[AMDGPU][gfx1250] Add cu-store subtarget feature

llvm/docs/AMDGPUUsage.rst

@@ -768,6 +768,9 @@ For example:
                                                   performant than code generated for XNACK replay
                                                   disabled.
 
+     cu-stores       TODO                         On GFX12.5, controls whether ``scope:SCOPE_CU`` stores may be used.
+                                                  If disabled, all stores will be done at ``scope:SCOPE_SE`` or greater.
+
      =============== ============================ ==================================================
 
 .. _amdgpu-target-id:
@@ -5107,7 +5110,9 @@ The fields used by CP for code objects before V3 also match those specified in
                                                      and must be 0,
      >454    1 bit   ENABLE_SGPR_PRIVATE_SEGMENT
                      _SIZE
-     457:455 3 bits                                  Reserved, must be 0.
+     455     1 bit   USES_CU_STORES                  GFX12.5: Whether the ``cu-stores`` target attribute is enabled.
+                                                     If 0, then all stores are ``SCOPE_SE`` or higher.
+     457:456 2 bits                                  Reserved, must be 0.
      458     1 bit   ENABLE_WAVEFRONT_SIZE32         GFX6-GFX9
                                                        Reserved, must be 0.
                                                      GFX10-GFX11
@@ -18188,6 +18193,8 @@ terminated by an ``.end_amdhsa_kernel`` directive.
                                                                                   GFX942)
      ``.amdhsa_user_sgpr_private_segment_size``               0                   GFX6-GFX12   Controls ENABLE_SGPR_PRIVATE_SEGMENT_SIZE in
                                                                                                :ref:`amdgpu-amdhsa-kernel-descriptor-v3-table`.
+     ``.amdhsa_uses_cu_stores``                               0                   GFX12.5      Controls USES_CU_STORES in
+                                                                                               :ref:`amdgpu-amdhsa-kernel-descriptor-v3-table`.
      ``.amdhsa_wavefront_size32``                             Target              GFX10-GFX12  Controls ENABLE_WAVEFRONT_SIZE32 in
                                                               Feature                          :ref:`amdgpu-amdhsa-kernel-descriptor-v3-table`.
                                                               Specific

llvm/include/llvm/Support/AMDHSAKernelDescriptor.h

@@ -223,7 +223,8 @@ enum : int32_t {
   KERNEL_CODE_PROPERTY(ENABLE_SGPR_DISPATCH_ID, 4, 1),
   KERNEL_CODE_PROPERTY(ENABLE_SGPR_FLAT_SCRATCH_INIT, 5, 1),
   KERNEL_CODE_PROPERTY(ENABLE_SGPR_PRIVATE_SEGMENT_SIZE, 6, 1),
-  KERNEL_CODE_PROPERTY(RESERVED0, 7, 3),
+  KERNEL_CODE_PROPERTY(RESERVED0, 7, 2),
+  KERNEL_CODE_PROPERTY(USES_CU_STORES, 9, 1),           // GFX12.5 +cu-stores
   KERNEL_CODE_PROPERTY(ENABLE_WAVEFRONT_SIZE32, 10, 1), // GFX10+
   KERNEL_CODE_PROPERTY(USES_DYNAMIC_STACK, 11, 1),
   KERNEL_CODE_PROPERTY(RESERVED1, 12, 4),

llvm/lib/Target/AMDGPU/AMDGPU.td

@@ -286,6 +286,12 @@ def FeatureSafeCUPrefetch : SubtargetFeature<"safe-cu-prefetch",
   "VMEM CU scope prefetches do not fail on illegal address"
 >;
 
+def FeatureCUStores : SubtargetFeature<"cu-stores",
+  "HasCUStores",
+  "true",
+  "Whether SCOPE_CU stores can be used on GFX12.5"
+>;
+
 def FeatureVcmpxExecWARHazard : SubtargetFeature<"vcmpx-exec-war-hazard",
   "HasVcmpxExecWARHazard",
   "true",
@@ -1988,6 +1994,7 @@ def FeatureISAVersion12 : FeatureSet<
 def FeatureISAVersion12_50 : FeatureSet<
   [FeatureGFX12,
    FeatureGFX1250Insts,
+   FeatureCUStores,
    FeatureCuMode,
    Feature64BitLiterals,
    FeatureLDSBankCount32,

llvm/lib/Target/AMDGPU/AMDGPUAsmPrinter.cpp

@@ -552,6 +552,7 @@ const MCExpr *AMDGPUAsmPrinter::getAmdhsaKernelCodeProperties(
   MCContext &Ctx = MF.getContext();
   uint16_t KernelCodeProperties = 0;
   const GCNUserSGPRUsageInfo &UserSGPRInfo = MFI.getUserSGPRInfo();
+  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
 
   if (UserSGPRInfo.hasPrivateSegmentBuffer()) {
     KernelCodeProperties |=
@@ -581,10 +582,13 @@ const MCExpr *AMDGPUAsmPrinter::getAmdhsaKernelCodeProperties(
     KernelCodeProperties |=
         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE;
   }
-  if (MF.getSubtarget<GCNSubtarget>().isWave32()) {
+  if (ST.isWave32()) {
     KernelCodeProperties |=
         amdhsa::KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32;
   }
+  if (isGFX1250(ST) && ST.hasCUStores()) {
+    KernelCodeProperties |= amdhsa::KERNEL_CODE_PROPERTY_USES_CU_STORES;
+  }
 
   // CurrentProgramInfo.DynamicCallStack is a MCExpr and could be
   // un-evaluatable at this point so it cannot be conditionally checked here.

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

@@ -6066,6 +6066,12 @@ bool AMDGPUAsmParser::ParseDirectiveAMDHSAKernel() {
                        ExprVal, ValRange);
       if (Val)
         ImpliedUserSGPRCount += 1;
+    } else if (ID == ".amdhsa_uses_cu_stores") {
+      if (!isGFX1250())
+        return Error(IDRange.Start, "directive requires gfx12.5", IDRange);
+
+      PARSE_BITS_ENTRY(KD.kernel_code_properties,
+                       KERNEL_CODE_PROPERTY_USES_CU_STORES, ExprVal, ValRange);
     } else if (ID == ".amdhsa_wavefront_size32") {
       EXPR_RESOLVE_OR_ERROR(EvaluatableExpr);
       if (IVersion.Major < 10)

llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp

@@ -2556,6 +2556,9 @@ Expected<bool> AMDGPUDisassembler::decodeKernelDescriptorDirective(
                       KERNEL_CODE_PROPERTY_ENABLE_SGPR_FLAT_SCRATCH_INIT);
     PRINT_DIRECTIVE(".amdhsa_user_sgpr_private_segment_size",
                     KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE);
+    if (isGFX1250())
+      PRINT_DIRECTIVE(".amdhsa_uses_cu_stores",
+                      KERNEL_CODE_PROPERTY_USES_CU_STORES);
 
     if (TwoByteBuffer & KERNEL_CODE_PROPERTY_RESERVED0)
       return createReservedKDBitsError(KERNEL_CODE_PROPERTY_RESERVED0,

llvm/lib/Target/AMDGPU/GCNSubtarget.h

@@ -248,6 +248,7 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   bool HasVmemPrefInsts = false;
   bool HasSafeSmemPrefetch = false;
   bool HasSafeCUPrefetch = false;
+  bool HasCUStores = false;
   bool HasVcmpxExecWARHazard = false;
   bool HasLdsBranchVmemWARHazard = false;
   bool HasNSAtoVMEMBug = false;
@@ -998,6 +999,8 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
 
   bool hasSafeCUPrefetch() const { return HasSafeCUPrefetch; }
 
+  bool hasCUStores() const { return HasCUStores; }
+
   // Has s_cmpk_* instructions.
   bool hasSCmpK() const { return getGeneration() < GFX12; }
 

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

@@ -440,6 +440,11 @@ void AMDGPUTargetAsmStreamer::EmitAmdhsaKernelDescriptor(
       amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE_SHIFT,
       amdhsa::KERNEL_CODE_PROPERTY_ENABLE_SGPR_PRIVATE_SEGMENT_SIZE,
       ".amdhsa_user_sgpr_private_segment_size");
+  if (isGFX1250(STI))
+    PrintField(KD.kernel_code_properties,
+               amdhsa::KERNEL_CODE_PROPERTY_USES_CU_STORES_SHIFT,
+               amdhsa::KERNEL_CODE_PROPERTY_USES_CU_STORES,
+               ".amdhsa_uses_cu_stores");
   if (IVersion.Major >= 10)
     PrintField(KD.kernel_code_properties,
                amdhsa::KERNEL_CODE_PROPERTY_ENABLE_WAVEFRONT_SIZE32_SHIFT,

llvm/lib/Target/AMDGPU/SIMemoryLegalizer.cpp

@@ -2564,7 +2564,9 @@ bool SIGfx12CacheControl::finalizeStore(MachineInstr &MI, bool Atomic) const {
 
   // GFX12.5 only: Require SCOPE_SE on stores that may hit the scratch address
   // space.
-  if (TII->mayAccessScratchThroughFlat(MI) && Scope == CPol::SCOPE_CU)
+  // We also require SCOPE_SE minimum if we not have the "cu-stores" feature.
+  if (Scope == CPol::SCOPE_CU &&
+      (!ST.hasCUStores() || TII->mayAccessScratchThroughFlat(MI)))
     return setScope(MI, CPol::SCOPE_SE);
 
   return false;

llvm/test/CodeGen/AMDGPU/gfx1250-no-scope-cu-stores.ll

@@ -0,0 +1,100 @@
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O3 -mcpu=gfx1250 < %s | FileCheck --check-prefixes=GCN,CU %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -O3 -mcpu=gfx1250 -mattr=-cu-stores < %s | FileCheck --check-prefixes=GCN,NOCU %s
+
+; Check that if -cu-stores is used, we use SCOPE_SE minimum on all stores.
+
+; GCN:     flat_store:
+; CU:        flat_store_b32 v{{.*}}, v{{.*}}, s{{.*}} scope:SCOPE_SE
+; NOCU:      flat_store_b32 v{{.*}}, v{{.*}}, s{{.*}} scope:SCOPE_SE
+; GCN:     .amdhsa_kernel flat_store
+; CU:       .amdhsa_uses_cu_stores 1
+; NOCU:     .amdhsa_uses_cu_stores 0
+define amdgpu_kernel void @flat_store(ptr %dst, i32 %val) {
+entry:
+  store i32 %val, ptr %dst
+  ret void
+}
+
+; GCN:     global_store:
+; CU:        global_store_b32 v{{.*}}, v{{.*}}, s{{.*}}{{$}}
+; NOCU:      global_store_b32 v{{.*}}, v{{.*}}, s{{.*}} scope:SCOPE_SE
+; GCN:     .amdhsa_kernel global_store
+; CU:        .amdhsa_uses_cu_stores 1
+; NOCU:      .amdhsa_uses_cu_stores 0
+define amdgpu_kernel void @global_store(ptr addrspace(1) %dst, i32 %val) {
+entry:
+  store i32 %val, ptr addrspace(1) %dst
+  ret void
+}
+
+; GCN:     local_store:
+; CU:        ds_store_b32 v{{.*}}, v{{.*}}{{$}}
+; NOCU:      ds_store_b32 v{{.*}}, v{{.*}}{{$}}
+; GCN:     .amdhsa_kernel local_store
+; CU:        .amdhsa_uses_cu_stores 1
+; NOCU:     .amdhsa_uses_cu_stores 0
+define amdgpu_kernel void @local_store(ptr addrspace(3) %dst, i32 %val) {
+entry:
+  store i32 %val, ptr addrspace(3) %dst
+  ret void
+}
+
+; GCN:     scratch_store:
+; CU:        scratch_store_b32 off, v{{.*}}, s{{.*}} scope:SCOPE_SE
+; NOCU:      scratch_store_b32 off, v{{.*}}, s{{.*}} scope:SCOPE_SE
+; GCN:     .amdhsa_kernel scratch_store
+; CU:        .amdhsa_uses_cu_stores 1
+; NOCU:      .amdhsa_uses_cu_stores 0
+define amdgpu_kernel void @scratch_store(ptr addrspace(5) %dst, i32 %val) {
+entry:
+  store i32 %val, ptr addrspace(5) %dst
+  ret void
+}
+
+; GCN:     flat_atomic_store:
+; CU:        flat_store_b32 v{{.*}}, v{{.*}}, s{{.*}} scope:SCOPE_SE
+; NOCU:      flat_store_b32 v{{.*}}, v{{.*}}, s{{.*}} scope:SCOPE_SE
+; GCN:     .amdhsa_kernel flat_atomic_store
+; CU:        .amdhsa_uses_cu_stores 1
+; NOCU:      .amdhsa_uses_cu_stores 0
+define amdgpu_kernel void @flat_atomic_store(ptr %dst, i32 %val) {
+entry:
+  store atomic i32 %val, ptr %dst syncscope("wavefront") unordered, align 4
+  ret void
+}
+
+; GCN:     global_atomic_store:
+; CU:        global_store_b32 v{{.*}}, v{{.*}}, s{{.*}}{{$}}
+; NOCU:      global_store_b32  v{{.*}}, v{{.*}}, s{{.*}} scope:SCOPE_SE
+; GCN:     .amdhsa_kernel global_atomic_store
+; CU:        .amdhsa_uses_cu_stores 1
+; NOCU:      .amdhsa_uses_cu_stores 0
+define amdgpu_kernel void @global_atomic_store(ptr addrspace(1) %dst, i32 %val) {
+entry:
+  store atomic i32 %val, ptr addrspace(1) %dst syncscope("wavefront") unordered, align 4
+  ret void
+}
+
+; GCN:     local_atomic_store:
+; CU:        ds_store_b32 v{{.*}}, v{{.*}}{{$}}
+; NOCU:      ds_store_b32 v{{.*}}, v{{.*}}{{$}}
+; GCN:     .amdhsa_kernel local_atomic_store
+; CU:        .amdhsa_uses_cu_stores 1
+; NOCU:      .amdhsa_uses_cu_stores 0
+define amdgpu_kernel void @local_atomic_store(ptr addrspace(3) %dst, i32 %val) {
+entry:
+  store atomic i32 %val, ptr addrspace(3) %dst syncscope("wavefront") unordered, align 4
+  ret void
+}
+
+; GCN:     scratch_atomic_store:
+; CU:        scratch_store_b32 off, v{{.*}}, s{{.*}} scope:SCOPE_SE
+; NOCU:      scratch_store_b32 off, v{{.*}}, s{{.*}} scope:SCOPE_SE
+; GCN:     .amdhsa_kernel scratch_atomic_store
+; CU:        .amdhsa_uses_cu_stores 1
+; NOCU:      .amdhsa_uses_cu_stores 0
+define amdgpu_kernel void @scratch_atomic_store(ptr addrspace(5) %dst, i32 %val) {
+entry:
+  store atomic i32 %val, ptr addrspace(5) %dst syncscope("wavefront") unordered, align 4
+  ret void
+}

llvm/test/MC/Disassembler/AMDGPU/kernel-descriptor-errors.test

@@ -13,10 +13,10 @@
 # RES_4_2: ; error decoding test.kd: kernel descriptor reserved bits in range (511:480) set
 # RES_4_2-NEXT: ; decoding failed region as bytes
 
-# RUN: yaml2obj %s -DGPU=GFX90A -DKD=00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000006000000000000 \
-# RUN:   | llvm-objdump --disassemble-symbols=test.kd - | FileCheck %s --check-prefix=RES_457
-# RES_457: ; error decoding test.kd: kernel descriptor reserved bits in range (457:455) set
-# RES_457-NEXT: ; decoding failed region as bytes
+# RUN: yaml2obj %s -DGPU=GFX90A -DKD=00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000003000000000000 \
+# RUN:   | llvm-objdump --disassemble-symbols=test.kd - | FileCheck %s --check-prefix=RES_456
+# RES_456: ; error decoding test.kd: kernel descriptor reserved bits in range (456:455) set
+# RES_456-NEXT: ; decoding failed region as bytes
 
 # RUN: yaml2obj %s -DGPU=GFX90A -DKD=0000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000c000000000000 \
 # RUN:   | llvm-objdump --disassemble-symbols=test.kd - | FileCheck %s --check-prefix=WF32