GlobalISel: Handle lowering non-power-of-2 extloads

[foad]

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir

15738

Why does this extend to s64 for the shifting and ORing?

[arsenm]

Use correct power of 2, not just ceil. Handling the case where the widened type is the same or not is more annoying than it should be

[foad]

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir

15995–16020

Why is this so much more convoluted than the GFX9-HSA case? The loads look the same, it's just all the shifting and ORing afterwards that looks crazy here.

[arsenm]

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir

15995–16020

This is an artifact from the terrible way we currently handle unaligned accesses. We treat it as a narrowScalar action, which doesn't really make sense. I'm trying to move towards making widenScalar/narrowScalar only touch the register size, and leave the memory access alone. Unaligned access decomposition is a kind of lowering, and only tangentially related to the register types needed after legalization.

The HSA case enables unaligned access and the mesa case doesn't, so we start out by reporting we need to narrow the s32 result to s24. When that load is legalized, it ends up producing this mess. Once lowering handles alignment decomposition they should look the same

[paquette]

I think this seems reasonable.

[arsenm]

47269da5d83e079a565439cbbacd4bb119ed0aff

llvm/lib/CodeGen/GlobalISel/LegalizerHelper.cpp

if (MemSizeInBits != MemStoreSizeInBits) {
 
     if (DstTy != LoadTy)
       MIRBuilder.buildTrunc(DstReg, LoadReg);
 
     MI.eraseFromParent();
     return Legalized;
   }
 
-  if (DstTy.getSizeInBits() != MMO.getSizeInBits())
-    return UnableToLegalize;
-
-  if (MI.getOpcode() == TargetOpcode::G_LOAD) {
-    // This load needs splitting into power of 2 sized loads.
-    if (DstTy.isVector())
-      return UnableToLegalize;
-    if (isPowerOf2_32(DstTy.getSizeInBits()))
-      return UnableToLegalize; // Don't know what we're being asked to do.
+  if (isPowerOf2_32(MemSizeInBits))
 
-    // Our strategy here is to generate anyextending loads for the smaller
-    // types up to next power-2 result type, and then combine the two larger
-    // result values together, before truncating back down to the non-pow-2
-    // type.
-    // E.g. v1 = i24 load =>
-    // v2 = i32 zextload (2 byte)
-    // v3 = i32 load (1 byte)
-    // v4 = i32 shl v3, 16
-    // v5 = i32 or v4, v2
-    // v1 = i24 trunc v5
-    // By doing this we generate the correct truncate which should get
-    // combined away as an artifact with a matching extend.
-    uint64_t LargeSplitSize = PowerOf2Floor(DstTy.getSizeInBits());
-    uint64_t SmallSplitSize = DstTy.getSizeInBits() - LargeSplitSize;
+  // Big endian lowering not implemented.
+  if (MIRBuilder.getDataLayout().isBigEndian())
+    return UnableToLegalize;
+
+  uint64_t LargeSplitSize = PowerOf2Floor(MemSizeInBits);
+  uint64_t SmallSplitSize = MemSizeInBits - LargeSplitSize;
 
-    MachineFunction &MF = MIRBuilder.getMF();
-    MachineMemOperand *LargeMMO =
-      MF.getMachineMemOperand(&MMO, 0, LargeSplitSize / 8);
-    MachineMemOperand *SmallMMO = MF.getMachineMemOperand(
-      &MMO, LargeSplitSize / 8, SmallSplitSize / 8);
 
-    LLT PtrTy = MRI.getType(PtrReg);
-    unsigned AnyExtSize = NextPowerOf2(DstTy.getSizeInBits());
-    LLT AnyExtTy = LLT::scalar(AnyExtSize);
-    auto LargeLoad = MIRBuilder.buildLoadInstr(
-      TargetOpcode::G_ZEXTLOAD, AnyExtTy, PtrReg, *LargeMMO);
 
-    auto OffsetCst = MIRBuilder.buildConstant(
-      LLT::scalar(PtrTy.getSizeInBits()), LargeSplitSize / 8);
-    Register PtrAddReg = MRI.createGenericVirtualRegister(PtrTy);
-    auto SmallPtr =
-      MIRBuilder.buildPtrAdd(PtrAddReg, PtrReg, OffsetCst);
-    auto SmallLoad = MIRBuilder.buildLoad(AnyExtTy, SmallPtr,
-                                          *SmallMMO);
+  auto SmallLoad = MIRBuilder.buildLoadInstr(
+    MI.getOpcode(), AnyExtTy, SmallPtr, *SmallMMO);
 
-    auto ShiftAmt = MIRBuilder.buildConstant(AnyExtTy, LargeSplitSize);
-    auto Shift = MIRBuilder.buildShl(AnyExtTy, SmallLoad, ShiftAmt);
-    auto Or = MIRBuilder.buildOr(AnyExtTy, Shift, LargeLoad);
-    MIRBuilder.buildTrunc(DstReg, {Or});
-    MI.eraseFromParent();
-    return Legalized;
-  }
 
-  return UnableToLegalize;
-}
-
 LegalizerHelper::LegalizeResult
 LegalizerHelper::lowerStore(MachineInstr &MI) {
   // Lower a non-power of 2 store into multiple pow-2 stores.
   // E.g. split an i24 store into an i16 store + i8 store.
   // We do this by first extending the stored value to the next largest power
   // of 2 type, and then using truncating stores to store the components.
   // By doing this, likewise with G_LOAD, generate an extend that can be
   // artifact-combined away instead of leaving behind extracts.

llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp

for (unsigned Op : {G_LOAD, G_STORE}) {
     .lowerIfMemSizeNotPow2()
     .minScalar(0, S32)
     .narrowScalarIf(isWideScalarExtLoadTruncStore(0), changeTo(0, S32))
     .widenScalarToNextPow2(0)
     .moreElementsIf(vectorSmallerThan(0, 32), moreEltsToNext32Bit(0))
     .lower();
   }
 
+  // FIXME: Unaligned accesses not lowered.
   auto &ExtLoads = getActionDefinitionsBuilder({G_SEXTLOAD, G_ZEXTLOAD})
                        .legalForTypesWithMemDesc({{S32, GlobalPtr, S8, 8},
                                                   {S32, GlobalPtr, S16, 2 * 8},
                                                   {S32, LocalPtr, S8, 8},
                                                   {S32, LocalPtr, S16, 16},
                                                   {S32, PrivatePtr, S8, 8},
                                                   {S32, PrivatePtr, S16, 16},
                                                   {S32, ConstantPtr, S8, 8},
AMDGPULegalizerInfo::AMDGPULegalizerInfo(const GCNSubtarget &ST_,
   // 64-bits.
   //
   // TODO: Should generalize bitcast action into coerce, which will also cover
   // inserting addrspacecasts.
   ExtLoads.customIf(typeIs(1, Constant32Ptr));
 
   ExtLoads.clampScalar(0, S32, S32)
           .widenScalarToNextPow2(0)
-          .unsupportedIfMemSizeNotPow2()
           .lower();
 
   auto &Atomics = getActionDefinitionsBuilder(
     {G_ATOMICRMW_XCHG, G_ATOMICRMW_ADD, G_ATOMICRMW_SUB,
      G_ATOMICRMW_AND, G_ATOMICRMW_OR, G_ATOMICRMW_XOR,
      G_ATOMICRMW_MAX, G_ATOMICRMW_MIN, G_ATOMICRMW_UMAX,
      G_ATOMICRMW_UMIN})
     .legalFor({{S32, GlobalPtr}, {S32, LocalPtr},

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-load-global.mir

bb.0:
     ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s16), align 4, addrspace 1)
     ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](s32)
     %0:_(p1) = COPY $vgpr0_vgpr1
     %1:_(s32) = G_LOAD %0 :: (load (s16), align 4, addrspace 1)
     $vgpr0 = COPY %1
 ...
 
 ---
+name: test_ext_load_global_s32_from_s24_align1
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_ext_load_global_s32_from_s24_align1
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s8), addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 1, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
+    ; SI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load (s8) from unknown-address + 2, addrspace 1)
+    ; SI: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
+    ; SI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C2]]
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C3]](s32)
+    ; SI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C2]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[DEF]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C4]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
+    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
+    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[OR2]](s32)
+    ; SI: $vgpr0 = COPY [[COPY4]](s32)
+    ; CI-HSA-LABEL: name: test_ext_load_global_s32_from_s24_align1
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load (s16), align 1, addrspace 1)
+    ; CI-HSA: [[ZEXT:%[0-9]+]]:_(s64) = G_ZEXT [[ZEXTLOAD]](s32)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, addrspace 1)
+    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
+    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CI-HSA: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[ANYEXT]], [[C1]](s32)
+    ; CI-HSA: [[OR:%[0-9]+]]:_(s64) = G_OR [[SHL]], [[ZEXT]]
+    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[OR]](s64)
+    ; CI-HSA: $vgpr0 = COPY [[TRUNC]](s32)
+    ; CI-MESA-LABEL: name: test_ext_load_global_s32_from_s24_align1
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s8), addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 1, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
+    ; CI-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load (s8) from unknown-address + 2, addrspace 1)
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C2]]
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C3]](s32)
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LOAD1]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C4]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C2]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[DEF]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY3]], [[C4]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[C3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
+    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C5]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[OR2]](s32)
+    ; CI-MESA: $vgpr0 = COPY [[COPY4]](s32)
+    ; VI-LABEL: name: test_ext_load_global_s32_from_s24_align1
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s8), addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 1, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
+    ; VI: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load (s8) from unknown-address + 2, addrspace 1)
+    ; VI: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
+    ; VI: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C2]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C2]]
+    ; VI: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C3]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C2]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[DEF]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C2]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C3]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
+    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
+    ; VI: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C4]](s32)
+    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[OR2]](s32)
+    ; VI: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-HSA-LABEL: name: test_ext_load_global_s32_from_s24_align1
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load (s16), align 1, addrspace 1)
+    ; GFX9-HSA: [[ZEXT:%[0-9]+]]:_(s64) = G_ZEXT [[ZEXTLOAD]](s32)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, addrspace 1)
+    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
+    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9-HSA: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[ANYEXT]], [[C1]](s32)
+    ; GFX9-HSA: [[OR:%[0-9]+]]:_(s64) = G_OR [[SHL]], [[ZEXT]]
+    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[OR]](s64)
+    ; GFX9-HSA: $vgpr0 = COPY [[TRUNC]](s32)
+    ; GFX9-MESA-LABEL: name: test_ext_load_global_s32_from_s24_align1
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s8), addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 1
+    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 1, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[PTR_ADD1:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C1]](s64)
+    ; GFX9-MESA: [[LOAD2:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD1]](p1) :: (load (s8) from unknown-address + 2, addrspace 1)
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C2]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C2]]
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C3]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD2]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C2]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[DEF]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C2]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C3]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
+    ; GFX9-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C4]](s32)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[OR2]](s32)
+    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s32) = G_LOAD %0 :: (load (s24), align 1, addrspace 1)
+    $vgpr0 = COPY %1
+...
+---
+name: test_ext_load_global_s32_from_s24_align2
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_ext_load_global_s32_from_s24_align2
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s16), addrspace 1)
+    ; SI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; SI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; SI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, align 2, addrspace 1)
+    ; SI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; SI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
+    ; SI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; SI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
+    ; SI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; SI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
+    ; SI: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
+    ; SI: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; SI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; SI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C4]]
+    ; SI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C1]](s32)
+    ; SI: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; SI: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
+    ; SI: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; SI: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; SI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; SI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; SI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; SI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C4]]
+    ; SI: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C1]](s32)
+    ; SI: [[COPY4:%[0-9]+]]:_(s32) = COPY [[DEF]](s32)
+    ; SI: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; SI: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; SI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; SI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; SI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
+    ; SI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
+    ; SI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C2]](s32)
+    ; SI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
+    ; SI: [[COPY5:%[0-9]+]]:_(s32) = COPY [[OR2]](s32)
+    ; SI: $vgpr0 = COPY [[COPY5]](s32)
+    ; CI-HSA-LABEL: name: test_ext_load_global_s32_from_s24_align2
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load (s16), addrspace 1)
+    ; CI-HSA: [[ZEXT:%[0-9]+]]:_(s64) = G_ZEXT [[ZEXTLOAD]](s32)
+    ; CI-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, align 2, addrspace 1)
+    ; CI-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
+    ; CI-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CI-HSA: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[ANYEXT]], [[C1]](s32)
+    ; CI-HSA: [[OR:%[0-9]+]]:_(s64) = G_OR [[SHL]], [[ZEXT]]
+    ; CI-HSA: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[OR]](s64)
+    ; CI-HSA: $vgpr0 = COPY [[TRUNC]](s32)
+    ; CI-MESA-LABEL: name: test_ext_load_global_s32_from_s24_align2
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s16), addrspace 1)
+    ; CI-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; CI-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; CI-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, align 2, addrspace 1)
+    ; CI-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; CI-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
+    ; CI-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; CI-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
+    ; CI-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; CI-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
+    ; CI-MESA: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
+    ; CI-MESA: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; CI-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; CI-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C4]]
+    ; CI-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[C1]](s32)
+    ; CI-MESA: [[C5:%[0-9]+]]:_(s32) = G_CONSTANT i32 255
+    ; CI-MESA: [[COPY2:%[0-9]+]]:_(s32) = COPY [[LSHR]](s32)
+    ; CI-MESA: [[AND1:%[0-9]+]]:_(s32) = G_AND [[COPY2]], [[C5]]
+    ; CI-MESA: [[SHL:%[0-9]+]]:_(s32) = G_SHL [[AND1]], [[COPY1]](s32)
+    ; CI-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[SHL]](s32)
+    ; CI-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[TRUNC1]]
+    ; CI-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; CI-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C4]]
+    ; CI-MESA: [[COPY3:%[0-9]+]]:_(s32) = COPY [[C1]](s32)
+    ; CI-MESA: [[COPY4:%[0-9]+]]:_(s32) = COPY [[DEF]](s32)
+    ; CI-MESA: [[AND3:%[0-9]+]]:_(s32) = G_AND [[COPY4]], [[C5]]
+    ; CI-MESA: [[SHL1:%[0-9]+]]:_(s32) = G_SHL [[AND3]], [[COPY3]](s32)
+    ; CI-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[SHL1]](s32)
+    ; CI-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[TRUNC3]]
+    ; CI-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
+    ; CI-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
+    ; CI-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C2]](s32)
+    ; CI-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
+    ; CI-MESA: [[COPY5:%[0-9]+]]:_(s32) = COPY [[OR2]](s32)
+    ; CI-MESA: $vgpr0 = COPY [[COPY5]](s32)
+    ; VI-LABEL: name: test_ext_load_global_s32_from_s24_align2
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s16), addrspace 1)
+    ; VI: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; VI: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; VI: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, align 2, addrspace 1)
+    ; VI: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; VI: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
+    ; VI: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; VI: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
+    ; VI: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; VI: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
+    ; VI: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
+    ; VI: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; VI: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; VI: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C4]]
+    ; VI: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; VI: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C4]]
+    ; VI: [[C5:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; VI: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C5]](s16)
+    ; VI: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; VI: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; VI: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C4]]
+    ; VI: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[DEF]](s32)
+    ; VI: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C4]]
+    ; VI: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C5]](s16)
+    ; VI: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; VI: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
+    ; VI: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
+    ; VI: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C2]](s32)
+    ; VI: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
+    ; VI: [[COPY1:%[0-9]+]]:_(s32) = COPY [[OR2]](s32)
+    ; VI: $vgpr0 = COPY [[COPY1]](s32)
+    ; GFX9-HSA-LABEL: name: test_ext_load_global_s32_from_s24_align2
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load (s16), addrspace 1)
+    ; GFX9-HSA: [[ZEXT:%[0-9]+]]:_(s64) = G_ZEXT [[ZEXTLOAD]](s32)
+    ; GFX9-HSA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-HSA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, align 2, addrspace 1)
+    ; GFX9-HSA: [[ANYEXT:%[0-9]+]]:_(s64) = G_ANYEXT [[LOAD]](s32)
+    ; GFX9-HSA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9-HSA: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[ANYEXT]], [[C1]](s32)
+    ; GFX9-HSA: [[OR:%[0-9]+]]:_(s64) = G_OR [[SHL]], [[ZEXT]]
+    ; GFX9-HSA: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[OR]](s64)
+    ; GFX9-HSA: $vgpr0 = COPY [[TRUNC]](s32)
+    ; GFX9-MESA-LABEL: name: test_ext_load_global_s32_from_s24_align2
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s16), addrspace 1)
+    ; GFX9-MESA: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX9-MESA: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; GFX9-MESA: [[LOAD1:%[0-9]+]]:_(s32) = G_LOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, align 2, addrspace 1)
+    ; GFX9-MESA: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 8
+    ; GFX9-MESA: [[LSHR:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C1]](s32)
+    ; GFX9-MESA: [[C2:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX9-MESA: [[LSHR1:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C2]](s32)
+    ; GFX9-MESA: [[C3:%[0-9]+]]:_(s32) = G_CONSTANT i32 24
+    ; GFX9-MESA: [[LSHR2:%[0-9]+]]:_(s32) = G_LSHR [[LOAD]], [[C3]](s32)
+    ; GFX9-MESA: [[DEF:%[0-9]+]]:_(s32) = G_IMPLICIT_DEF
+    ; GFX9-MESA: [[C4:%[0-9]+]]:_(s16) = G_CONSTANT i16 255
+    ; GFX9-MESA: [[TRUNC:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD]](s32)
+    ; GFX9-MESA: [[AND:%[0-9]+]]:_(s16) = G_AND [[TRUNC]], [[C4]]
+    ; GFX9-MESA: [[TRUNC1:%[0-9]+]]:_(s16) = G_TRUNC [[LSHR]](s32)
+    ; GFX9-MESA: [[AND1:%[0-9]+]]:_(s16) = G_AND [[TRUNC1]], [[C4]]
+    ; GFX9-MESA: [[C5:%[0-9]+]]:_(s16) = G_CONSTANT i16 8
+    ; GFX9-MESA: [[SHL:%[0-9]+]]:_(s16) = G_SHL [[AND1]], [[C5]](s16)
+    ; GFX9-MESA: [[OR:%[0-9]+]]:_(s16) = G_OR [[AND]], [[SHL]]
+    ; GFX9-MESA: [[TRUNC2:%[0-9]+]]:_(s16) = G_TRUNC [[LOAD1]](s32)
+    ; GFX9-MESA: [[AND2:%[0-9]+]]:_(s16) = G_AND [[TRUNC2]], [[C4]]
+    ; GFX9-MESA: [[TRUNC3:%[0-9]+]]:_(s16) = G_TRUNC [[DEF]](s32)
+    ; GFX9-MESA: [[AND3:%[0-9]+]]:_(s16) = G_AND [[TRUNC3]], [[C4]]
+    ; GFX9-MESA: [[SHL1:%[0-9]+]]:_(s16) = G_SHL [[AND3]], [[C5]](s16)
+    ; GFX9-MESA: [[OR1:%[0-9]+]]:_(s16) = G_OR [[AND2]], [[SHL1]]
+    ; GFX9-MESA: [[ZEXT:%[0-9]+]]:_(s32) = G_ZEXT [[OR]](s16)
+    ; GFX9-MESA: [[ZEXT1:%[0-9]+]]:_(s32) = G_ZEXT [[OR1]](s16)
+    ; GFX9-MESA: [[SHL2:%[0-9]+]]:_(s32) = G_SHL [[ZEXT1]], [[C2]](s32)
+    ; GFX9-MESA: [[OR2:%[0-9]+]]:_(s32) = G_OR [[ZEXT]], [[SHL2]]
+    ; GFX9-MESA: [[COPY1:%[0-9]+]]:_(s32) = COPY [[OR2]](s32)
+    ; GFX9-MESA: $vgpr0 = COPY [[COPY1]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s32) = G_LOAD %0 :: (load (s24), align 2, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
+name: test_ext_load_global_s32_from_s24_align4
+body: |
+  bb.0:
+    liveins: $vgpr0_vgpr1
+
+    ; SI-LABEL: name: test_ext_load_global_s32_from_s24_align4
+    ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; SI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s32), addrspace 1)
+    ; SI: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-HSA-LABEL: name: test_ext_load_global_s32_from_s24_align4
+    ; CI-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s32), addrspace 1)
+    ; CI-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; CI-MESA-LABEL: name: test_ext_load_global_s32_from_s24_align4
+    ; CI-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; CI-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s32), addrspace 1)
+    ; CI-MESA: $vgpr0 = COPY [[LOAD]](s32)
+    ; VI-LABEL: name: test_ext_load_global_s32_from_s24_align4
+    ; VI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; VI: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s32), addrspace 1)
+    ; VI: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-HSA-LABEL: name: test_ext_load_global_s32_from_s24_align4
+    ; GFX9-HSA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-HSA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s32), addrspace 1)
+    ; GFX9-HSA: $vgpr0 = COPY [[LOAD]](s32)
+    ; GFX9-MESA-LABEL: name: test_ext_load_global_s32_from_s24_align4
+    ; GFX9-MESA: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
+    ; GFX9-MESA: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p1) :: (load (s32), addrspace 1)
+    ; GFX9-MESA: $vgpr0 = COPY [[LOAD]](s32)
+    %0:_(p1) = COPY $vgpr0_vgpr1
+    %1:_(s32) = G_LOAD %0 :: (load (s24), align 4, addrspace 1)
+    $vgpr0 = COPY %1
+...
+
+---
 name: test_ext_load_global_s64_from_1_align4
 body: |
   bb.0:
     liveins: $vgpr0_vgpr1
 
 
     ; SI-LABEL: name: test_ext_load_global_s64_from_1_align4
     ; SI: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-sextload-global.mir

 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
 # RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -run-pass=legalizer -global-isel-abort=2 -pass-remarks-missed='gisel*' -o - 2> %t %s | FileCheck -check-prefix=GFX8 %s
 # RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -run-pass=legalizer -global-isel-abort=2 -o - %s | FileCheck -check-prefix=GFX6  %s
 # RUN: FileCheck -check-prefixes=ERR %s < %t
 
 # FIXME: Run with and without unaligned access turned on
 
 # ERR-NOT: remark
-# ERR: remark: <unknown>:0:0: unable to legalize instruction: %1:_(s32) = G_SEXTLOAD %0:_(p1) :: (load (s24), align 4, addrspace 1) (in function: test_sextload_global_i32_i24)
 # ERR: remark: <unknown>:0:0: unable to legalize instruction: %1:_(<2 x s16>) = G_SEXTLOAD %0:_(p1) :: (load (<2 x s8>), addrspace 1) (in function: test_sextload_global_v2i16_from_v2s8)
 # ERR-NEXT: remark: <unknown>:0:0: unable to legalize instruction: %1:_(<2 x s32>) = G_SEXTLOAD %0:_(p1) :: (load (<2 x s8>), addrspace 1) (in function: test_sextload_global_v2i32_from_v2s8)
 # ERR-NEXT: remark: <unknown>:0:0: unable to legalize instruction: %1:_(<2 x s32>) = G_SEXTLOAD %0:_(p1) :: (load (<2 x s16>), addrspace 1) (in function: test_sextload_global_v2i32_from_v2s16)
 # ERR-NEXT: remark: <unknown>:0:0: unable to legalize instruction: %1:_(<2 x s64>) = G_SEXTLOAD %0:_(p1) :: (load (<2 x s16>), addrspace 1) (in function: test_sextload_global_v2i64_from_v2s16)
 # ERR-NEXT: remark: <unknown>:0:0: unable to legalize instruction: %1:_(<2 x s64>) = G_SEXTLOAD %0:_(p1) :: (load (<2 x s32>), addrspace 1) (in function: test_sextload_global_v2i64_from_v2s32)
 # ERR-NEXT: remark: <unknown>:0:0: unable to legalize instruction: %1:_(s128) = G_SEXTLOAD %0:_(p1) :: (load (s64), addrspace 1) (in function: test_sextload_global_s128_8)
 # ERR-NOT: remark
 

 ---
 name: test_sextload_global_i32_i24
 body: |
   bb.0:
     liveins: $vgpr0_vgpr1
 
     ; GFX8-LABEL: name: test_sextload_global_i32_i24
     ; GFX8: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
-    ; GFX8: [[SEXTLOAD:%[0-9]+]]:_(s32) = G_SEXTLOAD [[COPY]](p1) :: (load (s24), align 4, addrspace 1)
-    ; GFX8: $vgpr0 = COPY [[SEXTLOAD]](s32)
+    ; GFX8: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load (s16), align 4, addrspace 1)
+    ; GFX8: [[ZEXT:%[0-9]+]]:_(s64) = G_ZEXT [[ZEXTLOAD]](s32)
+    ; GFX8: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX8: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; GFX8: [[SEXTLOAD:%[0-9]+]]:_(s32) = G_SEXTLOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, align 2, addrspace 1)
+    ; GFX8: [[SEXT:%[0-9]+]]:_(s64) = G_SEXT [[SEXTLOAD]](s32)
+    ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX8: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[SEXT]], [[C1]](s32)
+    ; GFX8: [[OR:%[0-9]+]]:_(s64) = G_OR [[SHL]], [[ZEXT]]
+    ; GFX8: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[OR]](s64)
+    ; GFX8: $vgpr0 = COPY [[TRUNC]](s32)
     ; GFX6-LABEL: name: test_sextload_global_i32_i24
     ; GFX6: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
-    ; GFX6: [[SEXTLOAD:%[0-9]+]]:_(s32) = G_SEXTLOAD [[COPY]](p1) :: (load (s24), align 4, addrspace 1)
-    ; GFX6: $vgpr0 = COPY [[SEXTLOAD]](s32)
+    ; GFX6: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load (s16), align 4, addrspace 1)
+    ; GFX6: [[ZEXT:%[0-9]+]]:_(s64) = G_ZEXT [[ZEXTLOAD]](s32)
+    ; GFX6: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX6: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; GFX6: [[SEXTLOAD:%[0-9]+]]:_(s32) = G_SEXTLOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, align 2, addrspace 1)
+    ; GFX6: [[SEXT:%[0-9]+]]:_(s64) = G_SEXT [[SEXTLOAD]](s32)
+    ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX6: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[SEXT]], [[C1]](s32)
+    ; GFX6: [[OR:%[0-9]+]]:_(s64) = G_OR [[SHL]], [[ZEXT]]
+    ; GFX6: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[OR]](s64)
+    ; GFX6: $vgpr0 = COPY [[TRUNC]](s32)
     %0:_(p1) = COPY $vgpr0_vgpr1
     %1:_(s32) = G_SEXTLOAD %0 :: (load (s24), addrspace 1)
     $vgpr0 = COPY %1
 ...
 
 ---
 name: test_sextload_global_i32_i30
 body: |

llvm/test/CodeGen/AMDGPU/GlobalISel/legalize-zextload-global.mir

 # NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
 # RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -run-pass=legalizer -global-isel-abort=2 -pass-remarks-missed='gisel*' -o - 2> %t %s | FileCheck -check-prefix=GFX8 %s
 # RUN: llc -mtriple=amdgcn-mesa-mesa3d -mcpu=tahiti -run-pass=legalizer -global-isel-abort=2 -o - %s | FileCheck -check-prefix=GFX6 %s
 # RUN: FileCheck -check-prefixes=ERR %s < %t
 
 # FIXME: Run with and without unaligned access turned on
 
 # ERR-NOT: remark
-# ERR: remark: <unknown>:0:0: unable to legalize instruction: %1:_(s32) = G_ZEXTLOAD %0:_(p1) :: (load (s24), align 4, addrspace 1) (in function: test_zextload_global_i32_i24)
 # ERR: remark: <unknown>:0:0: unable to legalize instruction: %1:_(<2 x s16>) = G_ZEXTLOAD %0:_(p1) :: (load (s16), addrspace 1) (in function: test_zextload_global_v2i16_from_2)
 # ERR-NEXT: remark: <unknown>:0:0: unable to legalize instruction: %1:_(<2 x s32>) = G_ZEXTLOAD %0:_(p1) :: (load (s16), addrspace 1) (in function: test_zextload_global_v2i32_from_2)
 # ERR-NEXT: remark: <unknown>:0:0: unable to legalize instruction: %1:_(<2 x s32>) = G_ZEXTLOAD %0:_(p1) :: (load (s32), addrspace 1) (in function: test_zextload_global_v2i32_from_4)
 # ERR-NEXT: remark: <unknown>:0:0: unable to legalize instruction: %1:_(<2 x s64>) = G_ZEXTLOAD %0:_(p1) :: (load (s32), addrspace 1) (in function: test_zextload_global_v2i64_from_4)
 # ERR-NEXT: remark: <unknown>:0:0: unable to legalize instruction: %1:_(<2 x s64>) = G_ZEXTLOAD %0:_(p1) :: (load (s64), addrspace 1) (in function: test_zextload_global_v2i64_from_8)
 # ERR-NEXT: remark: <unknown>:0:0: unable to legalize instruction: %1:_(s128) = G_ZEXTLOAD %0:_(p1) :: (load (s64), addrspace 1) (in function: test_zextload_global_s128_8)
 # ERR-NOT: remark
 

 ---
 name: test_zextload_global_i32_i24
 body: |
   bb.0:
     liveins: $vgpr0_vgpr1
 
     ; GFX8-LABEL: name: test_zextload_global_i32_i24
     ; GFX8: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
-    ; GFX8: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load (s24), align 4, addrspace 1)
-    ; GFX8: $vgpr0 = COPY [[ZEXTLOAD]](s32)
+    ; GFX8: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load (s16), align 4, addrspace 1)
+    ; GFX8: [[ZEXT:%[0-9]+]]:_(s64) = G_ZEXT [[ZEXTLOAD]](s32)
+    ; GFX8: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX8: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; GFX8: [[ZEXTLOAD1:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, align 2, addrspace 1)
+    ; GFX8: [[ZEXT1:%[0-9]+]]:_(s64) = G_ZEXT [[ZEXTLOAD1]](s32)
+    ; GFX8: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX8: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[ZEXT1]], [[C1]](s32)
+    ; GFX8: [[OR:%[0-9]+]]:_(s64) = G_OR [[SHL]], [[ZEXT]]
+    ; GFX8: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[OR]](s64)
+    ; GFX8: $vgpr0 = COPY [[TRUNC]](s32)
     ; GFX6-LABEL: name: test_zextload_global_i32_i24
     ; GFX6: [[COPY:%[0-9]+]]:_(p1) = COPY $vgpr0_vgpr1
-    ; GFX6: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load (s24), align 4, addrspace 1)
-    ; GFX6: $vgpr0 = COPY [[ZEXTLOAD]](s32)
+    ; GFX6: [[ZEXTLOAD:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[COPY]](p1) :: (load (s16), align 4, addrspace 1)
+    ; GFX6: [[ZEXT:%[0-9]+]]:_(s64) = G_ZEXT [[ZEXTLOAD]](s32)
+    ; GFX6: [[C:%[0-9]+]]:_(s64) = G_CONSTANT i64 2
+    ; GFX6: [[PTR_ADD:%[0-9]+]]:_(p1) = G_PTR_ADD [[COPY]], [[C]](s64)
+    ; GFX6: [[ZEXTLOAD1:%[0-9]+]]:_(s32) = G_ZEXTLOAD [[PTR_ADD]](p1) :: (load (s8) from unknown-address + 2, align 2, addrspace 1)
+    ; GFX6: [[ZEXT1:%[0-9]+]]:_(s64) = G_ZEXT [[ZEXTLOAD1]](s32)
+    ; GFX6: [[C1:%[0-9]+]]:_(s32) = G_CONSTANT i32 16
+    ; GFX6: [[SHL:%[0-9]+]]:_(s64) = G_SHL [[ZEXT1]], [[C1]](s32)
+    ; GFX6: [[OR:%[0-9]+]]:_(s64) = G_OR [[SHL]], [[ZEXT]]
+    ; GFX6: [[TRUNC:%[0-9]+]]:_(s32) = G_TRUNC [[OR]](s64)
+    ; GFX6: $vgpr0 = COPY [[TRUNC]](s32)
     %0:_(p1) = COPY $vgpr0_vgpr1
     %1:_(s32) = G_ZEXTLOAD %0 :: (load (s24), addrspace 1)
     $vgpr0 = COPY %1
 ...
 
 ---
 name: test_zextload_global_i32_i30
 body: |