[AMDGPU] Add 192-bit register classes

[arsenm]

llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp

1387

Should also AReg_192

1539

Trailing comma

[foad]

Update AMDGPURegisterBanks.td.
Add AReg_192 register class.
Remove trailing comma.

[foad]

Rebase.

[rampitec]

Same notes about missing places as in D78348.

llvm/lib/Target/AMDGPU/SIRegisterInfo.td

754–755

This renumbering becomes troublesome. Maybe we need a policy like AllocationPriority = Size/32.

[rampitec]

llvm/lib/Target/AMDGPU/SIRegisterInfo.td

755

I wander if we should just use for loop to define all the classes at once.

[kerbowa]

AMDGPUGenRegisterBankInfo.def probably should be updated as well.

[foad]

Rebase.
Update AMDGPUAsmPrinter, AMDGPUISelDAGToDAG, AMDGPUAsmParser,
SIMCCodeEmitter and SITargetLowering::getRegForInlineAsmConstraint.

[foad]

In D78312#1990637, @kerbowa wrote:

AMDGPUGenRegisterBankInfo.def probably should be updated as well.

Probably, but it is already missing 160-bit classes. Can we leave it for a separate patch?

[kerbowa]

In D78312#1994240, @foad wrote:

In D78312#1990637, @kerbowa wrote:

AMDGPUGenRegisterBankInfo.def probably should be updated as well.

Probably, but it is already missing 160-bit classes. Can we leave it for a separate patch?

Sounds good.

[rampitec]

LGTM

llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp

for (const MachineInstr &MI : MBB) {
           Width = 5;
         } else if (AMDGPU::SReg_160RegClass.contains(Reg)) {
           IsSGPR = true;
           Width = 5;
         } else if (AMDGPU::AReg_160RegClass.contains(Reg)) {
           IsSGPR = false;
           IsAGPR = true;
           Width = 5;
+        } else if (AMDGPU::VReg_192RegClass.contains(Reg)) {
+          IsSGPR = false;
+          Width = 6;
+        } else if (AMDGPU::SReg_192RegClass.contains(Reg)) {
+          IsSGPR = true;
+          Width = 6;
+        } else if (AMDGPU::AReg_192RegClass.contains(Reg)) {
+          IsSGPR = false;
+          IsAGPR = true;
+          Width = 6;
         } else if (AMDGPU::SReg_256RegClass.contains(Reg)) {
           assert(!AMDGPU::TTMP_256RegClass.contains(Reg) &&
             "trap handler registers should not be used");
           IsSGPR = true;
           Width = 8;
         } else if (AMDGPU::VReg_256RegClass.contains(Reg)) {
           IsSGPR = false;
           Width = 8;

llvm/lib/Target/AMDGPU/AMDGPURegisterBanks.td

 //=- AMDGPURegisterBank.td - Describe the AMDGPU Banks -------*- tablegen -*-=//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 
 def SGPRRegBank : RegisterBank<"SGPR",
-  [SGPR_LO16, SReg_32, SReg_64, SReg_128, SReg_256, SReg_512, SReg_1024]
+  [SGPR_LO16, SReg_32, SReg_64, SReg_128, SReg_160, SReg_192, SReg_256, SReg_512, SReg_1024]
 >;
 
 def VGPRRegBank : RegisterBank<"VGPR",
-  [VGPR_LO16, VGPR_HI16, VGPR_32, VReg_64, VReg_96, VReg_128, VReg_256, VReg_512, VReg_1024]
+  [VGPR_LO16, VGPR_HI16, VGPR_32, VReg_64, VReg_96, VReg_128, VReg_160, VReg_192, VReg_256, VReg_512, VReg_1024]
 >;
 
 // It is helpful to distinguish conditions from ordinary SGPRs.
 def VCCRegBank : RegisterBank <"VCC", [SReg_1]>;
 
 def AGPRRegBank : RegisterBank <"AGPR",
-  [AGPR_32, AReg_64, AReg_96, AReg_128, AReg_160, AReg_256, AReg_512, AReg_1024]
+  [AGPR_32, AReg_64, AReg_96, AReg_128, AReg_160, AReg_192, AReg_256, AReg_512, AReg_1024]
 >;

llvm/lib/Target/AMDGPU/AsmParser/AMDGPUAsmParser.cpp

public:
   }
 
   bool isVReg() const {
     return isRegClass(AMDGPU::VGPR_32RegClassID) ||
            isRegClass(AMDGPU::VReg_64RegClassID) ||
            isRegClass(AMDGPU::VReg_96RegClassID) ||
            isRegClass(AMDGPU::VReg_128RegClassID) ||
            isRegClass(AMDGPU::VReg_160RegClassID) ||
+           isRegClass(AMDGPU::VReg_192RegClassID) ||
            isRegClass(AMDGPU::VReg_256RegClassID) ||
            isRegClass(AMDGPU::VReg_512RegClassID) ||
            isRegClass(AMDGPU::VReg_1024RegClassID);
   }
 
   bool isVReg32() const {
     return isRegClass(AMDGPU::VGPR_32RegClassID);
   }
static int getRegClass(RegisterKind Is, unsigned RegWidth) {
   if (Is == IS_VGPR) {
     switch (RegWidth) {
       default: return -1;
       case 1: return AMDGPU::VGPR_32RegClassID;
       case 2: return AMDGPU::VReg_64RegClassID;
       case 3: return AMDGPU::VReg_96RegClassID;
       case 4: return AMDGPU::VReg_128RegClassID;
       case 5: return AMDGPU::VReg_160RegClassID;
+      case 6: return AMDGPU::VReg_192RegClassID;
       case 8: return AMDGPU::VReg_256RegClassID;
       case 16: return AMDGPU::VReg_512RegClassID;
       case 32: return AMDGPU::VReg_1024RegClassID;
     }
   } else if (Is == IS_TTMP) {
     switch (RegWidth) {
       default: return -1;
       case 1: return AMDGPU::TTMP_32RegClassID;
       case 2: return AMDGPU::TTMP_64RegClassID;
       case 4: return AMDGPU::TTMP_128RegClassID;
       case 8: return AMDGPU::TTMP_256RegClassID;
       case 16: return AMDGPU::TTMP_512RegClassID;
     }
   } else if (Is == IS_SGPR) {
     switch (RegWidth) {
       default: return -1;
       case 1: return AMDGPU::SGPR_32RegClassID;
       case 2: return AMDGPU::SGPR_64RegClassID;
+      case 3: return AMDGPU::SGPR_96RegClassID;
       case 4: return AMDGPU::SGPR_128RegClassID;
+      case 5: return AMDGPU::SGPR_160RegClassID;
+      case 6: return AMDGPU::SReg_192RegClassID;
       case 8: return AMDGPU::SGPR_256RegClassID;
       case 16: return AMDGPU::SGPR_512RegClassID;
     }
   } else if (Is == IS_AGPR) {
     switch (RegWidth) {
       default: return -1;
       case 1: return AMDGPU::AGPR_32RegClassID;
       case 2: return AMDGPU::AReg_64RegClassID;
       case 3: return AMDGPU::AReg_96RegClassID;
       case 4: return AMDGPU::AReg_128RegClassID;
       case 5: return AMDGPU::AReg_160RegClassID;
+      case 6: return AMDGPU::AReg_192RegClassID;
       case 8: return AMDGPU::AReg_256RegClassID;
       case 16: return AMDGPU::AReg_512RegClassID;
       case 32: return AMDGPU::AReg_1024RegClassID;
     }
   }
   return -1;
 }
 

llvm/lib/Target/AMDGPU/MCTargetDesc/SIMCCodeEmitter.cpp

SIMCCodeEmitter::getAVOperandEncoding(const MCInst &MI, unsigned OpNo,
   // VGPR and AGPR have the same encoding, but SrcA and SrcB operands of mfma
   // instructions use acc[0:1] modifier bits to distinguish. These bits are
   // encoded as a virtual 9th bit of the register for these operands.
   if (MRI.getRegClass(AMDGPU::AGPR_32RegClassID).contains(Reg) ||
       MRI.getRegClass(AMDGPU::AReg_64RegClassID).contains(Reg) ||
       MRI.getRegClass(AMDGPU::AReg_96RegClassID).contains(Reg) ||
       MRI.getRegClass(AMDGPU::AReg_128RegClassID).contains(Reg) ||
       MRI.getRegClass(AMDGPU::AReg_160RegClassID).contains(Reg) ||
+      MRI.getRegClass(AMDGPU::AReg_192RegClassID).contains(Reg) ||
       MRI.getRegClass(AMDGPU::AReg_256RegClassID).contains(Reg))
     Enc |= 512;
 
   return Enc;
 }
 
 static bool needsPCRel(const MCExpr *Expr) {
   switch (Expr->getKind()) {

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

case 'r':
         RC = &AMDGPU::SReg_96RegClass;
         break;
       case 128:
         RC = &AMDGPU::SGPR_128RegClass;
         break;
       case 160:
         RC = &AMDGPU::SReg_160RegClass;
         break;
+      case 192:
+        RC = &AMDGPU::SReg_192RegClass;
+        break;
       case 256:
         RC = &AMDGPU::SReg_256RegClass;
         break;
       case 512:
         RC = &AMDGPU::SReg_512RegClass;
         break;
       }
       break;
case 'v':
         RC = &AMDGPU::VReg_96RegClass;
         break;
       case 128:
         RC = &AMDGPU::VReg_128RegClass;
         break;
       case 160:
         RC = &AMDGPU::VReg_160RegClass;
         break;
+      case 192:
+        RC = &AMDGPU::VReg_192RegClass;
+        break;
       case 256:
         RC = &AMDGPU::VReg_256RegClass;
         break;
       case 512:
         RC = &AMDGPU::VReg_512RegClass;
         break;
       }
       break;
case 'a':
         RC = &AMDGPU::AReg_96RegClass;
         break;
       case 128:
         RC = &AMDGPU::AReg_128RegClass;
         break;
       case 160:
         RC = &AMDGPU::AReg_160RegClass;
         break;
+      case 192:
+        RC = &AMDGPU::AReg_192RegClass;
+        break;
       case 256:
         RC = &AMDGPU::AReg_256RegClass;
         break;
       case 512:
         RC = &AMDGPU::AReg_512RegClass;
         break;
       case 1024:
         RC = &AMDGPU::AReg_1024RegClass;

llvm/lib/Target/AMDGPU/SIRegisterInfo.cpp

SIRegisterInfo::getVGPRClassForBitWidth(unsigned BitWidth) {
   case 64:
     return &AMDGPU::VReg_64RegClass;
   case 96:
     return &AMDGPU::VReg_96RegClass;
   case 128:
     return &AMDGPU::VReg_128RegClass;
   case 160:
     return &AMDGPU::VReg_160RegClass;
+  case 192:
+    return &AMDGPU::VReg_192RegClass;
   case 256:
     return &AMDGPU::VReg_256RegClass;
   case 512:
     return &AMDGPU::VReg_512RegClass;
   case 1024:
     return &AMDGPU::VReg_1024RegClass;
   default:
     return nullptr;
   }
 }
 
 const TargetRegisterClass *
 SIRegisterInfo::getAGPRClassForBitWidth(unsigned BitWidth) {
   switch (BitWidth) {
   case 32:
     return &AMDGPU::AGPR_32RegClass;
   case 64:
     return &AMDGPU::AReg_64RegClass;
   case 96:
     return &AMDGPU::AReg_96RegClass;
   case 128:
     return &AMDGPU::AReg_128RegClass;
   case 160:
     return &AMDGPU::AReg_160RegClass;
+  case 192:
+    return &AMDGPU::AReg_192RegClass;
   case 256:
     return &AMDGPU::AReg_256RegClass;
   case 512:
     return &AMDGPU::AReg_512RegClass;
   case 1024:
     return &AMDGPU::AReg_1024RegClass;
   default:
     return nullptr;
SIRegisterInfo::getSGPRClassForBitWidth(unsigned BitWidth) {
   case 64:
     return &AMDGPU::SReg_64RegClass;
   case 96:
     return &AMDGPU::SReg_96RegClass;
   case 128:
     return &AMDGPU::SReg_128RegClass;
   case 160:
     return &AMDGPU::SReg_160RegClass;
+  case 192:
+    return &AMDGPU::SReg_192RegClass;
   case 256:
     return &AMDGPU::SReg_256RegClass;
   case 512:
     return &AMDGPU::SReg_512RegClass;
   case 1024:
     return &AMDGPU::SReg_1024RegClass;
   default:
     return nullptr;
static const TargetRegisterClass *const BaseClasses[] = {
     &AMDGPU::SReg_96RegClass,
     &AMDGPU::AReg_96RegClass,
     &AMDGPU::VReg_128RegClass,
     &AMDGPU::SReg_128RegClass,
     &AMDGPU::AReg_128RegClass,
     &AMDGPU::VReg_160RegClass,
     &AMDGPU::SReg_160RegClass,
     &AMDGPU::AReg_160RegClass,
+    &AMDGPU::VReg_192RegClass,
+    &AMDGPU::SReg_192RegClass,
+    &AMDGPU::AReg_192RegClass,
     &AMDGPU::VReg_256RegClass,
     &AMDGPU::SReg_256RegClass,
     &AMDGPU::AReg_256RegClass,
     &AMDGPU::VReg_512RegClass,
     &AMDGPU::SReg_512RegClass,
     &AMDGPU::AReg_512RegClass,
     &AMDGPU::SReg_1024RegClass,
     &AMDGPU::VReg_1024RegClass,
     &AMDGPU::AReg_1024RegClass,
if (MRI.isAllocatable(Reg) && !MRI.isPhysRegUsed(Reg))
       return Reg;
   return MCRegister();
 }
 
 ArrayRef<int16_t> SIRegisterInfo::getRegSplitParts(const TargetRegisterClass *RC,
                                                    unsigned EltSize) const {
   const unsigned RegBitWidth = AMDGPU::getRegBitWidth(*RC->MC);
   assert(RegBitWidth >= 32 && RegBitWidth <= 1024);
 
   const unsigned EltBitWidth = EltSize * 8;
   assert(EltBitWidth >= 32 && EltBitWidth < 1024 && isPowerOf2_32(EltBitWidth));
   const unsigned LogEltBitWidth = Log2_32(EltBitWidth);
 
   assert(RegBitWidth % EltBitWidth == 0);
 
   if (RegBitWidth == EltBitWidth)
     return {};

llvm/lib/Target/AMDGPU/SIRegisterInfo.td

 //  Helpers
 //===----------------------------------------------------------------------===//
 
 class getSubRegs<int size> {
   list<SubRegIndex> ret2 = [sub0, sub1];
   list<SubRegIndex> ret3 = [sub0, sub1, sub2];
   list<SubRegIndex> ret4 = [sub0, sub1, sub2, sub3];
   list<SubRegIndex> ret5 = [sub0, sub1, sub2, sub3, sub4];
+  list<SubRegIndex> ret6 = [sub0, sub1, sub2, sub3, sub4, sub5];
   list<SubRegIndex> ret8 = [sub0, sub1, sub2, sub3, sub4, sub5, sub6, sub7];
   list<SubRegIndex> ret16 = [sub0, sub1, sub2, sub3,
                              sub4, sub5, sub6, sub7,
                              sub8, sub9, sub10, sub11,
                              sub12, sub13, sub14, sub15];
   list<SubRegIndex> ret32 = [sub0, sub1, sub2, sub3,
                              sub4, sub5, sub6, sub7,
                              sub8, sub9, sub10, sub11,
                              sub12, sub13, sub14, sub15,
                              sub16, sub17, sub18, sub19,
                              sub20, sub21, sub22, sub23,
                              sub24, sub25, sub26, sub27,
                              sub28, sub29, sub30, sub31];
 
   list<SubRegIndex> ret = !if(!eq(size, 2), ret2,
                               !if(!eq(size, 3), ret3,
                                   !if(!eq(size, 4), ret4,
                                       !if(!eq(size, 5), ret5,
+                                          !if(!eq(size, 6), ret6,
                                               !if(!eq(size, 8), ret8,
-                                              !if(!eq(size, 16), ret16, ret32))))));
+                                                  !if(!eq(size, 16), ret16,
+                                                      ret32)))))));
 }
 
 // Generates list of sequential register tuple names.
 // E.g. RegSeq<3,2,2,"s">.ret -> [ "s[0:1]", "s[2:3]" ]
 class RegSeqNames<int last_reg, int stride, int size, string prefix,
                   int start = 0> {
   int next = !add(start, stride);
   int end_reg = !add(!add(start, size), -1);

 def SGPR_96Regs : SIRegisterTuples<getSubRegs<3>.ret, SGPR_32, 105, 3, 3, "s">;
 
 // SGPR 128-bit registers
 def SGPR_128Regs : SIRegisterTuples<getSubRegs<4>.ret, SGPR_32, 105, 4, 4, "s">;
 
 // SGPR 160-bit registers. No operations use these, but for symmetry with 160-bit VGPRs.
 def SGPR_160Regs : SIRegisterTuples<getSubRegs<5>.ret, SGPR_32, 105, 4, 5, "s">;
 
+// SGPR 192-bit registers
+def SGPR_192Regs : SIRegisterTuples<getSubRegs<6>.ret, SGPR_32, 105, 4, 6, "s">;
+
 // SGPR 256-bit registers
 def SGPR_256Regs : SIRegisterTuples<getSubRegs<8>.ret, SGPR_32, 105, 4, 8, "s">;
 
 // SGPR 512-bit registers
 def SGPR_512Regs : SIRegisterTuples<getSubRegs<16>.ret, SGPR_32, 105, 4, 16, "s">;
 
 // SGPR 1024-bit registers
 def SGPR_1024Regs : SIRegisterTuples<getSubRegs<32>.ret, SGPR_32, 105, 4, 32, "s">;

 def VGPR_96 : SIRegisterTuples<getSubRegs<3>.ret, VGPR_32, 255, 1, 3, "v">;
 
 // VGPR 128-bit registers
 def VGPR_128 : SIRegisterTuples<getSubRegs<4>.ret, VGPR_32, 255, 1, 4, "v">;
 
 // VGPR 160-bit registers
 def VGPR_160 : SIRegisterTuples<getSubRegs<5>.ret, VGPR_32, 255, 1, 5, "v">;
 
+// VGPR 192-bit registers
+def VGPR_192 : SIRegisterTuples<getSubRegs<6>.ret, VGPR_32, 255, 1, 6, "v">;
+
 // VGPR 256-bit registers
 def VGPR_256 : SIRegisterTuples<getSubRegs<8>.ret, VGPR_32, 255, 1, 8, "v">;
 
 // VGPR 512-bit registers
 def VGPR_512 : SIRegisterTuples<getSubRegs<16>.ret, VGPR_32, 255, 1, 16, "v">;
 
 // VGPR 1024-bit registers
 def VGPR_1024 : SIRegisterTuples<getSubRegs<32>.ret, VGPR_32, 255, 1, 32, "v">;

 def AGPR_96 : SIRegisterTuples<getSubRegs<3>.ret, AGPR_32, 255, 1, 3, "a">;
 
 // AGPR 128-bit registers
 def AGPR_128 : SIRegisterTuples<getSubRegs<4>.ret, AGPR_32, 255, 1, 4, "a">;
 
 // AGPR 160-bit registers
 def AGPR_160 : SIRegisterTuples<getSubRegs<5>.ret, AGPR_32, 255, 1, 5, "a">;
 
+// AGPR 192-bit registers
+def AGPR_192 : SIRegisterTuples<getSubRegs<6>.ret, AGPR_32, 255, 1, 6, "a">;
+
 // AGPR 256-bit registers
 def AGPR_256 : SIRegisterTuples<getSubRegs<8>.ret, AGPR_32, 255, 1, 8, "a">;
 
 // AGPR 512-bit registers
 def AGPR_512 : SIRegisterTuples<getSubRegs<16>.ret, AGPR_32, 255, 1, 16, "a">;
 
 // AGPR 1024-bit registers
 def AGPR_1024 : SIRegisterTuples<getSubRegs<32>.ret, AGPR_32, 255, 1, 32, "a">;
def SGPR_160 : RegisterClass<"AMDGPU", [v5i32, v5f32], 32,
   let AllocationPriority = 16;
 }
 
 def SReg_160 : RegisterClass<"AMDGPU", [v5i32, v5f32], 32,
                              (add SGPR_160)> {
   let AllocationPriority = 16;
 }
 
-def SGPR_256 : RegisterClass<"AMDGPU", [v8i32, v8f32, v4i64], 32, (add SGPR_256Regs)> {
+def SReg_192 : RegisterClass<"AMDGPU", [untyped], 32, (add SGPR_192Regs)> {
+  let Size = 192;
   let AllocationPriority = 17;
 }
 
+def SGPR_256 : RegisterClass<"AMDGPU", [v8i32, v8f32, v4i64], 32, (add SGPR_256Regs)> {
+  let AllocationPriority = 18;
+}
+
 def TTMP_256 : RegisterClass<"AMDGPU", [v8i32, v8f32], 32, (add TTMP_256Regs)> {
   let isAllocatable = 0;
 }
 
 def SReg_256 : RegisterClass<"AMDGPU", [v8i32, v8f32, v4i64], 32,
                              (add SGPR_256, TTMP_256)> {
   // Requires 4 s_mov_b64 to copy
   let CopyCost = 4;
-  let AllocationPriority = 17;
+  let AllocationPriority = 18;
 }
 
 def SGPR_512 : RegisterClass<"AMDGPU", [v16i32, v16f32], 32,
                              (add SGPR_512Regs)> {
-  let AllocationPriority = 18;
+  let AllocationPriority = 19;
 }
 
 def TTMP_512 : RegisterClass<"AMDGPU", [v16i32, v16f32], 32,
                              (add TTMP_512Regs)> {
   let isAllocatable = 0;
 }
 
 def SReg_512 : RegisterClass<"AMDGPU", [v16i32, v16f32], 32,
                              (add SGPR_512, TTMP_512)> {
   // Requires 8 s_mov_b64 to copy
   let CopyCost = 8;
-  let AllocationPriority = 18;
+  let AllocationPriority = 19;
 }
 
 def VRegOrLds_32 : RegisterClass<"AMDGPU", [i32, f32, i16, f16, v2i16, v2f16], 32,
                                  (add VGPR_32, LDS_DIRECT_CLASS)> {
   let isAllocatable = 0;
 }
 
 def SGPR_1024 : RegisterClass<"AMDGPU", [v32i32, v32f32], 32,
                               (add SGPR_1024Regs)> {
-  let AllocationPriority = 19;
+  let AllocationPriority = 20;
 }
 
 def SReg_1024 : RegisterClass<"AMDGPU", [v32i32, v32f32], 32,
                               (add SGPR_1024)> {
   let CopyCost = 16;
-  let AllocationPriority = 19;
+  let AllocationPriority = 20;
 }
 
 // Register class for all vector registers (VGPRs + Interploation Registers)
 class VRegClass<int numRegs, list<ValueType> regTypes, dag regList> :
     RegisterClass<"AMDGPU", regTypes, 32, regList> {
   let Size = !mul(numRegs, 32);
 
   // Requires n v_mov_b32 to copy
   let CopyCost = numRegs;
   let AllocationPriority = numRegs;
   let Weight = numRegs;
 }
 
 def VReg_64 : VRegClass<2, [i64, f64, v2i32, v2f32, v4f16, v4i16, p0, p1, p4],
                         (add VGPR_64)>;
 def VReg_96 : VRegClass<3, [v3i32, v3f32], (add VGPR_96)>;
 def VReg_128 : VRegClass<4, [v4i32, v4f32, v2i64, v2f64, i128], (add VGPR_128)>;
 def VReg_160 : VRegClass<5, [v5i32, v5f32], (add VGPR_160)>;
+def VReg_192 : VRegClass<6, [untyped], (add VGPR_192)>;
 def VReg_256 : VRegClass<8, [v8i32, v8f32], (add VGPR_256)>;
 def VReg_512 : VRegClass<16, [v16i32, v16f32], (add VGPR_512)>;
 def VReg_1024 : VRegClass<32, [v32i32, v32f32], (add VGPR_1024)>;
 
 class ARegClass<int numRegs, list<ValueType> regTypes, dag regList> :
     VRegClass<numRegs, regTypes, regList> {
   // Requires n v_accvgpr_write and n v_accvgpr_read to copy + burn 1 vgpr
   let CopyCost = !add(numRegs, numRegs, 1);
 }
 
 def AReg_64 : ARegClass<2, [i64, f64, v2i32, v2f32, v4f16, v4i16],
                         (add AGPR_64)>;
 def AReg_96 : ARegClass<3, [v3i32, v3f32], (add AGPR_96)>;
 def AReg_128 : ARegClass<4, [v4i32, v4f32, v2i64, v2f64], (add AGPR_128)>;
 def AReg_160 : ARegClass<5, [v5i32, v5f32], (add AGPR_160)>;
+def AReg_192 : ARegClass<6, [untyped], (add AGPR_192)>;
 def AReg_256 : ARegClass<8, [v8i32, v8f32], (add AGPR_256)>;
 def AReg_512 : ARegClass<16, [v16i32, v16f32], (add AGPR_512)>;
 def AReg_1024 : ARegClass<32, [v32i32, v32f32], (add AGPR_1024)>;
 
 } // End GeneratePressureSet = 0
 
 // This is not a real register. This is just to have a register to add
 // to VReg_1 that does not alias any real register that would

llvm/lib/Target/AMDGPU/Utils/AMDGPUBaseInfo.cpp

unsigned getRegBitWidth(unsigned RCID) {
   case AMDGPU::VReg_128RegClassID:
   case AMDGPU::AReg_128RegClassID:
     return 128;
   case AMDGPU::SGPR_160RegClassID:
   case AMDGPU::SReg_160RegClassID:
   case AMDGPU::VReg_160RegClassID:
   case AMDGPU::AReg_160RegClassID:
     return 160;
+  case AMDGPU::SReg_192RegClassID:
+  case AMDGPU::VReg_192RegClassID:
+  case AMDGPU::AReg_192RegClassID:
+    return 192;
   case AMDGPU::SReg_256RegClassID:
   case AMDGPU::VReg_256RegClassID:
   case AMDGPU::AReg_256RegClassID:
     return 256;
   case AMDGPU::SReg_512RegClassID:
   case AMDGPU::VReg_512RegClassID:
   case AMDGPU::AReg_512RegClassID:
     return 512;

llvm/test/CodeGen/AMDGPU/GlobalISel/cvt_f32_ubyte-xfail.ll

-; RUN: not --crash llc -global-isel -mtriple=amdgcn-- -mcpu=tahiti -verify-machineinstrs < %s
-; RUN: not --crash llc -global-isel -mtriple=amdgcn-- -mcpu=tonga -verify-machineinstrs < %s
-
-define <3 x float> @v_uitofp_v3i8_to_v3f32(i32 %arg0) nounwind {
-  %trunc = trunc i32 %arg0 to i24
-  %val = bitcast i24 %trunc to <3 x i8>
-  %cvt = uitofp <3 x i8> %val to <3 x float>
-  ret <3 x float> %cvt
-}
 

llvm/test/CodeGen/AMDGPU/GlobalISel/cvt_f32_ubyte.ll

 ; VI-NEXT:    v_cvt_f32_ubyte0_e32 v1, v0
 ; VI-NEXT:    v_mov_b32_e32 v0, v2
 ; VI-NEXT:    s_setpc_b64 s[30:31]
   %val = bitcast i16 %arg0 to <2 x i8>
   %cvt = uitofp <2 x i8> %val to <2 x float>
   ret <2 x float> %cvt
 }
 
+define <3 x float> @v_uitofp_v3i8_to_v3f32(i32 %arg0) nounwind {
+; SI-LABEL: v_uitofp_v3i8_to_v3f32:
+; SI:       ; %bb.0:
+; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; SI-NEXT:    v_and_b32_e32 v1, 0xffff, v0
+; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
+; SI-NEXT:    s_movk_i32 s4, 0xff
+; SI-NEXT:    v_lshrrev_b32_e32 v1, 8, v1
+; SI-NEXT:    v_and_b32_e32 v0, s4, v0
+; SI-NEXT:    v_and_b32_e32 v1, s4, v1
+; SI-NEXT:    v_and_b32_e32 v2, s4, v2
+; SI-NEXT:    v_cvt_f32_ubyte0_e32 v0, v0
+; SI-NEXT:    v_cvt_f32_ubyte0_e32 v1, v1
+; SI-NEXT:    v_cvt_f32_ubyte0_e32 v2, v2
+; SI-NEXT:    s_setpc_b64 s[30:31]
+;
+; VI-LABEL: v_uitofp_v3i8_to_v3f32:
+; VI:       ; %bb.0:
+; VI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; VI-NEXT:    s_movk_i32 s4, 0xff
+; VI-NEXT:    v_mov_b32_e32 v2, s4
+; VI-NEXT:    v_and_b32_sdwa v1, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:DWORD
+; VI-NEXT:    v_cvt_f32_ubyte0_sdwa v3, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0
+; VI-NEXT:    v_and_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; VI-NEXT:    v_cvt_f32_ubyte0_e32 v2, v0
+; VI-NEXT:    v_cvt_f32_ubyte0_e32 v1, v1
+; VI-NEXT:    v_mov_b32_e32 v0, v3
+; VI-NEXT:    s_setpc_b64 s[30:31]
+  %trunc = trunc i32 %arg0 to i24
+  %val = bitcast i24 %trunc to <3 x i8>
+  %cvt = uitofp <3 x i8> %val to <3 x float>
+  ret <3 x float> %cvt
+}
+
 define <4 x float> @v_uitofp_v4i8_to_v4f32(i32 %arg0) nounwind {
 ; SI-LABEL: v_uitofp_v4i8_to_v4f32:
 ; SI:       ; %bb.0:
 ; SI-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; SI-NEXT:    s_mov_b32 s4, 0xffff
 ; SI-NEXT:    v_lshrrev_b32_e32 v2, 16, v0
 ; SI-NEXT:    v_and_b32_e32 v1, s4, v0
 ; SI-NEXT:    v_and_b32_e32 v3, s4, v2

llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-concat-vectors.mir

 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
 # RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -run-pass=instruction-select -global-isel-abort=2 -pass-remarks-missed='gisel*'  -o - %s 2> %t | FileCheck -check-prefix=GCN  %s
 # RUN: FileCheck -check-prefix=ERR %s < %t
 
 # ERR-NOT: remark:
-# ERR: remark: <unknown>:0:0: cannot select: %3:sgpr(<12 x s16>) = G_CONCAT_VECTORS %0:sgpr(<4 x s16>), %1:sgpr(<4 x s16>), %2:sgpr(<4 x s16>) (in function: test_concat_vectors_s_v12s16_s_v4s16_s_v4s16_s_v4s16)
-# ERR-NEXT: remark: <unknown>:0:0: cannot select: %3:vgpr(<12 x s16>) = G_CONCAT_VECTORS %0:vgpr(<4 x s16>), %1:vgpr(<4 x s16>), %2:vgpr(<4 x s16>) (in function: test_concat_vectors_v_v12s16_v_v4s16_v_v4s16_v_v4s16)
-# ERR-NEXT: remark: <unknown>:0:0: cannot select: %2:sgpr(<6 x s64>) = G_CONCAT_VECTORS %0:sgpr(<3 x s64>), %1:sgpr(<3 x s64>) (in function: test_concat_vectors_s_v6s64_s_v3s64_s_v3s64)
+# ERR: remark: <unknown>:0:0: cannot select: %2:sgpr(<6 x s64>) = G_CONCAT_VECTORS %0:sgpr(<3 x s64>), %1:sgpr(<3 x s64>) (in function: test_concat_vectors_s_v6s64_s_v3s64_s_v3s64)
 # ERR-NOT: remark:
 
 ---
 name: test_concat_vectors_v_v4s16_v_v2s16_v_v2s16
 legalized:       true
 regBankSelected: true
 
 body: |

 legalized:       true
 regBankSelected: true
 
 body: |
   bb.0:
     liveins: $sgpr0_sgpr1, $sgpr2_sgpr3, $sgpr4_sgpr5
 
     ; GCN-LABEL: name: test_concat_vectors_s_v12s16_s_v4s16_s_v4s16_s_v4s16
-    ; GCN: [[COPY:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr0_sgpr1
-    ; GCN: [[COPY1:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr2_sgpr3
-    ; GCN: [[COPY2:%[0-9]+]]:sgpr(<4 x s16>) = COPY $sgpr4_sgpr5
-    ; GCN: [[CONCAT_VECTORS:%[0-9]+]]:sgpr(<12 x s16>) = G_CONCAT_VECTORS [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<4 x s16>)
-    ; GCN: S_ENDPGM 0, implicit [[CONCAT_VECTORS]](<12 x s16>)
+    ; GCN: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
+    ; GCN: [[COPY1:%[0-9]+]]:sreg_64 = COPY $sgpr2_sgpr3
+    ; GCN: [[COPY2:%[0-9]+]]:sreg_64 = COPY $sgpr4_sgpr5
+    ; GCN: [[REG_SEQUENCE:%[0-9]+]]:sreg_192 = REG_SEQUENCE [[COPY]], %subreg.sub0_sub1, [[COPY1]], %subreg.sub2_sub3, [[COPY2]], %subreg.sub4_sub5
+    ; GCN: S_ENDPGM 0, implicit [[REG_SEQUENCE]]
     %0:sgpr(<4 x s16>) = COPY $sgpr0_sgpr1
     %1:sgpr(<4 x s16>) = COPY $sgpr2_sgpr3
     %2:sgpr(<4 x s16>) = COPY $sgpr4_sgpr5
     %3:sgpr(<12 x s16>) = G_CONCAT_VECTORS %0, %1, %2
     S_ENDPGM 0, implicit %3
 ...
 
 ---
 name: test_concat_vectors_v_v12s16_v_v4s16_v_v4s16_v_v4s16
 legalized:       true
 regBankSelected: true
 
 body: |
   bb.0:
     liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $vgpr4_vgpr5
 
     ; GCN-LABEL: name: test_concat_vectors_v_v12s16_v_v4s16_v_v4s16_v_v4s16
-    ; GCN: [[COPY:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
-    ; GCN: [[COPY1:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
-    ; GCN: [[COPY2:%[0-9]+]]:vgpr(<4 x s16>) = COPY $vgpr4_vgpr5
-    ; GCN: [[CONCAT_VECTORS:%[0-9]+]]:vgpr(<12 x s16>) = G_CONCAT_VECTORS [[COPY]](<4 x s16>), [[COPY1]](<4 x s16>), [[COPY2]](<4 x s16>)
-    ; GCN: S_ENDPGM 0, implicit [[CONCAT_VECTORS]](<12 x s16>)
+    ; GCN: [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+    ; GCN: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr2_vgpr3
+    ; GCN: [[COPY2:%[0-9]+]]:vreg_64 = COPY $vgpr4_vgpr5
+    ; GCN: [[REG_SEQUENCE:%[0-9]+]]:vreg_192 = REG_SEQUENCE [[COPY]], %subreg.sub0_sub1, [[COPY1]], %subreg.sub2_sub3, [[COPY2]], %subreg.sub4_sub5
+    ; GCN: S_ENDPGM 0, implicit [[REG_SEQUENCE]]
     %0:vgpr(<4 x s16>) = COPY $vgpr0_vgpr1
     %1:vgpr(<4 x s16>) = COPY $vgpr2_vgpr3
     %2:vgpr(<4 x s16>) = COPY $vgpr4_vgpr5
     %3:vgpr(<12 x s16>) = G_CONCAT_VECTORS %0, %1, %2
     S_ENDPGM 0, implicit %3
 ...
 
 ---

llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-merge-values.mir

 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=2 -pass-remarks-missed='gisel*'  -o - %s 2> %t | FileCheck -check-prefix=GCN  %s
-# RUN: FileCheck -check-prefix=ERR %s < %t
-
-
-# ERR-NOT: remark:
-# ERR: remark: <unknown>:0:0: cannot select: %3:sgpr(s192) = G_MERGE_VALUES %0:sgpr(s64), %1:sgpr(s64), %2:sgpr(s64) (in function: test_merge_values_s_s192_s_s64_s_s64_s_s64)
-# ERR-NEXT: remark: <unknown>:0:0: cannot select: %3:vgpr(s192) = G_MERGE_VALUES %0:vgpr(s64), %1:vgpr(s64), %2:vgpr(s64) (in function: test_merge_values_v_s192_v_s64_v_s64_v_s64)
-# ERR-NOT: remark:
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -run-pass=instruction-select -verify-machineinstrs -global-isel-abort=2 -pass-remarks-missed='gisel*'  -o - %s | FileCheck -check-prefix=GCN  %s
 
 ---
 name: test_merge_values_v_s64_v_s32_v_s32
 legalized:       true
 regBankSelected: true
 tracksRegLiveness: true
 
 body: |

 tracksRegLiveness: true
 
 body: |
   bb.0:
     liveins: $sgpr0_sgpr1, $sgpr2_sgpr3, $sgpr4_sgpr5
 
     ; GCN-LABEL: name: test_merge_values_s_s192_s_s64_s_s64_s_s64
     ; GCN: liveins: $sgpr0_sgpr1, $sgpr2_sgpr3, $sgpr4_sgpr5
-    ; GCN: [[COPY:%[0-9]+]]:sgpr(s64) = COPY $sgpr0_sgpr1
-    ; GCN: [[COPY1:%[0-9]+]]:sgpr(s64) = COPY $sgpr2_sgpr3
-    ; GCN: [[COPY2:%[0-9]+]]:sgpr(s64) = COPY $sgpr4_sgpr5
-    ; GCN: [[MV:%[0-9]+]]:sgpr(s192) = G_MERGE_VALUES [[COPY]](s64), [[COPY1]](s64), [[COPY2]](s64)
-    ; GCN: S_ENDPGM 0, implicit [[MV]](s192)
+    ; GCN: [[COPY:%[0-9]+]]:sreg_64 = COPY $sgpr0_sgpr1
+    ; GCN: [[COPY1:%[0-9]+]]:sreg_64 = COPY $sgpr2_sgpr3
+    ; GCN: [[COPY2:%[0-9]+]]:sreg_64 = COPY $sgpr4_sgpr5
+    ; GCN: [[REG_SEQUENCE:%[0-9]+]]:sreg_192 = REG_SEQUENCE [[COPY]], %subreg.sub0_sub1, [[COPY1]], %subreg.sub2_sub3, [[COPY2]], %subreg.sub4_sub5
+    ; GCN: S_ENDPGM 0, implicit [[REG_SEQUENCE]]
     %0:sgpr(s64) = COPY $sgpr0_sgpr1
     %1:sgpr(s64) = COPY $sgpr2_sgpr3
     %2:sgpr(s64) = COPY $sgpr4_sgpr5
     %3:sgpr(s192) = G_MERGE_VALUES %0, %1, %2
     S_ENDPGM 0, implicit %3
 ...
 
 ---
 name: test_merge_values_v_s192_v_s64_v_s64_v_s64
 legalized:       true
 regBankSelected: true
 tracksRegLiveness: true
 
 body: |
   bb.0:
     liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $vgpr4_vgpr5
 
     ; GCN-LABEL: name: test_merge_values_v_s192_v_s64_v_s64_v_s64
     ; GCN: liveins: $vgpr0_vgpr1, $vgpr2_vgpr3, $vgpr4_vgpr5
-    ; GCN: [[COPY:%[0-9]+]]:vgpr(s64) = COPY $vgpr0_vgpr1
-    ; GCN: [[COPY1:%[0-9]+]]:vgpr(s64) = COPY $vgpr2_vgpr3
-    ; GCN: [[COPY2:%[0-9]+]]:vgpr(s64) = COPY $vgpr4_vgpr5
-    ; GCN: [[MV:%[0-9]+]]:vgpr(s192) = G_MERGE_VALUES [[COPY]](s64), [[COPY1]](s64), [[COPY2]](s64)
-    ; GCN: S_ENDPGM 0, implicit [[MV]](s192)
+    ; GCN: [[COPY:%[0-9]+]]:vreg_64 = COPY $vgpr0_vgpr1
+    ; GCN: [[COPY1:%[0-9]+]]:vreg_64 = COPY $vgpr2_vgpr3
+    ; GCN: [[COPY2:%[0-9]+]]:vreg_64 = COPY $vgpr4_vgpr5
+    ; GCN: [[REG_SEQUENCE:%[0-9]+]]:vreg_192 = REG_SEQUENCE [[COPY]], %subreg.sub0_sub1, [[COPY1]], %subreg.sub2_sub3, [[COPY2]], %subreg.sub4_sub5
+    ; GCN: S_ENDPGM 0, implicit [[REG_SEQUENCE]]
     %0:vgpr(s64) = COPY $vgpr0_vgpr1
     %1:vgpr(s64) = COPY $vgpr2_vgpr3
     %2:vgpr(s64) = COPY $vgpr4_vgpr5
     %3:vgpr(s192) = G_MERGE_VALUES %0, %1, %2
     S_ENDPGM 0, implicit %3
 ...
 
 ---

llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-unmerge-values.mir

 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -run-pass=instruction-select -global-isel-abort=2 -pass-remarks-missed='gisel*'  -o - %s 2> %t | FileCheck -check-prefix=GCN  %s
-# RUN: FileCheck -check-prefix=ERR %s < %t
-
-# ERR-NOT: remark:
-# ERR: remark: <unknown>:0:0: cannot select: %1:sgpr(s64), %2:sgpr(s64), %3:sgpr(s64) = G_UNMERGE_VALUES %0:sgpr(s192) (in function: test_unmerge_values_s_s64_s_s64_s64_s_s192)
-# ERR-NOT: remark:
+# RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -run-pass=instruction-select -global-isel-abort=2 -pass-remarks-missed='gisel*'  -o - %s | FileCheck -check-prefix=GCN  %s
 
 ---
 name: test_unmerge_values_v_s32_v_s32_v_s64
 legalized:       true
 regBankSelected: true
 tracksRegLiveness: true
 
 body: |

 tracksRegLiveness: true
 
 body: |
   bb.0:
     liveins: $sgpr0_sgpr1_sgpr2_sgpr3
 
     ; GCN-LABEL: name: test_unmerge_values_s_s64_s_s64_s64_s_s192
     ; GCN: liveins: $sgpr0_sgpr1_sgpr2_sgpr3
-    ; GCN: [[DEF:%[0-9]+]]:sgpr(s192) = G_IMPLICIT_DEF
-    ; GCN: [[UV:%[0-9]+]]:sgpr(s64), [[UV1:%[0-9]+]]:sgpr(s64), [[UV2:%[0-9]+]]:sgpr(s64) = G_UNMERGE_VALUES [[DEF]](s192)
-    ; GCN: S_ENDPGM 0, implicit [[UV]](s64), implicit [[UV1]](s64), implicit [[UV2]](s64)
+    ; GCN: [[DEF:%[0-9]+]]:sreg_192 = IMPLICIT_DEF
+    ; GCN: [[COPY:%[0-9]+]]:sreg_64 = COPY [[DEF]].sub0_sub1
+    ; GCN: [[COPY1:%[0-9]+]]:sreg_64 = COPY [[DEF]].sub2_sub3
+    ; GCN: [[COPY2:%[0-9]+]]:sreg_64 = COPY [[DEF]].sub4_sub5
+    ; GCN: S_ENDPGM 0, implicit [[COPY]], implicit [[COPY1]], implicit [[COPY2]]
     %0:sgpr(s192) = G_IMPLICIT_DEF
     %1:sgpr(s64), %2:sgpr(s64), %3:sgpr(s64) = G_UNMERGE_VALUES %0
     S_ENDPGM 0, implicit %1, implicit %2, implicit %3
 ...
 
 ---
 name: test_unmerge_values_rc_set_def_v_s32_v_s32_v_s64
 legalized:       true

llvm/test/CodeGen/AMDGPU/GlobalISel/zextload-xfail.ll

-; RUN: not --crash llc -global-isel -mtriple=amdgcn-amd-amdpal -mcpu=gfx900 < %s
-; RUN: not --crash llc -global-isel -mtriple=amdgcn-amd-amdpal -mcpu=fiji < %s
-; RUN: not --crash llc -global-isel -mtriple=amdgcn-amd-amdpal -mcpu=tahiti < %s
-
-define i96 @zextload_global_i32_to_i96(i32 addrspace(1)* %ptr) {
-  %load = load i32, i32 addrspace(1)* %ptr
-  %ext = zext i32 %load to i96
-  ret i96 %ext
-}

llvm/test/CodeGen/AMDGPU/GlobalISel/zextload.ll

 ; GFX6-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX6-NEXT:    s_waitcnt vmcnt(0)
 ; GFX6-NEXT:    s_setpc_b64 s[30:31]
   %load = load i32, i32 addrspace(1)* %ptr
   %ext = zext i32 %load to i64
   ret i64 %ext
 }
 
+define i96 @zextload_global_i32_to_i96(i32 addrspace(1)* %ptr) {
+; GFX9-LABEL: zextload_global_i32_to_i96:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_mov_b32_e32 v3, v1
+; GFX9-NEXT:    v_mov_b32_e32 v2, v0
+; GFX9-NEXT:    global_load_dword v0, v[2:3], off
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-NEXT:    v_mov_b32_e32 v3, 0
+; GFX9-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX8-LABEL: zextload_global_i32_to_i96:
+; GFX8:       ; %bb.0:
+; GFX8-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v3, v1
+; GFX8-NEXT:    v_mov_b32_e32 v2, v0
+; GFX8-NEXT:    flat_load_dword v0, v[2:3]
+; GFX8-NEXT:    v_mov_b32_e32 v2, 0
+; GFX8-NEXT:    v_mov_b32_e32 v1, 0
+; GFX8-NEXT:    v_mov_b32_e32 v3, 0
+; GFX8-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX8-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX6-LABEL: zextload_global_i32_to_i96:
+; GFX6:       ; %bb.0:
+; GFX6-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX6-NEXT:    v_mov_b32_e32 v3, v1
+; GFX6-NEXT:    v_mov_b32_e32 v2, v0
+; GFX6-NEXT:    s_mov_b32 s6, 0
+; GFX6-NEXT:    s_mov_b32 s7, 0xf000
+; GFX6-NEXT:    s_mov_b64 s[4:5], 0
+; GFX6-NEXT:    buffer_load_dword v0, v[2:3], s[4:7], 0 addr64
+; GFX6-NEXT:    v_mov_b32_e32 v2, 0
+; GFX6-NEXT:    v_mov_b32_e32 v1, 0
+; GFX6-NEXT:    v_mov_b32_e32 v3, 0
+; GFX6-NEXT:    s_waitcnt vmcnt(0)
+; GFX6-NEXT:    s_setpc_b64 s[30:31]
+  %load = load i32, i32 addrspace(1)* %ptr
+  %ext = zext i32 %load to i96
+  ret i96 %ext
+}
+
 define i128 @zextload_global_i32_to_i128(i32 addrspace(1)* %ptr) {
 ; GFX9-LABEL: zextload_global_i32_to_i128:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    global_load_dword v0, v[0:1], off
 ; GFX9-NEXT:    s_mov_b32 s4, 0
 ; GFX9-NEXT:    s_mov_b32 s5, 0
 ; GFX9-NEXT:    v_mov_b32_e32 v1, 0

llvm/test/CodeGen/AMDGPU/ipra-regmask.ll

 ; RUN: llc -mtriple=amdgcn-amd-amdhsa -enable-ipra -print-regusage -o /dev/null 2>&1 < %s | FileCheck %s
 ; Make sure the expected regmask is generated for sub/superregisters.
 
-; CHECK-DAG: csr Clobbered Registers: $vgpr0 $vgpr0_hi16 $vgpr0_lo16 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 $vgpr0_vgpr1_vgpr2_vgpr3 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 $vgpr0_vgpr1 $vgpr0_vgpr1_vgpr2 {{$}}
+; CHECK-DAG: csr Clobbered Registers: $vgpr0 $vgpr0_hi16 $vgpr0_lo16 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 $vgpr0_vgpr1_vgpr2_vgpr3 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 $vgpr0_vgpr1 $vgpr0_vgpr1_vgpr2 {{$}}
 define void @csr() #0 {
   call void asm sideeffect "", "~{v0},~{v36},~{v37}"() #0
   ret void
 }
 
-; CHECK-DAG: subregs_for_super Clobbered Registers: $vgpr0 $vgpr1 $vgpr0_hi16 $vgpr1_hi16 $vgpr0_lo16 $vgpr1_lo16 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32 $vgpr0_vgpr1_vgpr2_vgpr3 $vgpr1_vgpr2_vgpr3_vgpr4 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16 $vgpr0_vgpr1 $vgpr1_vgpr2 $vgpr0_vgpr1_vgpr2 $vgpr1_vgpr2_vgpr3 {{$}}
+; CHECK-DAG: subregs_for_super Clobbered Registers: $vgpr0 $vgpr1 $vgpr0_hi16 $vgpr1_hi16 $vgpr0_lo16 $vgpr1_lo16 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32 $vgpr0_vgpr1_vgpr2_vgpr3 $vgpr1_vgpr2_vgpr3_vgpr4 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16 $vgpr0_vgpr1 $vgpr1_vgpr2 $vgpr0_vgpr1_vgpr2 $vgpr1_vgpr2_vgpr3 {{$}}
 define void @subregs_for_super() #0 {
   call void asm sideeffect "", "~{v0},~{v1}"() #0
   ret void
 }
 
-; CHECK-DAG: Clobbered Registers: $vgpr0 $vgpr1 $vgpr0_hi16 $vgpr1_hi16 $vgpr0_lo16 $vgpr1_lo16 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32 $vgpr0_vgpr1_vgpr2_vgpr3 $vgpr1_vgpr2_vgpr3_vgpr4 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16 $vgpr0_vgpr1 $vgpr1_vgpr2 $vgpr0_vgpr1_vgpr2 $vgpr1_vgpr2_vgpr3 {{$}}
+; CHECK-DAG: clobbered_reg_with_sub Clobbered Registers: $vgpr0 $vgpr1 $vgpr0_hi16 $vgpr1_hi16 $vgpr0_lo16 $vgpr1_lo16 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16_vgpr17_vgpr18_vgpr19_vgpr20_vgpr21_vgpr22_vgpr23_vgpr24_vgpr25_vgpr26_vgpr27_vgpr28_vgpr29_vgpr30_vgpr31_vgpr32 $vgpr0_vgpr1_vgpr2_vgpr3 $vgpr1_vgpr2_vgpr3_vgpr4 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8 $vgpr0_vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15 $vgpr1_vgpr2_vgpr3_vgpr4_vgpr5_vgpr6_vgpr7_vgpr8_vgpr9_vgpr10_vgpr11_vgpr12_vgpr13_vgpr14_vgpr15_vgpr16 $vgpr0_vgpr1 $vgpr1_vgpr2 $vgpr0_vgpr1_vgpr2 $vgpr1_vgpr2_vgpr3 {{$}}
 define void @clobbered_reg_with_sub() #0 {
   call void asm sideeffect "", "~{v[0:1]}"() #0
   ret void
 }
 
 ; CHECK-DAG: nothing Clobbered Registers: {{$}}
 define void @nothing() #0 {
   ret void