[AMDGPU] MC support for gfx1250 scale_offset modifier (#149881)

rampitec · web-flow · commit a0b854d576c8 · 2025-07-21T15:04:59.000-07:00
diff --git a/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp b/llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp
@@ -5324,6 +5324,12 @@ bool AMDGPUAsmParser::validateCoherencyBits(const MCInst &Inst,
   unsigned CPol = Inst.getOperand(CPolPos).getImm();
 
   if (!isGFX1250()) {
+    if (CPol & CPol::SCAL) {
+      SMLoc S = getImmLoc(AMDGPUOperand::ImmTyCPol, Operands);
+      StringRef CStr(S.getPointer());
+      S = SMLoc::getFromPointer(&CStr.data()[CStr.find("scale_offset")]);
+      Error(S, "scale_offset is not supported on this GPU");
+    }
     if (CPol & CPol::NV) {
       SMLoc S = getImmLoc(AMDGPUOperand::ImmTyCPol, Operands);
       StringRef CStr(S.getPointer());
@@ -5332,6 +5338,13 @@ bool AMDGPUAsmParser::validateCoherencyBits(const MCInst &Inst,
     }
   }
 
+  if ((CPol & CPol::SCAL) && !supportsScaleOffset(MII, Inst.getOpcode())) {
+    SMLoc S = getImmLoc(AMDGPUOperand::ImmTyCPol, Operands);
+    StringRef CStr(S.getPointer());
+    S = SMLoc::getFromPointer(&CStr.data()[CStr.find("scale_offset")]);
+    Error(S, "scale_offset is not supported for this instruction");
+  }
+
   if (isGFX12Plus())
     return validateTHAndScopeBits(Inst, Operands, CPol);
 
@@ -7003,6 +7016,7 @@ ParseStatus AMDGPUAsmParser::parseCPol(OperandVector &Operands) {
     ParseStatus ResTH = ParseStatus::NoMatch;
     ParseStatus ResScope = ParseStatus::NoMatch;
     ParseStatus ResNV = ParseStatus::NoMatch;
+    ParseStatus ResScal = ParseStatus::NoMatch;
 
     for (;;) {
       if (ResTH.isNoMatch()) {
@@ -7041,10 +7055,22 @@ ParseStatus AMDGPUAsmParser::parseCPol(OperandVector &Operands) {
         }
       }
 
+      if (ResScal.isNoMatch()) {
+        if (trySkipId("scale_offset")) {
+          ResScal = ParseStatus::Success;
+          CPolVal |= CPol::SCAL;
+          continue;
+        } else if (trySkipId("no", "scale_offset")) {
+          ResScal = ParseStatus::Success;
+          continue;
+        }
+      }
+
       break;
     }
 
-    if (ResTH.isNoMatch() && ResScope.isNoMatch() && ResNV.isNoMatch())
+    if (ResTH.isNoMatch() && ResScope.isNoMatch() && ResNV.isNoMatch() &&
+        ResScal.isNoMatch())
       return ParseStatus::NoMatch;
 
     Operands.push_back(AMDGPUOperand::CreateImm(this, CPolVal, StringLoc,
diff --git a/llvm/lib/Target/AMDGPU/FLATInstructions.td b/llvm/lib/Target/AMDGPU/FLATInstructions.td
@@ -2941,6 +2941,7 @@ multiclass VFLAT_Real_gfx12 <bits<8> op, string name = get_FLAT_ps<NAME>.Mnemoni
     let DecoderNamespace = "GFX12";
 
     let Inst{25-24} = {ps.is_flat_global, ps.is_flat_scratch};
+    let Inst{48} = cpol{CPolBit.SCAL}; // scale offset
   }
 }
 
@@ -3170,6 +3171,7 @@ multiclass VFLAT_Real_gfx1250<bits<8> op,
     let DecoderNamespace = "GFX1250";
 
     let Inst{25-24} = {ps.is_flat_global, ps.is_flat_scratch};
+    let Inst{48} = cpol{CPolBit.SCAL}; // scale offset
   }
 }
 
diff --git a/llvm/lib/Target/AMDGPU/GCNSubtarget.h b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
@@ -1162,6 +1162,9 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
 
   bool hasLshlAddU64Inst() const { return HasLshlAddU64Inst; }
 
+  // Scalar and global loads support scale_offset bit.
+  bool hasScaleOffset() const { return GFX1250Insts; }
+
   bool hasFlatGVSMode() const { return FlatGVSMode; }
 
   bool enableSIScheduler() const {
diff --git a/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp b/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUInstPrinter.cpp
@@ -157,6 +157,9 @@ void AMDGPUInstPrinter::printCPol(const MCInst *MI, unsigned OpNo,
     const int64_t TH = Imm & CPol::TH;
     const int64_t Scope = Imm & CPol::SCOPE;
 
+    if (Imm & CPol::SCAL)
+      O << " scale_offset";
+
     printTH(MI, TH, Scope, O);
     printScope(Scope, O);
 
diff --git a/llvm/lib/Target/AMDGPU/SIDefines.h b/llvm/lib/Target/AMDGPU/SIDefines.h
@@ -402,6 +402,8 @@ enum CPol {
 
   SWZ = 1 << 6, // Swizzle bit
 
+  SCAL = 1 << 11, // Scale offset bit
+
   ALL = TH | SCOPE,
 
   // Helper bits
diff --git a/llvm/lib/Target/AMDGPU/SIInstrFormats.td b/llvm/lib/Target/AMDGPU/SIInstrFormats.td
@@ -318,6 +318,7 @@ def CPolBit {
   int DLC = 2;
   int SCC = 4;
   int NV = 5;
+  int SCAL = 11;
 }
 
 class VOPDstOperand <RegisterClass rc> : RegisterOperand <rc, "printVOPDst">;
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -5482,6 +5482,19 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
     }
   }
 
+  if (const MachineOperand *CPol = getNamedOperand(MI, AMDGPU::OpName::cpol)) {
+    if (CPol->getImm() & AMDGPU::CPol::SCAL) {
+      if (!ST.hasScaleOffset()) {
+        ErrInfo = "Subtarget does not support offset scaling";
+        return false;
+      }
+      if (!AMDGPU::supportsScaleOffset(*this, MI.getOpcode())) {
+        ErrInfo = "Instruction does not support offset scaling";
+        return false;
+      }
+    }
+  }
+
   return true;
 }
 
diff --git a/llvm/lib/Target/AMDGPU/SMInstructions.td b/llvm/lib/Target/AMDGPU/SMInstructions.td
@@ -1488,6 +1488,7 @@ class SMEM_Real_Load_gfx12<bits<6> op, string ps, string opName, OffsetMode offs
   let Inst{20} = cpol{CPolBit.NV}; // non-volatile
   let Inst{22-21} = cpol{4-3}; // scope
   let Inst{24-23} = cpol{1-0}; // th - only lower 2 bits are supported
+  let Inst{56} = cpol{CPolBit.SCAL}; // scale offset
 }
 
 multiclass SM_Real_Loads_gfx12<bits<6> op, string ps = NAME> {
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp
@@ -3228,6 +3228,25 @@ const GcnBufferFormatInfo *getGcnBufferFormatInfo(uint8_t Format,
                           : getGfx9BufferFormatInfo(Format);
 }
 
+bool supportsScaleOffset(const MCInstrInfo &MII, unsigned Opcode) {
+  uint64_t TSFlags = MII.get(Opcode).TSFlags;
+
+  if (TSFlags & SIInstrFlags::SMRD)
+    return !getSMEMIsBuffer(Opcode);
+  if (!(TSFlags & SIInstrFlags::FLAT))
+    return false;
+
+  // Only SV and SVS modes are supported.
+  if (TSFlags & SIInstrFlags::FlatScratch)
+    return hasNamedOperand(Opcode, OpName::vaddr);
+
+  // Only GVS mode is supported.
+  return hasNamedOperand(Opcode, OpName::vaddr) &&
+         hasNamedOperand(Opcode, OpName::saddr);
+
+  return false;
+}
+
 bool hasAny64BitVGPROperands(const MCInstrDesc &OpDesc) {
   for (auto OpName : {OpName::vdst, OpName::src0, OpName::src1, OpName::src2}) {
     int Idx = getNamedOperandIdx(OpDesc.getOpcode(), OpName);
diff --git a/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h b/llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.h
@@ -1757,6 +1757,9 @@ bool isIntrinsicSourceOfDivergence(unsigned IntrID);
 /// \returns true if the intrinsic is uniform
 bool isIntrinsicAlwaysUniform(unsigned IntrID);
 
+/// \returns true if a memory instruction supports scale_offset modifier.
+bool supportsScaleOffset(const MCInstrInfo &MII, unsigned Opcode);
+
 /// \returns lds block size in terms of dwords. \p
 /// This is used to calculate the lds size encoded for PAL metadata 3.0+ which
 /// must be defined in terms of bytes.
diff --git a/llvm/test/MC/AMDGPU/gfx1250_asm_smem.s b/llvm/test/MC/AMDGPU/gfx1250_asm_smem.s
@@ -12,3 +12,30 @@ s_buffer_load_i8 s5, s[4:7], s0 nv
 // GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: nv is not supported on this GPU
 // GFX12-ERR-NEXT:{{^}}s_buffer_load_i8 s5, s[4:7], s0 nv
 // GFX12-ERR-NEXT:{{^}}                                ^
+
+s_load_b32 s4, s[2:3], 0xa scale_offset
+// GFX1250: s_load_b32 s4, s[2:3], 0xa scale_offset ; encoding: [0x01,0x01,0x00,0xf4,0x0a,0x00,0x00,0xf9]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: scale_offset is not supported on this GPU
+// GFX12-ERR-NEXT:{{^}}s_load_b32 s4, s[2:3], 0xa scale_offset
+// GFX12-ERR-NEXT:{{^}}                           ^
+
+s_load_b32 s4, s[2:3], 0xa scale_offset nv
+// GFX1250: s_load_b32 s4, s[2:3], 0xa scale_offset nv ; encoding: [0x01,0x01,0x10,0xf4,0x0a,0x00,0x00,0xf9]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: scale_offset is not supported on this GPU
+// GFX12-ERR-NEXT:{{^}}s_load_b32 s4, s[2:3], 0xa scale_offset nv
+// GFX12-ERR-NEXT:{{^}}                           ^
+// GFX12-ERR-NEXT: error: nv is not supported on this GPU
+// GFX12-ERR-NEXT:{{^}}s_load_b32 s4, s[2:3], 0xa scale_offset nv
+// GFX12-ERR-NEXT:{{^}}                                        ^
+
+s_load_b32 s4, s[2:3], s5 offset:32 scale_offset
+// GFX1250: s_load_b32 s4, s[2:3], s5 offset:0x20 scale_offset ; encoding: [0x01,0x01,0x00,0xf4,0x20,0x00,0x00,0x0b]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: scale_offset is not supported on this GPU
+// GFX12-ERR-NEXT:{{^}}s_load_b32 s4, s[2:3], s5 offset:32 scale_offset
+// GFX12-ERR-NEXT:{{^}}                                    ^
+
+s_load_b32 s4, s[2:3], m0 offset:32 scale_offset
+// GFX1250: s_load_b32 s4, s[2:3], m0 offset:0x20 scale_offset ; encoding: [0x01,0x01,0x00,0xf4,0x20,0x00,0x00,0xfb]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: scale_offset is not supported on this GPU
+// GFX12-ERR-NEXT:{{^}}s_load_b32 s4, s[2:3], m0 offset:32 scale_offset
+// GFX12-ERR-NEXT:{{^}}                                    ^
diff --git a/llvm/test/MC/AMDGPU/gfx1250_asm_smem_err.s b/llvm/test/MC/AMDGPU/gfx1250_asm_smem_err.s
@@ -0,0 +1,16 @@
+// RUN: not llvm-mc -triple=amdgcn -mcpu=gfx1250 -show-encoding %s 2>&1 | FileCheck --check-prefix=GFX1250-ERR --implicit-check-not=error: --strict-whitespace %s
+
+s_buffer_load_i8 s5, s[4:7], s0 scale_offset
+// GFX1250-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: scale_offset is not supported for this instruction
+// GFX1250-ERR-NEXT:{{^}}s_buffer_load_i8 s5, s[4:7], s0 scale_offset
+// GFX1250-ERR-NEXT:{{^}}                                ^
+
+s_prefetch_data s[18:19], 100, s10, 7 nv
+// GFX1250-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: invalid operand for instruction
+// GFX1250-ERR-NEXT:{{^}}s_prefetch_data s[18:19], 100, s10, 7 nv
+// GFX1250-ERR-NEXT:{{^}}                                      ^
+
+s_prefetch_data s[18:19], 100, s10, 7 scale_offset
+// GFX1250-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: invalid operand for instruction
+// GFX1250-ERR-NEXT:{{^}}s_prefetch_data s[18:19], 100, s10, 7 scale_offset
+// GFX1250-ERR-NEXT:{{^}}                                      ^
diff --git a/llvm/test/MC/AMDGPU/gfx1250_asm_vbuffer_mubuf_err.s b/llvm/test/MC/AMDGPU/gfx1250_asm_vbuffer_mubuf_err.s
@@ -0,0 +1,6 @@
+// RUN: not llvm-mc -triple=amdgcn -mcpu=gfx1250 -show-encoding %s 2>&1 | FileCheck --check-prefix=GFX1250-ERR --implicit-check-not=error: --strict-whitespace %s
+
+buffer_load_b32 v5, v1, s[8:11], s3 offen offset:4095 scale_offset
+// GFX1250-ERR: :[[@LINE-1]]:{{[0-9]+}}: error: scale_offset is not supported for this instruction
+// GFX1250-ERR-NEXT:{{^}}buffer_load_b32 v5, v1, s[8:11], s3 offen offset:4095 scale_offset
+// GFX1250-ERR-NEXT:{{^}}                                                      ^
diff --git a/llvm/test/MC/AMDGPU/gfx1250_asm_vflat.s b/llvm/test/MC/AMDGPU/gfx1250_asm_vflat.s
diff --git a/llvm/test/MC/AMDGPU/gfx1250_asm_vflat_err.s b/llvm/test/MC/AMDGPU/gfx1250_asm_vflat_err.s
diff --git a/llvm/test/MC/Disassembler/AMDGPU/gfx1250_dasm_smem.txt b/llvm/test/MC/Disassembler/AMDGPU/gfx1250_dasm_smem.txt
diff --git a/llvm/test/MC/Disassembler/AMDGPU/gfx1250_dasm_vflat.txt b/llvm/test/MC/Disassembler/AMDGPU/gfx1250_dasm_vflat.txt

Original file line number	Diff line number	Diff line change
`@@ -2941,6 +2941,7 @@ multiclass VFLAT_Real_gfx12 <bits<8> op, string name = get_FLAT_ps<NAME>.Mnemoni`
`2941`	`2941`	`let DecoderNamespace = "GFX12";`
`2942`	`2942`
`2943`	`2943`	`let Inst{25-24} = {ps.is_flat_global, ps.is_flat_scratch};`
	`2944`	`+ let Inst{48} = cpol{CPolBit.SCAL}; // scale offset`
`2944`	`2945`	`}`
`2945`	`2946`	`}`
`2946`	`2947`
`@@ -3170,6 +3171,7 @@ multiclass VFLAT_Real_gfx1250<bits<8> op,`
`3170`	`3171`	`let DecoderNamespace = "GFX1250";`
`3171`	`3172`
`3172`	`3173`	`let Inst{25-24} = {ps.is_flat_global, ps.is_flat_scratch};`
	`3174`	`+ let Inst{48} = cpol{CPolBit.SCAL}; // scale offset`
`3173`	`3175`	`}`
`3174`	`3176`	`}`
`3175`	`3177`
Original file line number	Diff line number	Diff line change
`@@ -318,6 +318,7 @@ def CPolBit {`
`318`	`318`	`int DLC = 2;`
`319`	`319`	`int SCC = 4;`
`320`	`320`	`int NV = 5;`
	`321`	`+ int SCAL = 11;`
`321`	`322`	`}`
`322`	`323`
`323`	`324`	`class VOPDstOperand <RegisterClass rc> : RegisterOperand <rc, "printVOPDst">;`
Original file line number	Diff line number	Diff line change
`@@ -5482,6 +5482,19 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,`
`5482`	`5482`	`}`
`5483`	`5483`	`}`
`5484`	`5484`
	`5485`	`+ if (const MachineOperand *CPol = getNamedOperand(MI, AMDGPU::OpName::cpol)) {`
	`5486`	`+ if (CPol->getImm() & AMDGPU::CPol::SCAL) {`
	`5487`	`+ if (!ST.hasScaleOffset()) {`
	`5488`	`+ ErrInfo = "Subtarget does not support offset scaling";`
	`5489`	`+ return false;`
	`5490`	`+ }`
	`5491`	`+ if (!AMDGPU::supportsScaleOffset(*this, MI.getOpcode())) {`
	`5492`	`+ ErrInfo = "Instruction does not support offset scaling";`
	`5493`	`+ return false;`
	`5494`	`+ }`
	`5495`	`+ }`
	`5496`	`+ }`
	`5497`	`+`
`5485`	`5498`	`return true;`
`5486`	`5499`	`}`
`5487`	`5500`
Original file line number	Diff line number	Diff line change
`@@ -1488,6 +1488,7 @@ class SMEM_Real_Load_gfx12<bits<6> op, string ps, string opName, OffsetMode offs`
`1488`	`1488`	`let Inst{20} = cpol{CPolBit.NV}; // non-volatile`
`1489`	`1489`	`let Inst{22-21} = cpol{4-3}; // scope`
`1490`	`1490`	`let Inst{24-23} = cpol{1-0}; // th - only lower 2 bits are supported`
	`1491`	`+ let Inst{56} = cpol{CPolBit.SCAL}; // scale offset`
`1491`	`1492`	`}`
`1492`	`1493`
`1493`	`1494`	`multiclass SM_Real_Loads_gfx12<bits<6> op, string ps = NAME> {`