AMDGPU/NFC: Remove some bits from TSFlags (#81525)

- AMDGPU/NFC: Purge SOPK_ZEXT from TSFlags
  - Moved to helper function in SIInstInfo
- AMDGPU/NFC: Purge VOPAsmPrefer32Bit from TSFlags
  - This flag did not make sense / remnants of something else I think

Author: kzhuravl

Diff for: llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

@@ -3299,11 +3299,6 @@ unsigned AMDGPUAsmParser::checkTargetMatchPredicate(MCInst &Inst) {
       (isForcedSDWA() && !(TSFlags & SIInstrFlags::SDWA)) )
     return Match_InvalidOperand;
 
-  if ((TSFlags & SIInstrFlags::VOP3) &&
-      (TSFlags & SIInstrFlags::VOPAsmPrefer32Bit) &&
-      getForcedEncodingSize() != 64)
-    return Match_PreferE32;
-
   if (Inst.getOpcode() == AMDGPU::V_MAC_F32_sdwa_vi ||
       Inst.getOpcode() == AMDGPU::V_MAC_F16_sdwa_vi) {
     // v_mac_f32/16 allow only dst_sel == DWORD;

Diff for: llvm/lib/Target/AMDGPU/SIDefines.h

@@ -105,10 +105,16 @@ enum : uint64_t {
   WQM = UINT64_C(1) << 35,
   DisableWQM = UINT64_C(1) << 36,
   Gather4 = UINT64_C(1) << 37,
-  SOPK_ZEXT = UINT64_C(1) << 38,
+
+  // Reserved, must be 0.
+  Reserved0 = UINT64_C(1) << 38,
+
   SCALAR_STORE = UINT64_C(1) << 39,
   FIXED_SIZE = UINT64_C(1) << 40,
-  VOPAsmPrefer32Bit = UINT64_C(1) << 41,
+
+  // Reserved, must be 0.
+  Reserved1 = UINT64_C(1) << 41,
+
   VOP3_OPSEL = UINT64_C(1) << 42,
   maybeAtomic = UINT64_C(1) << 43,
   renamedInGFX9 = UINT64_C(1) << 44,

Diff for: llvm/lib/Target/AMDGPU/SIInstrFormats.td

@@ -69,10 +69,6 @@ class InstSI <dag outs, dag ins, string asm = "",
 
   field bit Gather4 = 0;
 
-  // Most sopk treat the immediate as a signed 16-bit, however some
-  // use it as unsigned.
-  field bit SOPKZext = 0;
-
   // This is an s_store_dword* instruction that requires a cache flush
   // on wave termination. It is necessary to distinguish from mayStore
   // SMEM instructions like the cache flush ones.
@@ -82,10 +78,6 @@ class InstSI <dag outs, dag ins, string asm = "",
   // instruction size.
   field bit FixedSize = 0;
 
-  // This bit tells the assembler to use the 32-bit encoding in case it
-  // is unable to infer the encoding from the operands.
-  field bit VOPAsmPrefer32Bit = 0;
-
   // This bit indicates that this is a VOP3 opcode which supports op_sel
   // modifier.
   field bit VOP3_OPSEL = 0;
@@ -209,10 +201,15 @@ class InstSI <dag outs, dag ins, string asm = "",
   let TSFlags{36} = DisableWQM;
   let TSFlags{37} = Gather4;
 
-  let TSFlags{38} = SOPKZext;
+  // Reserved, must be 0.
+  let TSFlags{38} = 0;
+
   let TSFlags{39} = ScalarStore;
   let TSFlags{40} = FixedSize;
-  let TSFlags{41} = VOPAsmPrefer32Bit;
+
+  // Reserved, must be 0.
+  let TSFlags{41} = 0;
+
   let TSFlags{42} = VOP3_OPSEL;
 
   let TSFlags{43} = maybeAtomic;

Diff for: llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

@@ -4918,7 +4918,7 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
       }
     } else {
       uint64_t Imm = Op->getImm();
-      if (sopkIsZext(MI)) {
+      if (sopkIsZext(Opcode)) {
         if (!isUInt<16>(Imm)) {
           ErrInfo = "invalid immediate for SOPK instruction";
           return false;

Diff for: llvm/lib/Target/AMDGPU/SIInstrInfo.h

@@ -842,12 +842,13 @@ class SIInstrInfo final : public AMDGPUGenInstrInfo {
     return MI.getDesc().TSFlags & SIInstrFlags::LGKM_CNT;
   }
 
-  static bool sopkIsZext(const MachineInstr &MI) {
-    return MI.getDesc().TSFlags & SIInstrFlags::SOPK_ZEXT;
-  }
-
-  bool sopkIsZext(uint16_t Opcode) const {
-    return get(Opcode).TSFlags & SIInstrFlags::SOPK_ZEXT;
+  // Most sopk treat the immediate as a signed 16-bit, however some
+  // use it as unsigned.
+  static bool sopkIsZext(unsigned Opcode) {
+    return Opcode == AMDGPU::S_CMPK_EQ_U32 || Opcode == AMDGPU::S_CMPK_LG_U32 ||
+           Opcode == AMDGPU::S_CMPK_GT_U32 || Opcode == AMDGPU::S_CMPK_GE_U32 ||
+           Opcode == AMDGPU::S_CMPK_LT_U32 || Opcode == AMDGPU::S_CMPK_LE_U32 ||
+           Opcode == AMDGPU::S_GETREG_B32;
   }
 
   /// \returns true if this is an s_store_dword* instruction. This is more

Diff for: llvm/lib/Target/AMDGPU/SIShrinkInstructions.cpp

@@ -251,9 +251,9 @@ void SIShrinkInstructions::shrinkScalarCompare(MachineInstr &MI) const {
 
   const MCInstrDesc &NewDesc = TII->get(SOPKOpc);
 
-  if ((TII->sopkIsZext(SOPKOpc) && isKUImmOperand(Src1)) ||
-      (!TII->sopkIsZext(SOPKOpc) && isKImmOperand(Src1))) {
-    if (!TII->sopkIsZext(SOPKOpc))
+  if ((SIInstrInfo::sopkIsZext(SOPKOpc) && isKUImmOperand(Src1)) ||
+      (!SIInstrInfo::sopkIsZext(SOPKOpc) && isKImmOperand(Src1))) {
+    if (!SIInstrInfo::sopkIsZext(SOPKOpc))
       Src1.setImm(SignExtend64(Src1.getImm(), 32));
     MI.setDesc(NewDesc);
   }

Diff for: llvm/lib/Target/AMDGPU/SOPInstructions.td

@@ -1078,14 +1078,12 @@ def S_CMPK_GE_I32 : SOPK_SCC <"s_cmpk_ge_i32", "s_cmp_ge_i32", 1>;
 def S_CMPK_LT_I32 : SOPK_SCC <"s_cmpk_lt_i32", "s_cmp_lt_i32", 1>;
 def S_CMPK_LE_I32 : SOPK_SCC <"s_cmpk_le_i32", "s_cmp_le_i32", 1>;
 
-let SOPKZext = 1 in {
 def S_CMPK_EQ_U32 : SOPK_SCC <"s_cmpk_eq_u32", "s_cmp_eq_u32", 0>;
 def S_CMPK_LG_U32 : SOPK_SCC <"s_cmpk_lg_u32", "s_cmp_lg_u32", 0>;
 def S_CMPK_GT_U32 : SOPK_SCC <"s_cmpk_gt_u32", "s_cmp_gt_u32", 0>;
 def S_CMPK_GE_U32 : SOPK_SCC <"s_cmpk_ge_u32", "s_cmp_ge_u32", 0>;
 def S_CMPK_LT_U32 : SOPK_SCC <"s_cmpk_lt_u32", "s_cmp_lt_u32", 0>;
 def S_CMPK_LE_U32 : SOPK_SCC <"s_cmpk_le_u32", "s_cmp_le_u32", 0>;
-} // End SOPKZext = 1
 } // End isCompare = 1
 
 let isCommutable = 1, DisableEncoding = "$src0",
@@ -1111,7 +1109,6 @@ def S_GETREG_B32 : SOPK_Pseudo <
   (outs SReg_32:$sdst), (ins hwreg:$simm16),
   "$sdst, $simm16",
   [(set i32:$sdst, (int_amdgcn_s_getreg (i32 timm:$simm16)))]> {
-  let SOPKZext = 1;
   let hasSideEffects = 1;
 }
 

Diff for: llvm/lib/Target/AMDGPU/VOP1Instructions.td

@@ -217,9 +217,7 @@ def VOP_I16_F16_SPECIAL_OMOD_t16 : VOPProfile_Fake16<VOP_I16_F16> {
 // VOP1 Instructions
 //===----------------------------------------------------------------------===//
 
-let VOPAsmPrefer32Bit = 1 in {
 defm V_NOP : VOP1Inst <"v_nop", VOP_NOP_PROFILE>;
-}
 
 def VOPProfile_MOV : VOPProfile <[i32, i32, untyped, untyped]> {
   let InsVOPDX = (ins Src0RC32:$src0X);
@@ -368,9 +366,7 @@ defm V_FREXP_EXP_I32_F32 : VOP1Inst <"v_frexp_exp_i32_f32", VOP_I32_F32, int_amd
 defm V_FREXP_MANT_F32 : VOP1Inst <"v_frexp_mant_f32", VOP_F32_F32, int_amdgcn_frexp_mant>;
 } // End isReMaterializable = 1
 
-let VOPAsmPrefer32Bit = 1 in {
 defm V_CLREXCP : VOP1Inst <"v_clrexcp", VOP_NO_EXT<VOP_NONE>>;
-}
 
 // Restrict src0 to be VGPR
 def VOP_MOVRELS : VOPProfile<[i32, i32, untyped, untyped]> {

Diff for: llvm/lib/Target/AMDGPU/VOPInstructions.td

@@ -16,7 +16,6 @@ class LetDummies {
   bit isMoveImm;
   bit isReMaterializable;
   bit isAsCheapAsAMove;
-  bit VOPAsmPrefer32Bit;
   bit FPDPRounding;
   Predicate SubtargetPredicate;
   string Constraints;