store_dwordx2 s[2:3], s[4:5], 0x0
 // GFX10: encoding: [0x82,0x00,0x44,0xf4,0x00,0x00,0x00,0xfa]
 
+s_store_dwordx2 s[2:3], s[4:5], s0 offset:0x12345
+// GFX10: encoding: [0x82,0x00,0x44,0xf4,0x45,0x23,0x01,0x00]
+
 s_store_dwordx2 s[2:3], s[4:5], s0 glc
 // GFX10: encoding: [0x82,0x00,0x45,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dwordx2 s[2:3], s[4:5], s0 dlc
 // GFX10: encoding: [0x82,0x40,0x44,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dwordx2 s[2:3], s[4:5], s0 glc dlc
 // GFX10: encoding: [0x82,0x40,0x45,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dwordx2 s[2:3], s[4:5], 0x1234 glc dlc
 // GFX10: encoding: [0x82,0x40,0x45,0xf4,0x34,0x12,0x00,0xfa]
 
+s_store_dwordx2 s[2:3], s[4:5], s0 offset:0x12345 glc dlc
+// GFX10: encoding: [0x82,0x40,0x45,0xf4,0x45,0x23,0x01,0x00]
+
 s_store_dwordx4 s[4:7], s[4:5], s0
 // GFX10: encoding: [0x02,0x01,0x48,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dwordx4 s[8:11], s[4:5], s0
 // GFX10: encoding: [0x02,0x02,0x48,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dwordx4 s[96:99], s[4:5], s0
 // GFX10: encoding: [0x02,0x18,0x48,0xf4,0x00,0x00,0x00,0x00]

 // GFX10: encoding: [0x02,0x01,0x48,0xf4,0x00,0x00,0x00,0xf8]
 
 s_store_dwordx4 s[4:7], s[4:5], null
 // GFX10: encoding: [0x02,0x01,0x48,0xf4,0x00,0x00,0x00,0xfa]
 
 s_store_dwordx4 s[4:7], s[4:5], 0x0
 // GFX10: encoding: [0x02,0x01,0x48,0xf4,0x00,0x00,0x00,0xfa]
 
+s_store_dwordx4 s[4:7], s[4:5], s0 offset:0x12345
+// GFX10: encoding: [0x02,0x01,0x48,0xf4,0x45,0x23,0x01,0x00]
+
 s_store_dwordx4 s[4:7], s[4:5], s0 glc
 // GFX10: encoding: [0x02,0x01,0x49,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dwordx4 s[4:7], s[4:5], s0 dlc
 // GFX10: encoding: [0x02,0x41,0x48,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dwordx4 s[4:7], s[4:5], s0 glc dlc
 // GFX10: encoding: [0x02,0x41,0x49,0xf4,0x00,0x00,0x00,0x00]
 
 s_store_dwordx4 s[4:7], s[4:5], 0x1234 glc dlc
 // GFX10: encoding: [0x02,0x41,0x49,0xf4,0x34,0x12,0x00,0xfa]
 
+s_store_dwordx4 s[4:7], s[4:5], s0 offset:0x12345 glc dlc
+// GFX10: encoding: [0x02,0x41,0x49,0xf4,0x45,0x23,0x01,0x00]
+
 s_buffer_store_dword s1, s[8:11], s0
 // GFX10: encoding: [0x44,0x00,0x60,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dword s101, s[8:11], s0
 // GFX10: encoding: [0x44,0x19,0x60,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dword vcc_lo, s[8:11], s0
 // GFX10: encoding: [0x84,0x1a,0x60,0xf4,0x00,0x00,0x00,0x00]

 // GFX10: encoding: [0x44,0x00,0x60,0xf4,0x00,0x00,0x00,0xf8]
 
 s_buffer_store_dword s1, s[8:11], null
 // GFX10: encoding: [0x44,0x00,0x60,0xf4,0x00,0x00,0x00,0xfa]
 
 s_buffer_store_dword s1, s[8:11], 0x0
 // GFX10: encoding: [0x44,0x00,0x60,0xf4,0x00,0x00,0x00,0xfa]
 
+s_buffer_store_dword s1, s[8:11], s0 offset:0x12345
+// GFX10: encoding: [0x44,0x00,0x60,0xf4,0x45,0x23,0x01,0x00]
+
 s_buffer_store_dword s1, s[8:11], s0 glc
 // GFX10: encoding: [0x44,0x00,0x61,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dword s1, s[8:11], s0 dlc
 // GFX10: encoding: [0x44,0x40,0x60,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dword s1, s[8:11], s0 glc dlc
 // GFX10: encoding: [0x44,0x40,0x61,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dword s1, s[8:11], 0x1234 glc dlc
 // GFX10: encoding: [0x44,0x40,0x61,0xf4,0x34,0x12,0x00,0xfa]
 
+s_buffer_store_dword s1, s[8:11], s0 offset:0x12345 glc dlc
+// GFX10: encoding: [0x44,0x40,0x61,0xf4,0x45,0x23,0x01,0x00]
+
 s_buffer_store_dwordx2 s[2:3], s[8:11], s0
 // GFX10: encoding: [0x84,0x00,0x64,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dwordx2 s[4:5], s[8:11], s0
 // GFX10: encoding: [0x04,0x01,0x64,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dwordx2 s[100:101], s[8:11], s0
 // GFX10: encoding: [0x04,0x19,0x64,0xf4,0x00,0x00,0x00,0x00]

 // GFX10: encoding: [0x84,0x00,0x64,0xf4,0x00,0x00,0x00,0xf8]
 
 s_buffer_store_dwordx2 s[2:3], s[8:11], null
 // GFX10: encoding: [0x84,0x00,0x64,0xf4,0x00,0x00,0x00,0xfa]
 
 s_buffer_store_dwordx2 s[2:3], s[8:11], 0x0
 // GFX10: encoding: [0x84,0x00,0x64,0xf4,0x00,0x00,0x00,0xfa]
 
+s_buffer_store_dwordx2 s[2:3], s[8:11], s0 offset:0x12345
+// GFX10: encoding: [0x84,0x00,0x64,0xf4,0x45,0x23,0x01,0x00]
+
 s_buffer_store_dwordx2 s[2:3], s[8:11], s0 glc
 // GFX10: encoding: [0x84,0x00,0x65,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dwordx2 s[2:3], s[8:11], s0 dlc
 // GFX10: encoding: [0x84,0x40,0x64,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dwordx2 s[2:3], s[8:11], s0 glc dlc
 // GFX10: encoding: [0x84,0x40,0x65,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dwordx2 s[2:3], s[8:11], 0x1234 glc dlc
 // GFX10: encoding: [0x84,0x40,0x65,0xf4,0x34,0x12,0x00,0xfa]
 
+s_buffer_store_dwordx2 s[2:3], s[8:11], s0 offset:0x12345 glc dlc
+// GFX10: encoding: [0x84,0x40,0x65,0xf4,0x45,0x23,0x01,0x00]
+
 s_buffer_store_dwordx4 s[4:7], s[8:11], s0
 // GFX10: encoding: [0x04,0x01,0x68,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dwordx4 s[8:11], s[8:11], s0
 // GFX10: encoding: [0x04,0x02,0x68,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dwordx4 s[96:99], s[8:11], s0
 // GFX10: encoding: [0x04,0x18,0x68,0xf4,0x00,0x00,0x00,0x00]

 // GFX10: encoding: [0x04,0x01,0x68,0xf4,0x00,0x00,0x00,0xf8]
 
 s_buffer_store_dwordx4 s[4:7], s[8:11], null
 // GFX10: encoding: [0x04,0x01,0x68,0xf4,0x00,0x00,0x00,0xfa]
 
 s_buffer_store_dwordx4 s[4:7], s[8:11], 0x0
 // GFX10: encoding: [0x04,0x01,0x68,0xf4,0x00,0x00,0x00,0xfa]
 
+s_buffer_store_dwordx4 s[4:7], s[8:11], s0 offset:0x12345
+// GFX10: encoding: [0x04,0x01,0x68,0xf4,0x45,0x23,0x01,0x00]
+
 s_buffer_store_dwordx4 s[4:7], s[8:11], s0 glc
 // GFX10: encoding: [0x04,0x01,0x69,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dwordx4 s[4:7], s[8:11], s0 dlc
 // GFX10: encoding: [0x04,0x41,0x68,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dwordx4 s[4:7], s[8:11], s0 glc dlc
 // GFX10: encoding: [0x04,0x41,0x69,0xf4,0x00,0x00,0x00,0x00]
 
 s_buffer_store_dwordx4 s[4:7], s[8:11], 0x1234 glc dlc
 // GFX10: encoding: [0x04,0x41,0x69,0xf4,0x34,0x12,0x00,0xfa]
 
+s_buffer_store_dwordx4 s[4:7], s[8:11], s0 offset:0x12345 glc dlc
+// GFX10: encoding: [0x04,0x41,0x69,0xf4,0x45,0x23,0x01,0x00]
+
 s_memrealtime s[10:11]
 // GFX10: encoding: [0x80,0x02,0x94,0xf4,0x00,0x00,0x00,0x00]
 
 s_memrealtime s[12:13]
 // GFX10: encoding: [0x00,0x03,0x94,0xf4,0x00,0x00,0x00,0x00]
 
 s_memrealtime s[100:101]
 // GFX10: encoding: [0x00,0x19,0x94,0xf4,0x00,0x00,0x00,0x00]

 // GFX10: encoding: [0x02,0x05,0x1c,0xf4,0x00,0x00,0x00,0x00]
 
 s_scratch_store_dword s101, s[4:5], s0
 // GFX10: encoding: [0x42,0x19,0x54,0xf4,0x00,0x00,0x00,0x00]
 
 s_scratch_store_dword s1, s[4:5], 0x123 glc
 // GFX10: encoding: [0x42,0x00,0x55,0xf4,0x23,0x01,0x00,0xfa]
 
+s_scratch_store_dword s1, s[4:5], s0 offset:0x12345 glc
+// GFX10: encoding: [0x42,0x00,0x55,0xf4,0x45,0x23,0x01,0x00]
+
 s_scratch_store_dwordx2 s[2:3], s[4:5], s101 glc
 // GFX10: encoding: [0x82,0x00,0x59,0xf4,0x00,0x00,0x00,0xca]
 
 s_scratch_store_dwordx4 s[4:7], s[4:5], s0 glc
 // GFX10: encoding: [0x02,0x01,0x5d,0xf4,0x00,0x00,0x00,0x00]
 
 s_atc_probe 0x7, s[4:5], s2
 // GFX10: s_atc_probe 7, s[4:5], s2 ; encoding: [0xc2,0x01,0x98,0xf4,0x00,0x00,0x00,0x04]

llvm/test/MC/Disassembler/AMDGPU/gfx10_dasm_all.txt

 0x46,0x00,0x60,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_buffer_store_dword s1, s[8:11], null  ; encoding: [0x44,0x00,0x60,0xf4,0x00,0x00,0x00,0xfa]
 0x44,0x00,0x60,0xf4,0x00,0x00,0x00,0xfa
 
 # GFX10: s_buffer_store_dword s1, s[8:11], 0x1234 glc dlc ; encoding: [0x44,0x40,0x61,0xf4,0x34,0x12,0x00,0xfa]
 0x44,0x40,0x61,0xf4,0x34,0x12,0x00,0xfa
 
+# GFX10: s_buffer_store_dword s1, s[8:11], s0 offset:0x12345 glc dlc ; encoding: [0x44,0x40,0x61,0xf4,0x45,0x23,0x01,0x00]
+0x44,0x40,0x61,0xf4,0x45,0x23,0x01,0x00
+
 # GFX10: s_buffer_store_dword s1, s[8:11], m0    ; encoding: [0x44,0x00,0x60,0xf4,0x00,0x00,0x00,0xf8]
 0x44,0x00,0x60,0xf4,0x00,0x00,0x00,0xf8
 
 # GFX10: s_buffer_store_dword s1, s[8:11], s0    ; encoding: [0x44,0x00,0x60,0xf4,0x00,0x00,0x00,0x00]
 0x44,0x00,0x60,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_buffer_store_dword s1, s[8:11], s0 dlc ; encoding: [0x44,0x40,0x60,0xf4,0x00,0x00,0x00,0x00]
 0x44,0x40,0x60,0xf4,0x00,0x00,0x00,0x00

 0x86,0x00,0x64,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_buffer_store_dwordx2 s[2:3], s[8:11], null ; encoding: [0x84,0x00,0x64,0xf4,0x00,0x00,0x00,0xfa]
 0x84,0x00,0x64,0xf4,0x00,0x00,0x00,0xfa
 
 # GFX10: s_buffer_store_dwordx2 s[2:3], s[8:11], 0x1234 glc dlc ; encoding: [0x84,0x40,0x65,0xf4,0x34,0x12,0x00,0xfa]
 0x84,0x40,0x65,0xf4,0x34,0x12,0x00,0xfa
 
+# GFX10: s_buffer_store_dwordx2 s[2:3], s[8:11], s0 offset:0x12345 glc dlc ; encoding: [0x84,0x40,0x65,0xf4,0x45,0x23,0x01,0x00]
+0x84,0x40,0x65,0xf4,0x45,0x23,0x01,0x00
+
 # GFX10: s_buffer_store_dwordx2 s[2:3], s[8:11], m0 ; encoding: [0x84,0x00,0x64,0xf4,0x00,0x00,0x00,0xf8]
 0x84,0x00,0x64,0xf4,0x00,0x00,0x00,0xf8
 
 # GFX10: s_buffer_store_dwordx2 s[2:3], s[8:11], s0 ; encoding: [0x84,0x00,0x64,0xf4,0x00,0x00,0x00,0x00]
 0x84,0x00,0x64,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_buffer_store_dwordx2 s[2:3], s[8:11], s0 dlc ; encoding: [0x84,0x40,0x64,0xf4,0x00,0x00,0x00,0x00]
 0x84,0x40,0x64,0xf4,0x00,0x00,0x00,0x00

 0x06,0x01,0x68,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_buffer_store_dwordx4 s[4:7], s[8:11], null ; encoding: [0x04,0x01,0x68,0xf4,0x00,0x00,0x00,0xfa]
 0x04,0x01,0x68,0xf4,0x00,0x00,0x00,0xfa
 
 # GFX10: s_buffer_store_dwordx4 s[4:7], s[8:11], 0x1234 glc dlc ; encoding: [0x04,0x41,0x69,0xf4,0x34,0x12,0x00,0xfa]
 0x04,0x41,0x69,0xf4,0x34,0x12,0x00,0xfa
 
+# GFX10: s_buffer_store_dwordx4 s[4:7], s[8:11], s0 offset:0x12345 glc dlc ; encoding: [0x04,0x41,0x69,0xf4,0x45,0x23,0x01,0x00]
+0x04,0x41,0x69,0xf4,0x45,0x23,0x01,0x00
+
 # GFX10: s_buffer_store_dwordx4 s[4:7], s[8:11], m0 ; encoding: [0x04,0x01,0x68,0xf4,0x00,0x00,0x00,0xf8]
 0x04,0x01,0x68,0xf4,0x00,0x00,0x00,0xf8
 
 # GFX10: s_buffer_store_dwordx4 s[4:7], s[8:11], s0 ; encoding: [0x04,0x01,0x68,0xf4,0x00,0x00,0x00,0x00]
 0x04,0x01,0x68,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_buffer_store_dwordx4 s[4:7], s[8:11], s0 dlc ; encoding: [0x04,0x41,0x68,0xf4,0x00,0x00,0x00,0x00]
 0x04,0x41,0x68,0xf4,0x00,0x00,0x00,0x00

 0x81,0x02,0x19,0xf4,0x01,0x00,0x00,0xfa
 
 # GFX10: s_scratch_load_dwordx4 s[20:23], s[4:5], s0 ; encoding: [0x02,0x05,0x1c,0xf4,0x00,0x00,0x00,0x00]
 0x02,0x05,0x1c,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_scratch_store_dword s1, s[4:5], 0x123 glc ; encoding: [0x42,0x00,0x55,0xf4,0x23,0x01,0x00,0xfa]
 0x42,0x00,0x55,0xf4,0x23,0x01,0x00,0xfa
 
+# GFX10: s_scratch_store_dword s1, s[4:5], s0 offset:0x12345 glc ; encoding: [0x42,0x00,0x55,0xf4,0x45,0x23,0x01,0x00]
+0x42,0x00,0x55,0xf4,0x45,0x23,0x01,0x00
+
 # GFX10: s_scratch_store_dword s101, s[4:5], s0  ; encoding: [0x42,0x19,0x54,0xf4,0x00,0x00,0x00,0x00]
 0x42,0x19,0x54,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_scratch_store_dwordx2 s[2:3], s[4:5], s101 glc ; encoding: [0x82,0x00,0x59,0xf4,0x00,0x00,0x00,0xca]
 0x82,0x00,0x59,0xf4,0x00,0x00,0x00,0xca
 
 # GFX10: s_scratch_store_dwordx4 s[4:7], s[4:5], s0 glc ; encoding: [0x02,0x01,0x5d,0xf4,0x00,0x00,0x00,0x00]
 0x02,0x01,0x5d,0xf4,0x00,0x00,0x00,0x00

 0x72,0x00,0x40,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_store_dword s1, s[4:5], null          ; encoding: [0x42,0x00,0x40,0xf4,0x00,0x00,0x00,0xfa]
 0x42,0x00,0x40,0xf4,0x00,0x00,0x00,0xfa
 
 # GFX10: s_store_dword s1, s[4:5], 0x1234 glc dlc ; encoding: [0x42,0x40,0x41,0xf4,0x34,0x12,0x00,0xfa]
 0x42,0x40,0x41,0xf4,0x34,0x12,0x00,0xfa
 
+# GFX10: s_store_dword s1, s[4:5], s0 offset:0x12345 glc dlc ; encoding: [0x42,0x40,0x41,0xf4,0x45,0x23,0x01,0x00]
+0x42,0x40,0x41,0xf4,0x45,0x23,0x01,0x00
+
 # GFX10: s_store_dword s1, s[4:5], m0            ; encoding: [0x42,0x00,0x40,0xf4,0x00,0x00,0x00,0xf8]
 0x42,0x00,0x40,0xf4,0x00,0x00,0x00,0xf8
 
 # GFX10: s_store_dword s1, s[4:5], s0            ; encoding: [0x42,0x00,0x40,0xf4,0x00,0x00,0x00,0x00]
 0x42,0x00,0x40,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_store_dword s1, s[4:5], s0 dlc        ; encoding: [0x42,0x40,0x40,0xf4,0x00,0x00,0x00,0x00]
 0x42,0x40,0x40,0xf4,0x00,0x00,0x00,0x00

 0xb2,0x00,0x44,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_store_dwordx2 s[2:3], s[4:5], null    ; encoding: [0x82,0x00,0x44,0xf4,0x00,0x00,0x00,0xfa]
 0x82,0x00,0x44,0xf4,0x00,0x00,0x00,0xfa
 
 # GFX10: s_store_dwordx2 s[2:3], s[4:5], 0x1234 glc dlc ; encoding: [0x82,0x40,0x45,0xf4,0x34,0x12,0x00,0xfa]
 0x82,0x40,0x45,0xf4,0x34,0x12,0x00,0xfa
 
+# GFX10: s_store_dwordx2 s[2:3], s[4:5], s0 offset:0x12345 glc dlc ; encoding: [0x82,0x40,0x45,0xf4,0x45,0x23,0x01,0x00]
+0x82,0x40,0x45,0xf4,0x45,0x23,0x01,0x00
+
 # GFX10: s_store_dwordx2 s[2:3], s[4:5], m0      ; encoding: [0x82,0x00,0x44,0xf4,0x00,0x00,0x00,0xf8]
 0x82,0x00,0x44,0xf4,0x00,0x00,0x00,0xf8
 
 # GFX10: s_store_dwordx2 s[2:3], s[4:5], s0      ; encoding: [0x82,0x00,0x44,0xf4,0x00,0x00,0x00,0x00]
 0x82,0x00,0x44,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_store_dwordx2 s[2:3], s[4:5], s0 dlc  ; encoding: [0x82,0x40,0x44,0xf4,0x00,0x00,0x00,0x00]
 0x82,0x40,0x44,0xf4,0x00,0x00,0x00,0x00

 0x32,0x01,0x48,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_store_dwordx4 s[4:7], s[4:5], null    ; encoding: [0x02,0x01,0x48,0xf4,0x00,0x00,0x00,0xfa]
 0x02,0x01,0x48,0xf4,0x00,0x00,0x00,0xfa
 
 # GFX10: s_store_dwordx4 s[4:7], s[4:5], 0x1234 glc dlc ; encoding: [0x02,0x41,0x49,0xf4,0x34,0x12,0x00,0xfa]
 0x02,0x41,0x49,0xf4,0x34,0x12,0x00,0xfa
 
+# GFX10: s_store_dwordx4 s[4:7], s[4:5], s0 offset:0x12345 glc dlc ; encoding: [0x02,0x41,0x49,0xf4,0x45,0x23,0x01,0x00]
+0x02,0x41,0x49,0xf4,0x45,0x23,0x01,0x00
+
 # GFX10: s_store_dwordx4 s[4:7], s[4:5], m0      ; encoding: [0x02,0x01,0x48,0xf4,0x00,0x00,0x00,0xf8]
 0x02,0x01,0x48,0xf4,0x00,0x00,0x00,0xf8
 
 # GFX10: s_store_dwordx4 s[4:7], s[4:5], s0      ; encoding: [0x02,0x01,0x48,0xf4,0x00,0x00,0x00,0x00]
 0x02,0x01,0x48,0xf4,0x00,0x00,0x00,0x00
 
 # GFX10: s_store_dwordx4 s[4:7], s[4:5], s0 dlc  ; encoding: [0x02,0x41,0x48,0xf4,0x00,0x00,0x00,0x00]
 0x02,0x41,0x48,0xf4,0x00,0x00,0x00,0x00