[AMDGPU] gfx1250 flat and global prefetch MC support (#150455)

5 files changed, 147 insertions, 0 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index d6298c4..b72e6f0 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -268,6 +268,12 @@ def FeatureInstFwdPrefetchBug : SubtargetFeature<"inst-fwd-prefetch-bug",
   "S_INST_PREFETCH instruction causes shader to hang"
 >;
 
+def FeatureVmemPrefInsts : SubtargetFeature<"vmem-pref-insts",
+  "HasVmemPrefInsts",
+  "true",
+  "Has flat_prefect_b8 and global_prefetch_b8 instructions"
+>;
+
 def FeatureSafeSmemPrefetch : SubtargetFeature<"safe-smem-prefetch",
   "HasSafeSmemPrefetch",
   "true",
@@ -2027,6 +2033,7 @@ def FeatureISAVersion12_50 : FeatureSet<
    FeatureFlatBufferGlobalAtomicFaddF64Inst,
    FeatureMemoryAtomicFAddF32DenormalSupport,
    FeatureKernargPreload,
+   FeatureVmemPrefInsts,
    FeatureLshlAddU64Inst,
    FeatureAddSubU64Insts,
    FeatureLdsBarrierArriveAtomic,
@@ -2807,6 +2814,9 @@ def HasScalarDwordx3Loads : Predicate<"Subtarget->hasScalarDwordx3Loads()">;
 def HasXF32Insts : Predicate<"Subtarget->hasXF32Insts()">,
    AssemblerPredicate<(all_of FeatureXF32Insts)>;
 
+def HasVmemPrefInsts : Predicate<"Subtarget->hasVmemPrefInsts()">,
+  AssemblerPredicate<(all_of FeatureVmemPrefInsts)>;
+
 def HasAshrPkInsts : Predicate<"Subtarget->hasAshrPkInsts()">,
   AssemblerPredicate<(all_of FeatureAshrPkInsts)>;
 
diff --git a/llvm/lib/Target/AMDGPU/FLATInstructions.td b/llvm/lib/Target/AMDGPU/FLATInstructions.td
index 679c55d..db827f4 100644
--- a/llvm/lib/Target/AMDGPU/FLATInstructions.td
+++ b/llvm/lib/Target/AMDGPU/FLATInstructions.td
@@ -464,6 +464,37 @@ class FLAT_Global_Invalidate_Writeback<string opName, SDPatternOperator node = n
   let sve = 0;
 }
 
+class FLAT_Prefetch_Pseudo<string opName, dag addr = (ins VReg_64:$vaddr), string asm = " $vaddr"> :
+  FLAT_Pseudo<opName, (outs), !con(addr, (ins flat_offset:$offset, CPol_0:$cpol)), asm#"$offset$cpol"> {
+  let has_vdst = 0;
+  let has_data = 0;
+  let mayLoad = 1;
+  let mayStore = 1;
+  let VM_CNT = 0;
+  let LGKM_CNT = 0;
+}
+
+multiclass FLAT_Flat_Prefetch_Pseudo<string opName> {
+  def "" : FLAT_Prefetch_Pseudo<opName>,
+    GlobalSaddrTable<0, opName>;
+  def _SADDR : FLAT_Prefetch_Pseudo<opName, (ins SReg_64:$saddr, VGPR_32:$vaddr), " $vaddr, $saddr">,
+    GlobalSaddrTable<1, opName> {
+    let OtherPredicates = [HasFlatGVSMode];
+    let enabled_saddr = 1;
+  }
+}
+
+multiclass FLAT_Global_Prefetch_Pseudo<string opName> {
+  let is_flat_global = 1, has_saddr = 1 in {
+    def "" : FLAT_Prefetch_Pseudo<opName, (ins VReg_64:$vaddr), " $vaddr, off">,
+      GlobalSaddrTable<0, opName>;
+    def _SADDR : FLAT_Prefetch_Pseudo<opName, (ins SReg_64:$saddr, VGPR_32:$vaddr), " $vaddr, $saddr">,
+      GlobalSaddrTable<1, opName> {
+      let enabled_saddr = 1;
+    }
+  }
+}
+
 class FlatScratchInst <string sv_op, string mode> {
   string SVOp = sv_op;
   string Mode = mode;
@@ -1218,6 +1249,11 @@ let OtherPredicates = [HasAtomicBufferGlobalPkAddF16Insts] in
     "global_atomic_pk_add_f16", VGPR_32, v2f16
   >;
 
+let SubtargetPredicate = HasVmemPrefInsts in {
+  defm FLAT_PREFETCH_B8   : FLAT_Flat_Prefetch_Pseudo<"flat_prefetch_b8">;
+  defm GLOBAL_PREFETCH_B8 : FLAT_Global_Prefetch_Pseudo<"global_prefetch_b8">;
+}
+
 //===----------------------------------------------------------------------===//
 // Flat Patterns
 //===----------------------------------------------------------------------===//
@@ -3210,6 +3246,9 @@ multiclass VFLAT_Real_Atomics_gfx1250<bits<8> op, string name = get_FLAT_ps<NAME
 defm TENSOR_SAVE                      : VFLAT_Real_gfx1250<0x06e>;
 defm TENSOR_STOP                      : VFLAT_Real_gfx1250<0x06f>;
 
+defm FLAT_PREFETCH_B8                 : VFLAT_Real_AllAddr_gfx1250<0x05d>;
+defm GLOBAL_PREFETCH_B8               : VFLAT_Real_AllAddr_gfx1250<0x05d>;
+
 defm GLOBAL_LOAD_TR_B128_w32          : VFLAT_Real_AllAddr_gfx1250<0x057, "global_load_tr16_b128">;
 defm GLOBAL_LOAD_TR_B64_w32           : VFLAT_Real_AllAddr_gfx1250<0x058, "global_load_tr8_b64">;
 
diff --git a/llvm/lib/Target/AMDGPU/GCNSubtarget.h b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
index 0683f02..4d76b79 100644
--- a/llvm/lib/Target/AMDGPU/GCNSubtarget.h
+++ b/llvm/lib/Target/AMDGPU/GCNSubtarget.h
@@ -245,6 +245,7 @@ protected:
   bool HasVMEMtoScalarWriteHazard = false;
   bool HasSMEMtoVectorWriteHazard = false;
   bool HasInstFwdPrefetchBug = false;
+  bool HasVmemPrefInsts = false;
   bool HasSafeSmemPrefetch = false;
   bool HasVcmpxExecWARHazard = false;
   bool HasLdsBranchVmemWARHazard = false;
@@ -990,6 +991,8 @@ public:
 
   bool hasPrefetch() const { return GFX12Insts; }
 
+  bool hasVmemPrefInsts() const { return HasVmemPrefInsts; }
+
   bool hasSafeSmemPrefetch() const { return HasSafeSmemPrefetch; }
 
   // Has s_cmpk_* instructions.
diff --git a/llvm/test/MC/AMDGPU/gfx1250_asm_vflat.s b/llvm/test/MC/AMDGPU/gfx1250_asm_vflat.s
index d3a49f2..f073e90 100644
--- a/llvm/test/MC/AMDGPU/gfx1250_asm_vflat.s
+++ b/llvm/test/MC/AMDGPU/gfx1250_asm_vflat.s
@@ -109,6 +109,58 @@ scratch_store_b32 v2, v5, s1 scale_offset
 // GFX12-ERR-NEXT:{{^}}scratch_store_b32 v2, v5, s1 scale_offset
 // GFX12-ERR-NEXT:{{^}}                             ^
 
+flat_prefetch_b8 v[2:3]
+// GFX1250: flat_prefetch_b8 v[2:3]                 ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x02,0x00,0x00,0x00]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+flat_prefetch_b8 v[2:3] offset:1024    ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x02,0x00,0xfc,0xff]
+// GFX1250: flat_prefetch_b8 v[2:3] offset:1024     ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x02,0x00,0x04,0x00]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+flat_prefetch_b8 v[2:3] offset:-1024    ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x02,0x00,0xfc,0xff]
+// GFX1250: flat_prefetch_b8 v[2:3] offset:-1024    ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x02,0x00,0xfc,0xff]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+flat_prefetch_b8 v[2:3] offset:-1024 th:TH_LOAD_NT scope:SCOPE_SE ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x14,0x00,0x02,0x00,0xfc,0xff]
+// GFX1250: flat_prefetch_b8 v[2:3] offset:-1024 th:TH_LOAD_NT scope:SCOPE_SE ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x14,0x00,0x02,0x00,0xfc,0xff]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+flat_prefetch_b8 v[2:3] th:TH_LOAD_HT scope:SCOPE_CU ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x24,0x00,0x02,0x00,0xfc,0xff]
+// GFX1250: flat_prefetch_b8 v[2:3] th:TH_LOAD_HT   ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x20,0x00,0x02,0x00,0x00,0x00]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+flat_prefetch_b8 v[2:3] offset:64 th:TH_LOAD_NT_RT scope:SCOPE_DEV ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x24,0x00,0x02,0x00,0xfc,0xff]
+// GFX1250: flat_prefetch_b8 v[2:3] offset:64 th:TH_LOAD_NT_RT scope:SCOPE_DEV ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x48,0x00,0x02,0x40,0x00,0x00]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+flat_prefetch_b8 v[2:3] th:TH_LOAD_HT
+// GFX1250: flat_prefetch_b8 v[2:3] th:TH_LOAD_HT   ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x20,0x00,0x02,0x00,0x00,0x00]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+flat_prefetch_b8 v[2:3] th:TH_LOAD_BYPASS scope:SCOPE_SYS
+// GFX1250: flat_prefetch_b8 v[2:3] th:TH_LOAD_BYPASS scope:SCOPE_SYS ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x3c,0x00,0x02,0x00,0x00,0x00]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+global_prefetch_b8 v[2:3], off offset:-1024 th:TH_LOAD_HT scope:SCOPE_SE
+// GFX1250: global_prefetch_b8 v[2:3], off offset:-1024 th:TH_LOAD_HT scope:SCOPE_SE ; encoding: [0x7c,0x40,0x17,0xee,0x00,0x00,0x24,0x00,0x02,0x00,0xfc,0xff]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+global_prefetch_b8 v4, s[2:3] offset:-1024 th:TH_LOAD_NT scope:SCOPE_DEV
+// GFX1250: global_prefetch_b8 v4, s[2:3] offset:-1024 th:TH_LOAD_NT scope:SCOPE_DEV ; encoding: [0x02,0x40,0x17,0xee,0x00,0x00,0x18,0x00,0x04,0x00,0xfc,0xff]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+global_prefetch_b8 v4, s[2:3] th:TH_LOAD_RT_NT scope:SCOPE_CU
+// GFX1250: global_prefetch_b8 v4, s[2:3] th:TH_LOAD_RT_NT ; encoding: [0x02,0x40,0x17,0xee,0x00,0x00,0x50,0x00,0x04,0x00,0x00,0x00]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+global_prefetch_b8 v[2:3], off th:TH_LOAD_BYPASS scope:SCOPE_SYS
+// GFX1250: global_prefetch_b8 v[2:3], off th:TH_LOAD_BYPASS scope:SCOPE_SYS ; encoding: [0x7c,0x40,0x17,0xee,0x00,0x00,0x3c,0x00,0x02,0x00,0x00,0x00]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
+global_prefetch_b8 v[2:3], off offset:64 th:TH_LOAD_NT_RT scope:SCOPE_DEV
+// GFX1250: global_prefetch_b8 v[2:3], off offset:64 th:TH_LOAD_NT_RT scope:SCOPE_DEV ; encoding: [0x7c,0x40,0x17,0xee,0x00,0x00,0x48,0x00,0x02,0x40,0x00,0x00]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
+
 tensor_save s[0:1]
 // GFX1250: tensor_save s[0:1] ; encoding: [0x00,0x80,0x1b,0xee,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00]
 // GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
@@ -484,3 +536,7 @@ flat_store_d16_hi_b16 v2, v2, s[2:3] offset:64 scale_offset
 flat_store_d16_hi_b8 v2, v2, s[2:3] offset:64 scale_offset
 // GFX1250: flat_store_d16_hi_b8 v2, v2, s[2:3] offset:64 scale_offset ; encoding: [0x02,0x00,0x09,0xec,0x00,0x00,0x01,0x01,0x02,0x40,0x00,0x00]
 // GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: not a valid operand.
+
+flat_prefetch_b8 v3, s[2:3]
+// GFX1250: flat_prefetch_b8 v3, s[2:3]             ; encoding: [0x02,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x03,0x00,0x00,0x00]
+// GFX12-ERR: :[[@LINE-2]]:{{[0-9]+}}: error: instruction not supported on this GPU
diff --git a/llvm/test/MC/Disassembler/AMDGPU/gfx1250_dasm_vflat.txt b/llvm/test/MC/Disassembler/AMDGPU/gfx1250_dasm_vflat.txt
index 3455f4c..2f74c69 100644
--- a/llvm/test/MC/Disassembler/AMDGPU/gfx1250_dasm_vflat.txt
+++ b/llvm/test/MC/Disassembler/AMDGPU/gfx1250_dasm_vflat.txt
@@ -240,6 +240,9 @@
 # GFX1250: flat_load_u8 v1, v2, s[2:3] offset:64 scale_offset ; encoding: [0x02,0x00,0x04,0xec,0x01,0x00,0x01,0x00,0x02,0x40,0x00,0x00]
 0x02,0x00,0x04,0xec,0x01,0x00,0x01,0x00,0x02,0x40,0x00,0x00
 
+# GFX1250: flat_prefetch_b8 v3, s[2:3]             ; encoding: [0x02,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x03,0x00,0x00,0x00]
+0x02,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x03,0x00,0x00,0x00
+
 # GFX1250: flat_store_b128 v2, v[2:5], s[2:3] offset:64 scale_offset ; encoding: [0x02,0x40,0x07,0xec,0x00,0x00,0x01,0x01,0x02,0x40,0x00,0x00]
 0x02,0x40,0x07,0xec,0x00,0x00,0x01,0x01,0x02,0x40,0x00,0x00
 
@@ -3045,6 +3048,42 @@
 # GFX1250: scratch_store_b32 v2, v5, s1 scale_offset ; encoding: [0x01,0x80,0x06,0xed,0x00,0x00,0x83,0x02,0x02,0x00,0x00,0x00]
 0x01,0x80,0x06,0xed,0x00,0x00,0x83,0x02,0x02,0x00,0x00,0x00
 
+# GFX1250: flat_prefetch_b8 v[2:3]                 ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x02,0x00,0x00,0x00]
+0x7c,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x02,0x00,0x00,0x00
+
+# GFX1250: flat_prefetch_b8 v[2:3] offset:-1024    ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x02,0x00,0xfc,0xff]
+0x7c,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x02,0x00,0xfc,0xff
+
+# GFX1250: flat_prefetch_b8 v[2:3] offset:-1024 th:TH_LOAD_NT scope:SCOPE_SE ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x14,0x00,0x02,0x00,0xfc,0xff]
+0x7c,0x40,0x17,0xec,0x00,0x00,0x14,0x00,0x02,0x00,0xfc,0xff
+
+# GFX1250: flat_prefetch_b8 v[2:3] offset:-64 th:TH_LOAD_RT_NT scope:SCOPE_SYS ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x5c,0x00,0x02,0xc0,0xff,0xff]
+0x7c,0x40,0x17,0xec,0x00,0x00,0x5c,0x00,0x02,0xc0,0xff,0xff
+
+# GFX1250: flat_prefetch_b8 v[2:3] offset:1024     ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x02,0x00,0x04,0x00]
+0x7c,0x40,0x17,0xec,0x00,0x00,0x00,0x00,0x02,0x00,0x04,0x00
+
+# GFX1250: flat_prefetch_b8 v[2:3] th:TH_LOAD_BYPASS scope:SCOPE_SYS ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x3c,0x00,0x02,0x00,0x00,0x00]
+0x7c,0x40,0x17,0xec,0x00,0x00,0x3c,0x00,0x02,0x00,0x00,0x00
+
+# GFX1250: flat_prefetch_b8 v[2:3] th:TH_LOAD_HT   ; encoding: [0x7c,0x40,0x17,0xec,0x00,0x00,0x20,0x00,0x02,0x00,0x00,0x00]
+0x7c,0x40,0x17,0xec,0x00,0x00,0x20,0x00,0x02,0x00,0x00,0x00
+
+# GFX1250: global_prefetch_b8 v4, s[2:3] offset:-1024 th:TH_LOAD_NT scope:SCOPE_DEV ; encoding: [0x02,0x40,0x17,0xee,0x00,0x00,0x18,0x00,0x04,0x00,0xfc,0xff]
+0x02,0x40,0x17,0xee,0x00,0x00,0x18,0x00,0x04,0x00,0xfc,0xff
+
+# GFX1250: global_prefetch_b8 v4, s[2:3] th:TH_LOAD_RT_NT ; encoding: [0x02,0x40,0x17,0xee,0x00,0x00,0x50,0x00,0x04,0x00,0x00,0x00]
+0x02,0x40,0x17,0xee,0x00,0x00,0x50,0x00,0x04,0x00,0x00,0x00
+
+# GFX1250: global_prefetch_b8 v[2:3], off offset:-1024 th:TH_LOAD_HT scope:SCOPE_SE ; encoding: [0x7c,0x40,0x17,0xee,0x00,0x00,0x24,0x00,0x02,0x00,0xfc,0xff]
+0x7c,0x40,0x17,0xee,0x00,0x00,0x24,0x00,0x02,0x00,0xfc,0xff
+
+# GFX1250: global_prefetch_b8 v[2:3], off offset:64 th:TH_LOAD_NT_RT scope:SCOPE_DEV ; encoding: [0x7c,0x40,0x17,0xee,0x00,0x00,0x48,0x00,0x02,0x40,0x00,0x00]
+0x7c,0x40,0x17,0xee,0x00,0x00,0x48,0x00,0x02,0x40,0x00,0x00
+
+# GFX1250: global_prefetch_b8 v[2:3], off th:TH_LOAD_BYPASS scope:SCOPE_SYS ; encoding: [0x7c,0x40,0x17,0xee,0x00,0x00,0x3c,0x00,0x02,0x00,0x00,0x00]
+0x7c,0x40,0x17,0xee,0x00,0x00,0x3c,0x00,0x02,0x00,0x00,0x00
+
 # GFX1250: tensor_save s[0:1] ; encoding: [0x00,0x80,0x1b,0xee,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00]
 0x00,0x80,0x1b,0xee,0x00,0x00,0x00,0x00,0x00,0x00,0x00,0x00