[AMDGPU] Add SIMemoryLegalizer comments to clarify bit usage

[critson]

Summary

Attempt to further document the intended cache policies requested
by different combinations of GLC, SLC and DLC bits.

[foad]

LGTM.

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

881

Not really related to your patch, but why do we return here? Doesn't that mean that IsNonTemporal is effectively ignored if IsVolatile is true? Wouldn't it be both better and simpler to fall through to the IsNonTemporal handling here?

[critson]

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

881

This is a good point. From a bit setting perspective it would be fine.
Of course there is the question of what it semantically means to have a volatile nontemporal access when we seem to define volatile as "bypasses all caches".

[t-tye]

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

881

I suspect this is because at one time relaxed atomics were marked as volatile. This may have been because the C/C++/OpenCL standards defined them that way, or because LLVM back then did not fully support atomics so marking them all as volatile made the existing passes "do the right thing". So this code may have been an attempt not to pessimize normal relaxed atomics. LLVM does not support non-temporal atomics currently.

I am not sure if these reasons are still the case so would be good to investigate and potentially fix this code (or at least document why it is the way it is with a FIXME). That can be a separate review I think.

[t-tye]

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

844

Please add back:

/// There is no bypass control for the L2 cache at the isa level.

The modified comment is only explaining the L1 cache and both caches are involved for system scope.

867–868

How about:

// Request L1 cache policy to be MISS_EVICT for load instructions and MISS_LRU for store instructions. Note that there is no L2 cache bypass policy at the isa level.

885–887

I had not remembered that the GLC and SLC bits are used together to set the L1 cache policy. So how about:

// Setting GLC and SLC both to 1 sets the L1 cache policy to MISS_EVICT for both loads and stores, and the L2 cache policy to STREAM.

and delete the comment below.

1108

MISS_EVICT is a policy that only applies when both SLC and GLC are set. Here only GLC is being set which specifies MISS_LRU and L2 LRU. How about:

// Set the L1 cache policy to MISS_LRU. Note that there is no L2 cache bypass policy at the isa level.

1219–1220

How about:

// Request L1 cache policy to be MISS_EVICT for load instructions and MISS_LRU for store instructions. Note that there is no L2 cache bypass policy at the isa level.

1237–1240

How about:

// Setting GLC and SLC both to 1 sets the L1 cache policy to MISS_EVICT for both loads and stores, and the L2 cache policy to STREAM.

1400

How about:

// Set the L0 and L1 cache policies to MISS_EVICT. Note that there is no L2 cache bypass policy at the isa level.

1450–1452

How about:

// Request L0 and L1 cache policy to be MISS_EVICT for load instructions and MISS_LRU for store instructions. Note that there is no L2 cache coherent bypass policy at the isa level.

1469–1476

It appears that this should be setting GLC=1 for stores so that L0 will be HIT_EVICT instead of MISS_EVICT. This must not be done for loads as that would make Lo MISS_EVICT. How about:

// For loads setting GLC to 1 sets the L0 and L1 cache policy to HIT_EVICT and the L2 cache policy to STREAM. For stores setting GLC and SLC both to 1 sets the L0 and L1 cache policy to MISS_EVICT and the L2 cache policy to STREAM.

[critson]

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

844

I deleted that text because there is a bypass for L2 stores and atomics on GFX10: SLC=0 DLC=1.
I can put it back but only contextualised for targets before GFX10?
(And the same for all the similar references in comments below.)

881

Yes, a separate investigation and review.

1469–1476

Do you have MISS_EVICT and HIT_EVICT flipped in your description?

Do you mean:
// For loads setting SLC to 1 sets the L0 and L1 cache policy to HIT_EVICT and the L2 cache policy to STREAM. For stores setting GLC and SLC both to 1 sets the L0 cache policy to MISS_EVICT and the L2 cache policy to STREAM. L1 is always bypassed for stores.

I can add the GLC bit for stores and this ceases to be NFC.

[t-tye]

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

844

In GFX10 the bypass is only available for stores and not loads, and is not coherent so cannot be used anyway. That is why I added the word "coherent" in one of the comments below. So probably should do that here too.

1469–1476

I believe I have it right according to the hardware GFX10 memory model spec.

// For loads setting SLC to 1 sets the L0 and L1 cache policy to HIT_EVICT and the L2 cache policy to STREAM. For stores setting GLC and SLC both to 1 sets the L0 and L1 cache policy to MISS_EVICT and the L2 cache policy to STREAM.

We have to state the policy for L1 too even though the hardware documentation does not state it. The L1 MUST be evict or a subsequent load could see stale data.

Yes this changes ceases to be NFC and will need thorough testing.

[critson]

• Address reviewer comments.
• Normalise language to use "set" instead of "request".
• Add GLC to non-temporal stores on GFX10 meaning this is no longer NFC.

[critson]

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

1108

"The load intentionally misses the GPU L1 and reads from L2. If there was a line in the GPU L1 that matched, it is invalidated; L2 is reread." -- (CDNA1 Shader ISA, p67).

This sounds like MISS_EVICT to me.
As far as I know MISS_LRU only exists for stores and means write-combine, whereas MISS_EVICT is write-through.

1469–1476

Technically the L1 only has one policy described as "bypassed (but is coherent)" (RDNA1 Shader ISA p69), but on paper the behaviour of this looks the same as MISS_EVICT. So I guess I can accept just calling it that.

Do you have any test cases which use non-temporal?

[t-tye]

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

1108

GFX90A cache policies are different to other GFX9 and GFX10. If GLC=1 then the L1 policy is MISS_LRU for loads: any existing line is invalidated, then the cache line is loaded and remains in the cache with LRU policy.

1469–1476

In the GFX10 memory model spec I have, it does not state the L1 policy for stores, but my understanding is that it behaves as MISS_EVICT. To me "bypass but coherent" is semantically the same thing as MISS_EVICT and it seems odd not to use the terminology that is already well defined.

[critson]

• Update GFX90A comment inline with Tony's comments

[t-tye]

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

868

I would not add "{write-combine)" as it really is not an exact match for that term.

1220

Would eliminate "(write-combine)" as mentioned above.

1451

Eliminate "(write-cobine)". Add:

// Note that there is no L2 cache coherent bypass policy at the isa level.

[critson]

• Address reviewer comments

[t-tye]

LGTM

[critson]

• Update tests for non-temporal GLC bit addition in GFX10

[foad]

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

1474

Does this change the documented code sequences: https://llvm.org/docs/AMDGPUUsage.html#amdgpu-amdhsa-memory-model-code-sequences-gfx10-table ?

[foad]

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

1474

D114707

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

bool SIGfx6CacheControl::enableLoadCacheBypass(
     SIAtomicAddrSpace AddrSpace) const {
   assert(MI->mayLoad() && !MI->mayStore());
   bool Changed = false;
 
   if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) {
     switch (Scope) {
     case SIAtomicScope::SYSTEM:
     case SIAtomicScope::AGENT:
+      // Set L1 cache policy to MISS_EVICT.
+      // Note: there is no L2 cache bypass policy at the ISA level.
       Changed |= enableGLCBit(MI);
       break;
     case SIAtomicScope::WORKGROUP:
     case SIAtomicScope::WAVEFRONT:
     case SIAtomicScope::SINGLETHREAD:
       // No cache to bypass.
       break;
     default:

 
 bool SIGfx6CacheControl::enableRMWCacheBypass(
     const MachineBasicBlock::iterator &MI,
     SIAtomicScope Scope,
     SIAtomicAddrSpace AddrSpace) const {
   assert(MI->mayLoad() && MI->mayStore());
   bool Changed = false;
 
-  /// The L1 cache is write through so does not need to be bypassed. There is no
-  /// bypass control for the L2 cache at the isa level.
+  /// Do not set GLC for RMW atomic operations as L0/L1 cache is automatically
+  /// bypassed, and the GLC bit is instead used to indicate if they are
+  /// return or no-return.
+  /// Note: there is no L2 cache coherent bypass control at the ISA level.
 
   return Changed;
 }
 
 bool SIGfx6CacheControl::enableVolatileAndOrNonTemporal(
     MachineBasicBlock::iterator &MI, SIAtomicAddrSpace AddrSpace, SIMemOp Op,
     bool IsVolatile, bool IsNonTemporal) const {
   // Only handle load and store, not atomic read-modify-write insructions. The
   // latter use glc to indicate if the atomic returns a result and so must not
   // be used for cache control.
   assert(MI->mayLoad() ^ MI->mayStore());
 
   // Only update load and store, not LLVM IR atomic read-modify-write
   // instructions. The latter are always marked as volatile so cannot sensibly
   // handle it as do not want to pessimize all atomics. Also they do not support
   // the nontemporal attribute.
   assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE);
 
   bool Changed = false;
 
   if (IsVolatile) {
+    // Set L1 cache policy to be MISS_EVICT for load instructions
+    // and MISS_LRU for store instructions.
+    // Note: there is no L2 cache bypass policy at the ISA level.
     if (Op == SIMemOp::LOAD)
       Changed |= enableGLCBit(MI);
 
     // Ensure operation has completed at system scope to cause all volatile
     // operations to be visible outside the program in a global order. Do not
     // request cross address space as only the global address space can be
     // observable outside the program, so no need to cause a waitcnt for LDS
     // address space operations.
     Changed |= insertWait(MI, SIAtomicScope::SYSTEM, AddrSpace, Op, false,
                           Position::AFTER);
 
     return Changed;
   }
 
   if (IsNonTemporal) {
-    // Request L1 MISS_EVICT and L2 STREAM for load and store instructions.
+    // Setting both GLC and SLC configures L1 cache policy to MISS_EVICT
+    // for both loads and stores, and the L2 cache policy to STREAM.
     Changed |= enableGLCBit(MI);
     Changed |= enableSLCBit(MI);
     return Changed;
   }
 
   return Changed;
 }
 
 bool SIGfx6CacheControl::insertWait(MachineBasicBlock::iterator &MI,
bool SIGfx90ACacheControl::enableLoadCacheBypass(
     SIAtomicAddrSpace AddrSpace) const {
   assert(MI->mayLoad() && !MI->mayStore());
   bool Changed = false;
 
   if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) {
     switch (Scope) {
     case SIAtomicScope::SYSTEM:
     case SIAtomicScope::AGENT:
+      // Set the L1 cache policy to MISS_LRU.
+      // Note: there is no L2 cache bypass policy at the ISA level.
       Changed |= enableGLCBit(MI);
       break;
     case SIAtomicScope::WORKGROUP:
       // In threadgroup split mode the waves of a work-group can be executing on
       // different CUs. Therefore need to bypass the L1 which is per CU.
       // Otherwise in non-threadgroup split mode all waves of a work-group are
       // on the same CU, and so the L1 does not need to be bypassed.
       if (ST.isTgSplitEnabled())
bool SIGfx90ACacheControl::enableVolatileAndOrNonTemporal(
   // instructions. The latter are always marked as volatile so cannot sensibly
   // handle it as do not want to pessimize all atomics. Also they do not support
   // the nontemporal attribute.
   assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE);
 
   bool Changed = false;
 
   if (IsVolatile) {
+    // Set L1 cache policy to be MISS_EVICT for load instructions
+    // and MISS_LRU for store instructions.
+    // Note: there is no L2 cache bypass policy at the ISA level.
     if (Op == SIMemOp::LOAD)
       Changed |= enableGLCBit(MI);
 
     // Ensure operation has completed at system scope to cause all volatile
     // operations to be visible outside the program in a global order. Do not
     // request cross address space as only the global address space can be
     // observable outside the program, so no need to cause a waitcnt for LDS
     // address space operations.
     Changed |= insertWait(MI, SIAtomicScope::SYSTEM, AddrSpace, Op, false,
                           Position::AFTER);
 
     return Changed;
   }
 
   if (IsNonTemporal) {
-    // Request L1 MISS_EVICT and L2 STREAM for load and store instructions.
+    // Setting both GLC and SLC configures L1 cache policy to MISS_EVICT
+    // for both loads and stores, and the L2 cache policy to STREAM.
     Changed |= enableGLCBit(MI);
     Changed |= enableSLCBit(MI);
     return Changed;
   }
 
   return Changed;
 }
 
 bool SIGfx90ACacheControl::insertWait(MachineBasicBlock::iterator &MI,
                                       SIAtomicScope Scope,

 bool SIGfx10CacheControl::enableLoadCacheBypass(
     const MachineBasicBlock::iterator &MI,
     SIAtomicScope Scope,
     SIAtomicAddrSpace AddrSpace) const {
   assert(MI->mayLoad() && !MI->mayStore());
   bool Changed = false;
 
   if ((AddrSpace & SIAtomicAddrSpace::GLOBAL) != SIAtomicAddrSpace::NONE) {
-    /// TODO Do not set glc for rmw atomic operations as they
-    /// implicitly bypass the L0/L1 caches.
-
     switch (Scope) {
     case SIAtomicScope::SYSTEM:
     case SIAtomicScope::AGENT:
+      // Set the L0 and L1 cache policies to MISS_EVICT.
+      // Note: there is no L2 cache coherent bypass control at the ISA level.
       Changed |= enableGLCBit(MI);
       Changed |= enableDLCBit(MI);
       break;
     case SIAtomicScope::WORKGROUP:
       // In WGP mode the waves of a work-group can be executing on either CU of
       // the WGP. Therefore need to bypass the L0 which is per CU. Otherwise in
       // CU mode all waves of a work-group are on the same CU, and so the L0
       // does not need to be bypassed.
bool SIGfx10CacheControl::enableVolatileAndOrNonTemporal(
   // instructions. The latter are always marked as volatile so cannot sensibly
   // handle it as do not want to pessimize all atomics. Also they do not support
   // the nontemporal attribute.
   assert(Op == SIMemOp::LOAD || Op == SIMemOp::STORE);
 
   bool Changed = false;
 
   if (IsVolatile) {
+    // Set L0 and L1 cache policy to be MISS_EVICT for load instructions
+    // and MISS_LRU for store instructions.
+    // Note: there is no L2 cache coherent bypass control at the ISA level.
     if (Op == SIMemOp::LOAD) {
       Changed |= enableGLCBit(MI);
       Changed |= enableDLCBit(MI);
     }
 
     // Ensure operation has completed at system scope to cause all volatile
     // operations to be visible outside the program in a global order. Do not
     // request cross address space as only the global address space can be
     // observable outside the program, so no need to cause a waitcnt for LDS
     // address space operations.
     Changed |= insertWait(MI, SIAtomicScope::SYSTEM, AddrSpace, Op, false,
                           Position::AFTER);
     return Changed;
   }
 
   if (IsNonTemporal) {
-    // Request L0/L1 HIT_EVICT and L2 STREAM for load and store instructions.
+    // For loads setting SLC configures L0 and L1 cache policy to HIT_EVICT
+    // and L2 cache policy to STREAM.
+    // For stores setting both GLC and SLC configures L0 and L1 cache policy
+    // to MISS_EVICT and the L2 cache policy to STREAM.
+    if (Op == SIMemOp::STORE)
+      Changed |= enableGLCBit(MI);
     Changed |= enableSLCBit(MI);
+
     return Changed;
   }
 
   return Changed;
 }
 
 bool SIGfx10CacheControl::insertWait(MachineBasicBlock::iterator &MI,
                                      SIAtomicScope Scope,

llvm/test/CodeGen/AMDGPU/memory-legalizer-flat-nontemporal.ll

 ; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s1
 ; GFX10-WGP-NEXT:    flat_load_dword v2, v[0:1]
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s3
 ; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2 slc
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2 glc slc
 ; GFX10-WGP-NEXT:    s_endpgm
 ;
 ; GFX10-CU-LABEL: flat_nontemporal_store_0:
 ; GFX10-CU:       ; %bb.0: ; %entry
 ; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s1
 ; GFX10-CU-NEXT:    flat_load_dword v2, v[0:1]
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s2
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s3
 ; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2 slc
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2 glc slc
 ; GFX10-CU-NEXT:    s_endpgm
 ;
 ; SKIP-CACHE-INV-LABEL: flat_nontemporal_store_0:
 ; SKIP-CACHE-INV:       ; %bb.0: ; %entry
 ; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
 ; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v0, s0
 ; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s1

 ; GFX10-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v2, s1
 ; GFX10-WGP-NEXT:    v_add_co_u32 v0, s0, s2, v0
 ; GFX10-WGP-NEXT:    flat_load_dword v2, v[1:2]
 ; GFX10-WGP-NEXT:    v_add_co_ci_u32_e64 v1, s0, s3, 0, s0
 ; GFX10-WGP-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2 slc
+; GFX10-WGP-NEXT:    flat_store_dword v[0:1], v2 glc slc
 ; GFX10-WGP-NEXT:    s_endpgm
 ;
 ; GFX10-CU-LABEL: flat_nontemporal_store_1:
 ; GFX10-CU:       ; %bb.0: ; %entry
 ; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; GFX10-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v2, s1
 ; GFX10-CU-NEXT:    v_add_co_u32 v0, s0, s2, v0
 ; GFX10-CU-NEXT:    flat_load_dword v2, v[1:2]
 ; GFX10-CU-NEXT:    v_add_co_ci_u32_e64 v1, s0, s3, 0, s0
 ; GFX10-CU-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2 slc
+; GFX10-CU-NEXT:    flat_store_dword v[0:1], v2 glc slc
 ; GFX10-CU-NEXT:    s_endpgm
 ;
 ; SKIP-CACHE-INV-LABEL: flat_nontemporal_store_1:
 ; SKIP-CACHE-INV:       ; %bb.0: ; %entry
 ; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; SKIP-CACHE-INV-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)
 ; SKIP-CACHE-INV-NEXT:    v_mov_b32_e32 v1, s0

llvm/test/CodeGen/AMDGPU/memory-legalizer-global-nontemporal.ll

 ; GFX10-WGP-LABEL: global_nontemporal_store_0:
 ; GFX10-WGP:       ; %bb.0: ; %entry
 ; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-WGP-NEXT:    s_load_dword s0, s[0:1], 0x0
 ; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
-; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3] slc
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3] glc slc
 ; GFX10-WGP-NEXT:    s_endpgm
 ;
 ; GFX10-CU-LABEL: global_nontemporal_store_0:
 ; GFX10-CU:       ; %bb.0: ; %entry
 ; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-CU-NEXT:    s_load_dword s0, s[0:1], 0x0
 ; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
-; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3] slc
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3] glc slc
 ; GFX10-CU-NEXT:    s_endpgm
 ;
 ; SKIP-CACHE-INV-LABEL: global_nontemporal_store_0:
 ; SKIP-CACHE-INV:       ; %bb.0: ; %entry
 ; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
 ; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, -1
 ; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)

 ; GFX10-WGP-LABEL: global_nontemporal_store_1:
 ; GFX10-WGP:       ; %bb.0: ; %entry
 ; GFX10-WGP-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; GFX10-WGP-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-WGP-NEXT:    s_load_dword s0, s[0:1], 0x0
 ; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
-; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3] slc
+; GFX10-WGP-NEXT:    global_store_dword v0, v1, s[2:3] glc slc
 ; GFX10-WGP-NEXT:    s_endpgm
 ;
 ; GFX10-CU-LABEL: global_nontemporal_store_1:
 ; GFX10-CU:       ; %bb.0: ; %entry
 ; GFX10-CU-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; GFX10-CU-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-CU-NEXT:    s_load_dword s0, s[0:1], 0x0
 ; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
-; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3] slc
+; GFX10-CU-NEXT:    global_store_dword v0, v1, s[2:3] glc slc
 ; GFX10-CU-NEXT:    s_endpgm
 ;
 ; SKIP-CACHE-INV-LABEL: global_nontemporal_store_1:
 ; SKIP-CACHE-INV:       ; %bb.0: ; %entry
 ; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x9
 ; SKIP-CACHE-INV-NEXT:    s_mov_b32 s7, 0xf000
 ; SKIP-CACHE-INV-NEXT:    s_mov_b32 s6, 0
 ; SKIP-CACHE-INV-NEXT:    v_lshlrev_b32_e32 v0, 2, v0

llvm/test/CodeGen/AMDGPU/memory-legalizer-private-nontemporal.ll

 ; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
 ; GFX10-WGP-NEXT:    s_add_u32 s8, s8, s7
 ; GFX10-WGP-NEXT:    s_addc_u32 s9, s9, 0
 ; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-WGP-NEXT:    s_load_dword s0, s[0:1], 0x0
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s2
 ; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v0, s0
-; GFX10-WGP-NEXT:    buffer_store_dword v0, v1, s[8:11], 0 offen slc
+; GFX10-WGP-NEXT:    buffer_store_dword v0, v1, s[8:11], 0 offen glc slc
 ; GFX10-WGP-NEXT:    s_endpgm
 ;
 ; GFX10-CU-LABEL: private_nontemporal_store_0:
 ; GFX10-CU:       ; %bb.0: ; %entry
 ; GFX10-CU-NEXT:    s_mov_b64 s[10:11], s[2:3]
 ; GFX10-CU-NEXT:    s_mov_b64 s[8:9], s[0:1]
 ; GFX10-CU-NEXT:    s_clause 0x1
 ; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
 ; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
 ; GFX10-CU-NEXT:    s_add_u32 s8, s8, s7
 ; GFX10-CU-NEXT:    s_addc_u32 s9, s9, 0
 ; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-CU-NEXT:    s_load_dword s0, s[0:1], 0x0
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s2
 ; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v0, s0
-; GFX10-CU-NEXT:    buffer_store_dword v0, v1, s[8:11], 0 offen slc
+; GFX10-CU-NEXT:    buffer_store_dword v0, v1, s[8:11], 0 offen glc slc
 ; GFX10-CU-NEXT:    s_endpgm
 ;
 ; SKIP-CACHE-INV-LABEL: private_nontemporal_store_0:
 ; SKIP-CACHE-INV:       ; %bb.0: ; %entry
 ; SKIP-CACHE-INV-NEXT:    s_getpc_b64 s[4:5]
 ; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s0
 ; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
 ; SKIP-CACHE-INV-NEXT:    s_waitcnt lgkmcnt(0)

 ; GFX10-WGP-NEXT:    s_load_dword s2, s[4:5], 0x8
 ; GFX10-WGP-NEXT:    s_add_u32 s8, s8, s7
 ; GFX10-WGP-NEXT:    s_addc_u32 s9, s9, 0
 ; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-WGP-NEXT:    s_load_dword s0, s[0:1], 0x0
 ; GFX10-WGP-NEXT:    v_lshl_add_u32 v0, v0, 2, s2
 ; GFX10-WGP-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-WGP-NEXT:    v_mov_b32_e32 v1, s0
-; GFX10-WGP-NEXT:    buffer_store_dword v1, v0, s[8:11], 0 offen slc
+; GFX10-WGP-NEXT:    buffer_store_dword v1, v0, s[8:11], 0 offen glc slc
 ; GFX10-WGP-NEXT:    s_endpgm
 ;
 ; GFX10-CU-LABEL: private_nontemporal_store_1:
 ; GFX10-CU:       ; %bb.0: ; %entry
 ; GFX10-CU-NEXT:    s_mov_b64 s[10:11], s[2:3]
 ; GFX10-CU-NEXT:    s_mov_b64 s[8:9], s[0:1]
 ; GFX10-CU-NEXT:    s_clause 0x1
 ; GFX10-CU-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x0
 ; GFX10-CU-NEXT:    s_load_dword s2, s[4:5], 0x8
 ; GFX10-CU-NEXT:    s_add_u32 s8, s8, s7
 ; GFX10-CU-NEXT:    s_addc_u32 s9, s9, 0
 ; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-CU-NEXT:    s_load_dword s0, s[0:1], 0x0
 ; GFX10-CU-NEXT:    v_lshl_add_u32 v0, v0, 2, s2
 ; GFX10-CU-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-CU-NEXT:    v_mov_b32_e32 v1, s0
-; GFX10-CU-NEXT:    buffer_store_dword v1, v0, s[8:11], 0 offen slc
+; GFX10-CU-NEXT:    buffer_store_dword v1, v0, s[8:11], 0 offen glc slc
 ; GFX10-CU-NEXT:    s_endpgm
 ;
 ; SKIP-CACHE-INV-LABEL: private_nontemporal_store_1:
 ; SKIP-CACHE-INV:       ; %bb.0: ; %entry
 ; SKIP-CACHE-INV-NEXT:    s_getpc_b64 s[4:5]
 ; SKIP-CACHE-INV-NEXT:    s_mov_b32 s4, s0
 ; SKIP-CACHE-INV-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x0
 ; SKIP-CACHE-INV-NEXT:    v_lshlrev_b32_e32 v0, 2, v0