[AMDGPU][GFX12] Add Atomic cond_sub_u32 #76224
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
mariusz-sikora-at-amd
requested review from
jayfoad,
arsenm,
piotrAMD and
mbrkusanin
|
@llvm/pr-subscribers-llvm-analysis @llvm/pr-subscribers-backend-amdgpu Author: Mariusz Sikora (mariusz-sikora-at-amd) ChangesPatch is 66.35 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/76224.diff 23 Files Affected:
diff --git a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
index cb48f54b13a6cd..2d066350ee9f84 100644
--- a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
+++ b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
@@ -10,6 +10,10 @@
//
//===----------------------------------------------------------------------===//
+def flat_ptr_ty : LLVMQualPointerType<0>;
+def global_ptr_ty : LLVMQualPointerType<1>;
+def local_ptr_ty : LLVMQualPointerType<3>;
+
class AMDGPUReadPreloadRegisterIntrinsic
: DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrNoMem, IntrSpeculatable]>;
@@ -1243,6 +1247,7 @@ def int_amdgcn_raw_buffer_atomic_or : AMDGPURawBufferAtomic;
def int_amdgcn_raw_buffer_atomic_xor : AMDGPURawBufferAtomic;
def int_amdgcn_raw_buffer_atomic_inc : AMDGPURawBufferAtomic;
def int_amdgcn_raw_buffer_atomic_dec : AMDGPURawBufferAtomic;
+def int_amdgcn_raw_buffer_atomic_cond_sub_u32 : AMDGPURawBufferAtomic;
def int_amdgcn_raw_buffer_atomic_cmpswap : Intrinsic<
[llvm_anyint_ty],
[LLVMMatchType<0>, // src(VGPR)
@@ -1279,6 +1284,7 @@ def int_amdgcn_raw_ptr_buffer_atomic_or : AMDGPURawPtrBufferAtomic;
def int_amdgcn_raw_ptr_buffer_atomic_xor : AMDGPURawPtrBufferAtomic;
def int_amdgcn_raw_ptr_buffer_atomic_inc : AMDGPURawPtrBufferAtomic;
def int_amdgcn_raw_ptr_buffer_atomic_dec : AMDGPURawPtrBufferAtomic;
+def int_amdgcn_raw_ptr_buffer_atomic_cond_sub_u32 : AMDGPURawPtrBufferAtomic;
def int_amdgcn_raw_ptr_buffer_atomic_cmpswap : Intrinsic<
[llvm_anyint_ty],
[LLVMMatchType<0>, // src(VGPR)
@@ -1317,6 +1323,7 @@ def int_amdgcn_struct_buffer_atomic_or : AMDGPUStructBufferAtomic;
def int_amdgcn_struct_buffer_atomic_xor : AMDGPUStructBufferAtomic;
def int_amdgcn_struct_buffer_atomic_inc : AMDGPUStructBufferAtomic;
def int_amdgcn_struct_buffer_atomic_dec : AMDGPUStructBufferAtomic;
+def int_amdgcn_struct_buffer_atomic_cond_sub_u32 : AMDGPUStructBufferAtomic;
def int_amdgcn_struct_buffer_atomic_cmpswap : Intrinsic<
[llvm_anyint_ty],
[LLVMMatchType<0>, // src(VGPR)
@@ -1352,6 +1359,7 @@ def int_amdgcn_struct_ptr_buffer_atomic_or : AMDGPUStructPtrBufferAtomic;
def int_amdgcn_struct_ptr_buffer_atomic_xor : AMDGPUStructPtrBufferAtomic;
def int_amdgcn_struct_ptr_buffer_atomic_inc : AMDGPUStructPtrBufferAtomic;
def int_amdgcn_struct_ptr_buffer_atomic_dec : AMDGPUStructPtrBufferAtomic;
+def int_amdgcn_struct_ptr_buffer_atomic_cond_sub_u32 : AMDGPUStructPtrBufferAtomic;
def int_amdgcn_struct_ptr_buffer_atomic_cmpswap : Intrinsic<
[llvm_anyint_ty],
[LLVMMatchType<0>, // src(VGPR)
@@ -2353,10 +2361,10 @@ def int_amdgcn_s_get_waveid_in_workgroup :
Intrinsic<[llvm_i32_ty], [],
[IntrNoMem, IntrHasSideEffects, IntrWillReturn, IntrNoCallback, IntrNoFree]>;
-class AMDGPUAtomicRtn<LLVMType vt> : Intrinsic <
+class AMDGPUAtomicRtn<LLVMType vt, LLVMType pt = llvm_anyptr_ty> : Intrinsic <
[vt],
- [llvm_anyptr_ty, // vaddr
- vt], // vdata(VGPR)
+ [pt, // vaddr
+ vt], // vdata(VGPR)
[IntrArgMemOnly, IntrWillReturn, NoCapture<ArgIndex<0>>, IntrNoCallback, IntrNoFree], "",
[SDNPMemOperand]>;
@@ -2491,6 +2499,11 @@ def int_amdgcn_flat_atomic_fmax_num : AMDGPUAtomicRtn<llvm_anyfloat_ty>;
def int_amdgcn_global_atomic_fmin_num : AMDGPUAtomicRtn<llvm_anyfloat_ty>;
def int_amdgcn_global_atomic_fmax_num : AMDGPUAtomicRtn<llvm_anyfloat_ty>;
+def int_amdgcn_flat_atomic_cond_sub_u32 : AMDGPUAtomicRtn<llvm_i32_ty, flat_ptr_ty>;
+def int_amdgcn_global_atomic_cond_sub_u32 : AMDGPUAtomicRtn<llvm_i32_ty, global_ptr_ty>;
+
+def int_amdgcn_ds_cond_sub_u32 : AMDGPUAtomicRtn<llvm_i32_ty, local_ptr_ty>;
+
//===----------------------------------------------------------------------===//
// Deep learning intrinsics.
//===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUGISel.td b/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
index 2b85024a9b40be..801c5fa2e1565d 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
@@ -264,6 +264,7 @@ def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_FADD, SIbuffer_atomic_fadd>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_FMIN, SIbuffer_atomic_fmin>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_FMAX, SIbuffer_atomic_fmax>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_CMPSWAP, SIbuffer_atomic_cmpswap>;
+def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_COND_SUB_U32, SIbuffer_atomic_cond_sub_u32>;
def : GINodeEquiv<G_AMDGPU_S_BUFFER_LOAD, SIsbuffer_load>;
def : GINodeEquiv<G_FPTRUNC_ROUND_UPWARD, SIfptrunc_round_upward>;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 4bf4707553e5fe..609c70a60d6463 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -5438,6 +5438,7 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
NODE_NAME_CASE(BUFFER_ATOMIC_FADD)
NODE_NAME_CASE(BUFFER_ATOMIC_FMIN)
NODE_NAME_CASE(BUFFER_ATOMIC_FMAX)
+ NODE_NAME_CASE(BUFFER_ATOMIC_COND_SUB_U32)
case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break;
}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index 827fb106b55199..6828db6e0220d5 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -589,6 +589,7 @@ enum NodeType : unsigned {
BUFFER_ATOMIC_FADD,
BUFFER_ATOMIC_FMIN,
BUFFER_ATOMIC_FMAX,
+ BUFFER_ATOMIC_COND_SUB_U32,
LAST_AMDGPU_ISD_NUMBER
};
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
index eaf72d7157ee2d..e3c6f46c758e89 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
@@ -642,10 +642,14 @@ defm int_amdgcn_global_atomic_fmax : noret_op;
defm int_amdgcn_global_atomic_csub : noret_op;
defm int_amdgcn_flat_atomic_fadd : local_addr_space_atomic_op;
defm int_amdgcn_ds_fadd_v2bf16 : noret_op;
+defm int_amdgcn_flat_atomic_cond_sub_u32 : noret_op;
+defm int_amdgcn_global_atomic_cond_sub_u32 : noret_op;
defm int_amdgcn_flat_atomic_fmin_num : noret_op;
defm int_amdgcn_flat_atomic_fmax_num : noret_op;
defm int_amdgcn_global_atomic_fmin_num : noret_op;
defm int_amdgcn_global_atomic_fmax_num : noret_op;
+defm int_amdgcn_ds_cond_sub_u32 : noret_op;
+defm int_amdgcn_ds_cond_sub_u32 : local_addr_space_atomic_op;
multiclass noret_binary_atomic_op<SDNode atomic_op, bit IsInt = 1> {
let HasNoUse = true in
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index fbee2888945185..e605986564f2fc 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -5882,6 +5882,9 @@ static unsigned getBufferAtomicPseudo(Intrinsic::ID IntrID) {
case Intrinsic::amdgcn_struct_buffer_atomic_fmax:
case Intrinsic::amdgcn_struct_ptr_buffer_atomic_fmax:
return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_FMAX;
+ case Intrinsic::amdgcn_raw_buffer_atomic_cond_sub_u32:
+ case Intrinsic::amdgcn_struct_buffer_atomic_cond_sub_u32:
+ return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_COND_SUB_U32;
default:
llvm_unreachable("unhandled atomic opcode");
}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
index c9412f720c62ec..51d2bb130774f9 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -4690,6 +4690,9 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
case Intrinsic::amdgcn_flat_atomic_fmax_num:
case Intrinsic::amdgcn_global_atomic_fadd_v2bf16:
case Intrinsic::amdgcn_flat_atomic_fadd_v2bf16:
+ case Intrinsic::amdgcn_flat_atomic_cond_sub_u32:
+ case Intrinsic::amdgcn_global_atomic_cond_sub_u32:
+ case Intrinsic::amdgcn_ds_cond_sub_u32:
return getDefaultMappingAllVGPR(MI);
case Intrinsic::amdgcn_ds_ordered_add:
case Intrinsic::amdgcn_ds_ordered_swap:
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td b/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td
index beb670669581f1..1a28197fd90265 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td
@@ -243,17 +243,20 @@ def : SourceOfDivergence<int_amdgcn_global_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_global_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_global_atomic_fmin_num>;
def : SourceOfDivergence<int_amdgcn_global_atomic_fmax_num>;
+def : SourceOfDivergence<int_amdgcn_global_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fadd>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fmin_num>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fmax_num>;
+def : SourceOfDivergence<int_amdgcn_flat_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_global_atomic_fadd_v2bf16>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fadd_v2bf16>;
def : SourceOfDivergence<int_amdgcn_ds_fadd>;
def : SourceOfDivergence<int_amdgcn_ds_fmin>;
def : SourceOfDivergence<int_amdgcn_ds_fmax>;
def : SourceOfDivergence<int_amdgcn_ds_fadd_v2bf16>;
+def : SourceOfDivergence<int_amdgcn_ds_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_buffer_atomic_swap>;
def : SourceOfDivergence<int_amdgcn_buffer_atomic_add>;
def : SourceOfDivergence<int_amdgcn_buffer_atomic_sub>;
@@ -281,6 +284,7 @@ def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_fadd>;
def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_cmpswap>;
+def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_swap>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_add>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_sub>;
@@ -297,6 +301,7 @@ def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_fadd>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_cmpswap>;
+def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_swap>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_add>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_sub>;
@@ -313,6 +318,7 @@ def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_fadd>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_cmpswap>;
+def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_swap>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_add>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_sub>;
@@ -329,6 +335,7 @@ def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_fadd>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_cmpswap>;
+def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_buffer_atomic_csub>;
def : SourceOfDivergence<int_amdgcn_ps_live>;
def : SourceOfDivergence<int_amdgcn_live_mask>;
diff --git a/llvm/lib/Target/AMDGPU/BUFInstructions.td b/llvm/lib/Target/AMDGPU/BUFInstructions.td
index 43d35fa5291ca0..fb4ef8620b795d 100644
--- a/llvm/lib/Target/AMDGPU/BUFInstructions.td
+++ b/llvm/lib/Target/AMDGPU/BUFInstructions.td
@@ -1245,6 +1245,12 @@ defm BUFFER_ATOMIC_PK_ADD_F16 : MUBUF_Pseudo_Atomics_RTN <
"buffer_atomic_pk_add_f16", VGPR_32, v2f16, null_frag
>;
+let SubtargetPredicate = isGFX12Plus in {
+defm BUFFER_ATOMIC_COND_SUB_U32 : MUBUF_Pseudo_Atomics <
+ "buffer_atomic_cond_sub_u32", VGPR_32, i32
+>;
+}
+
//===----------------------------------------------------------------------===//
// MTBUF Instructions
//===----------------------------------------------------------------------===//
@@ -1708,6 +1714,13 @@ defm : SIBufferAtomicPat<"SIbuffer_atomic_dec", i64, "BUFFER_ATOMIC_DEC_X2">;
let SubtargetPredicate = HasAtomicCSubNoRtnInsts in
defm : SIBufferAtomicPat<"SIbuffer_atomic_csub", i32, "BUFFER_ATOMIC_CSUB", ["noret"]>;
+let SubtargetPredicate = isGFX12Plus in {
+ defm : SIBufferAtomicPat_Common<"SIbuffer_atomic_cond_sub_u32", i32, "BUFFER_ATOMIC_COND_SUB_U32_VBUFFER", ["ret"]>;
+
+ let OtherPredicates = [HasAtomicCSubNoRtnInsts] in
+ defm : SIBufferAtomicPat_Common<"SIbuffer_atomic_cond_sub_u32", i32, "BUFFER_ATOMIC_COND_SUB_U32_VBUFFER", ["noret"]>;
+}
+
let SubtargetPredicate = isGFX6GFX7GFX10Plus in {
defm : SIBufferAtomicPat<"SIbuffer_atomic_fmin", f32, "BUFFER_ATOMIC_FMIN">;
defm : SIBufferAtomicPat<"SIbuffer_atomic_fmax", f32, "BUFFER_ATOMIC_FMAX">;
@@ -2610,6 +2623,7 @@ defm BUFFER_ATOMIC_AND_X2 : MUBUF_Real_Atomic_gfx11_gfx12_Renamed<0x049,
defm BUFFER_ATOMIC_CMPSWAP : MUBUF_Real_Atomic_gfx11_gfx12_Renamed<0x034, "buffer_atomic_cmpswap_b32">;
defm BUFFER_ATOMIC_CMPSWAP_X2 : MUBUF_Real_Atomic_gfx11_gfx12_Renamed<0x042, "buffer_atomic_cmpswap_b64">;
defm BUFFER_ATOMIC_FCMPSWAP : MUBUF_Real_Atomic_gfx11_Renamed<0x050, "buffer_atomic_cmpswap_f32">;
+defm BUFFER_ATOMIC_COND_SUB_U32 : MUBUF_Real_Atomic_gfx12<0x050>;
defm BUFFER_ATOMIC_CSUB : MUBUF_Real_Atomic_gfx11_gfx12_Renamed_gfx12_Renamed<0x037, "buffer_atomic_sub_clamp_u32", "buffer_atomic_csub_u32">;
def : Mnem_gfx11_gfx12<"buffer_atomic_csub", "buffer_atomic_csub_u32">;
defm BUFFER_ATOMIC_DEC : MUBUF_Real_Atomic_gfx11_gfx12_Renamed<0x040, "buffer_atomic_dec_u32">;
diff --git a/llvm/lib/Target/AMDGPU/DSInstructions.td b/llvm/lib/Target/AMDGPU/DSInstructions.td
index bc9049b4ef33c0..fb79d9264a3583 100644
--- a/llvm/lib/Target/AMDGPU/DSInstructions.td
+++ b/llvm/lib/Target/AMDGPU/DSInstructions.td
@@ -438,6 +438,12 @@ class DS_1A1D_PERMUTE <string opName, SDPatternOperator node = null_frag,
let has_gds = 0;
}
+class DSAtomicRetPat<DS_Pseudo inst, ValueType vt, PatFrag frag, int complexity = 0,
+ bit gds=0> : GCNPat <(frag (DS1Addr1Offset i32:$ptr, i32:$offset), vt:$value),
+ (inst $ptr, getVregSrcForVT<vt>.ret:$value, offset:$offset, (i1 gds))> {
+ let AddedComplexity = complexity;
+}
+
defm DS_ADD_U32 : DS_1A1D_NORET_mc<"ds_add_u32">;
defm DS_SUB_U32 : DS_1A1D_NORET_mc<"ds_sub_u32">;
defm DS_RSUB_U32 : DS_1A1D_NORET_mc<"ds_rsub_u32">;
@@ -733,9 +739,22 @@ def DS_BVH_STACK_RTN_B32 : DS_BVH_STACK<"ds_bvh_stack_rtn_b32">;
let SubtargetPredicate = isGFX12Plus in {
+defm DS_COND_SUB_U32 : DS_1A1D_NORET_mc<"ds_cond_sub_u32">;
+defm DS_COND_SUB_RTN_U32 : DS_1A1D_RET_mc<"ds_cond_sub_rtn_u32", VGPR_32, "ds_cond_sub_u32">;
defm DS_SUB_CLAMP_U32 : DS_1A1D_NORET_mc<"ds_sub_clamp_u32">;
defm DS_SUB_CLAMP_RTN_U32 : DS_1A1D_RET_mc<"ds_sub_clamp_rtn_u32", VGPR_32, "ds_sub_clamp_u32">;
+multiclass DSAtomicRetNoRetPatIntrinsic_mc<DS_Pseudo inst, DS_Pseudo noRetInst,
+ ValueType vt, string frag> {
+ def : DSAtomicRetPat<inst, vt,
+ !cast<PatFrag>(frag#"_local_addrspace")>;
+
+ let OtherPredicates = [HasAtomicCSubNoRtnInsts] in
+ def : DSAtomicRetPat<noRetInst, vt,
+ !cast<PatFrag>(frag#"_noret_local_addrspace"), /* complexity */ 1>;
+}
+
+defm : DSAtomicRetNoRetPatIntrinsic_mc<DS_COND_SUB_RTN_U32, DS_COND_SUB_U32, i32, "int_amdgcn_ds_cond_sub_u32">;
} // let SubtargetPredicate = isGFX12Plus
//===----------------------------------------------------------------------===//
@@ -955,12 +974,6 @@ defm : DSWritePat_mc <DS_WRITE_B128, vt, "store_align_less_than_4_local">;
} // End AddedComplexity = 100
-class DSAtomicRetPat<DS_Pseudo inst, ValueType vt, PatFrag frag, int complexity = 0,
- bit gds=0> : GCNPat <(frag (DS1Addr1Offset i32:$ptr, i32:$offset), vt:$value),
- (inst $ptr, getVregSrcForVT<vt>.ret:$value, offset:$offset, (i1 gds))> {
- let AddedComplexity = complexity;
-}
-
multiclass DSAtomicRetPat_mc<DS_Pseudo inst, ValueType vt, string frag> {
let OtherPredicates = [LDSRequiresM0Init] in {
def : DSAtomicRetPat<inst, vt, !cast<PatFrag>(frag#"_local_m0_"#vt.Size)>;
@@ -1238,7 +1251,9 @@ defm DS_MIN_NUM_F64 : DS_Real_Renamed_gfx12<0x052, DS_MIN_F64, "ds_min_num
defm DS_MAX_NUM_F64 : DS_Real_Renamed_gfx12<0x053, DS_MAX_F64, "ds_max_num_f64">;
defm DS_MIN_NUM_RTN_F64 : DS_Real_Renamed_gfx12<0x072, DS_MIN_RTN_F64, "ds_min_num_rtn_f64">;
defm DS_MAX_NUM_RTN_F64 : DS_Real_Renamed_gfx12<0x073, DS_MAX_RTN_F64, "ds_max_num_rtn_f64">;
+defm DS_COND_SUB_U32 : DS_Real_gfx12<0x098>;
defm DS_SUB_CLAMP_U32 : DS_Real_gfx12<0x099>;
+defm DS_COND_SUB_RTN_U32 : DS_Real_gfx12<0x0a8>;
defm DS_SUB_CLAMP_RTN_U32 : DS_Real_gfx12<0x0a9>;
//===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/AMDGPU/FLATInstructions.td b/llvm/lib/Target/AMDGPU/FLATInstructions.td
index 0dd2b3f5c2c912..4d5ebf11d7232f 100644
--- a/llvm/lib/Target/AMDGPU/FLATInstructions.td
+++ b/llvm/lib/Target/AMDGPU/FLATInstructions.td
@@ -801,6 +801,7 @@ let SubtargetPredicate = HasFlatAtomicFaddF32Inst in {
let SubtargetPredicate = isGFX12Plus in {
defm FLAT_ATOMIC_CSUB_U32 : FLAT_Atomic_Pseudo <"flat_atomic_csub_u32", VGPR_32, i32>;
+ defm FLAT_ATOMIC_COND_SUB_U32 : FLAT_Atomic_Pseudo <"flat_atomic_cond_sub_u32", VGPR_32, i32>;
} // End SubtargetPredicate = isGFX12Plus
defm GLOBAL_LOAD_UBYTE : FLAT_Global_Load_Pseudo <"global_load_ubyte", VGPR_32>;
@@ -926,6 +927,10 @@ defm GLOBAL_LOAD_LDS_USHORT : FLAT_Global_Load_LDS_Pseudo <"global_load_lds_usho
defm GLOBAL_LOAD_LDS_SSHORT : FLAT_Global_Load_LDS_Pseudo <"global_load_lds_sshort">;
defm GLOBAL_LOAD_LDS_DWORD : FLAT_Global_Load_LDS_Pseudo <"global_load_lds_dword">;
+let SubtargetPredicate = isGFX12Plus in {
+ defm GLOBAL_ATOMIC_COND_SUB_U32 : FLAT_Global_Atomic_Pseudo <"global_atomic_cond_sub_u32", VGPR_32, i32>;
+} // End SubtargetPredicate = isGFX12Plus
+
} // End is_flat_global = 1
@@ -1268,6 +1273,13 @@ defm : FlatAtomicPat <"FLAT_ATOMIC_CMPSWAP_X2", "AMDGPUatomic_cmp_swap_"#as, i64
defm : FlatAtomicPat <"FLAT_ATOMIC_XOR_X2", "atomic_load_xor_"#as, i64>;
} // end foreach as
+let SubtargetPredicate = isGFX12Plus in {
+ defm : FlatAtomicIntrRtnPat<"FLAT_ATOMIC_COND_SUB_U32", "int_amdgcn_flat_atomic_cond_sub_u32", i32>;
+
+ let OtherPredicates = [HasAtomicCSubNoRtnInsts] in
+ defm : FlatAtomicIntrNoRtnPat<"FLAT_ATOMIC_COND_SUB_U32", "int_amdgcn_flat_atomic_cond_sub_u32", i32>;
+}
+
def : FlatStorePat <FLAT_STORE_BYTE, truncstorei8_flat, i16>;
def : FlatStorePat <FLAT_STORE_SHORT, store_flat, i16>;
@@ -1377,6 +1389,16 @@ multiclass GlobalFLATAtomicIntrPats<string inst, string node, ValueType vt,
defm : GlobalFLATAtomicPats<inst, node, vt, data_vt, /* isIntr */ 1>;
}
+multiclass GlobalFLATAtomicIntrPatsNoRtn<string inst, string node, ValueType vt,
+ ValueType data_vt = vt> {
+ defm : GlobalFLATAtomicPatsNoRtn<inst, node, vt, data_vt, /* isIntr */ 1>;
+}
+
+multiclass GlobalFLATAtomicIntrPatsRtn<string inst, string node, ValueType vt,
+ ValueType data_vt = vt> {
+ defm : GlobalFLATAtomicPatsRtn<inst, node, vt, data_vt, /* isIntr */ 1>;
+}
+
multiclass ScratchFLATLoadPats<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
def : ScratchLoadSignedPat <inst, node, vt> {
let AddedComplexity = 25;
@@ -1529,6 +1551,13 @@ defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SWAP_X2", "atomic_swap_global", i64>
defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_CMPSWAP_X2", "AMDGPUatomic_cmp_swap_global", i64, v2i64>;
defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_XOR_X2", "atomic_load_xor_global", i64>;
+let SubtargetPredicate = isGFX12Plus in {
+ defm : GlobalFLATAtomicIntrPatsRtn <"GLOBAL_ATOMIC_COND_SUB_U32", "int_amdgcn_global_atomi...
[truncated]
|
|
@llvm/pr-subscribers-llvm-ir Author: Mariusz Sikora (mariusz-sikora-at-amd) ChangesPatch is 66.35 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/76224.diff 23 Files Affected:
diff --git a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
index cb48f54b13a6cd..2d066350ee9f84 100644
--- a/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
+++ b/llvm/include/llvm/IR/IntrinsicsAMDGPU.td
@@ -10,6 +10,10 @@
//
//===----------------------------------------------------------------------===//
+def flat_ptr_ty : LLVMQualPointerType<0>;
+def global_ptr_ty : LLVMQualPointerType<1>;
+def local_ptr_ty : LLVMQualPointerType<3>;
+
class AMDGPUReadPreloadRegisterIntrinsic
: DefaultAttrsIntrinsic<[llvm_i32_ty], [], [IntrNoMem, IntrSpeculatable]>;
@@ -1243,6 +1247,7 @@ def int_amdgcn_raw_buffer_atomic_or : AMDGPURawBufferAtomic;
def int_amdgcn_raw_buffer_atomic_xor : AMDGPURawBufferAtomic;
def int_amdgcn_raw_buffer_atomic_inc : AMDGPURawBufferAtomic;
def int_amdgcn_raw_buffer_atomic_dec : AMDGPURawBufferAtomic;
+def int_amdgcn_raw_buffer_atomic_cond_sub_u32 : AMDGPURawBufferAtomic;
def int_amdgcn_raw_buffer_atomic_cmpswap : Intrinsic<
[llvm_anyint_ty],
[LLVMMatchType<0>, // src(VGPR)
@@ -1279,6 +1284,7 @@ def int_amdgcn_raw_ptr_buffer_atomic_or : AMDGPURawPtrBufferAtomic;
def int_amdgcn_raw_ptr_buffer_atomic_xor : AMDGPURawPtrBufferAtomic;
def int_amdgcn_raw_ptr_buffer_atomic_inc : AMDGPURawPtrBufferAtomic;
def int_amdgcn_raw_ptr_buffer_atomic_dec : AMDGPURawPtrBufferAtomic;
+def int_amdgcn_raw_ptr_buffer_atomic_cond_sub_u32 : AMDGPURawPtrBufferAtomic;
def int_amdgcn_raw_ptr_buffer_atomic_cmpswap : Intrinsic<
[llvm_anyint_ty],
[LLVMMatchType<0>, // src(VGPR)
@@ -1317,6 +1323,7 @@ def int_amdgcn_struct_buffer_atomic_or : AMDGPUStructBufferAtomic;
def int_amdgcn_struct_buffer_atomic_xor : AMDGPUStructBufferAtomic;
def int_amdgcn_struct_buffer_atomic_inc : AMDGPUStructBufferAtomic;
def int_amdgcn_struct_buffer_atomic_dec : AMDGPUStructBufferAtomic;
+def int_amdgcn_struct_buffer_atomic_cond_sub_u32 : AMDGPUStructBufferAtomic;
def int_amdgcn_struct_buffer_atomic_cmpswap : Intrinsic<
[llvm_anyint_ty],
[LLVMMatchType<0>, // src(VGPR)
@@ -1352,6 +1359,7 @@ def int_amdgcn_struct_ptr_buffer_atomic_or : AMDGPUStructPtrBufferAtomic;
def int_amdgcn_struct_ptr_buffer_atomic_xor : AMDGPUStructPtrBufferAtomic;
def int_amdgcn_struct_ptr_buffer_atomic_inc : AMDGPUStructPtrBufferAtomic;
def int_amdgcn_struct_ptr_buffer_atomic_dec : AMDGPUStructPtrBufferAtomic;
+def int_amdgcn_struct_ptr_buffer_atomic_cond_sub_u32 : AMDGPUStructPtrBufferAtomic;
def int_amdgcn_struct_ptr_buffer_atomic_cmpswap : Intrinsic<
[llvm_anyint_ty],
[LLVMMatchType<0>, // src(VGPR)
@@ -2353,10 +2361,10 @@ def int_amdgcn_s_get_waveid_in_workgroup :
Intrinsic<[llvm_i32_ty], [],
[IntrNoMem, IntrHasSideEffects, IntrWillReturn, IntrNoCallback, IntrNoFree]>;
-class AMDGPUAtomicRtn<LLVMType vt> : Intrinsic <
+class AMDGPUAtomicRtn<LLVMType vt, LLVMType pt = llvm_anyptr_ty> : Intrinsic <
[vt],
- [llvm_anyptr_ty, // vaddr
- vt], // vdata(VGPR)
+ [pt, // vaddr
+ vt], // vdata(VGPR)
[IntrArgMemOnly, IntrWillReturn, NoCapture<ArgIndex<0>>, IntrNoCallback, IntrNoFree], "",
[SDNPMemOperand]>;
@@ -2491,6 +2499,11 @@ def int_amdgcn_flat_atomic_fmax_num : AMDGPUAtomicRtn<llvm_anyfloat_ty>;
def int_amdgcn_global_atomic_fmin_num : AMDGPUAtomicRtn<llvm_anyfloat_ty>;
def int_amdgcn_global_atomic_fmax_num : AMDGPUAtomicRtn<llvm_anyfloat_ty>;
+def int_amdgcn_flat_atomic_cond_sub_u32 : AMDGPUAtomicRtn<llvm_i32_ty, flat_ptr_ty>;
+def int_amdgcn_global_atomic_cond_sub_u32 : AMDGPUAtomicRtn<llvm_i32_ty, global_ptr_ty>;
+
+def int_amdgcn_ds_cond_sub_u32 : AMDGPUAtomicRtn<llvm_i32_ty, local_ptr_ty>;
+
//===----------------------------------------------------------------------===//
// Deep learning intrinsics.
//===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUGISel.td b/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
index 2b85024a9b40be..801c5fa2e1565d 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUGISel.td
@@ -264,6 +264,7 @@ def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_FADD, SIbuffer_atomic_fadd>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_FMIN, SIbuffer_atomic_fmin>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_FMAX, SIbuffer_atomic_fmax>;
def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_CMPSWAP, SIbuffer_atomic_cmpswap>;
+def : GINodeEquiv<G_AMDGPU_BUFFER_ATOMIC_COND_SUB_U32, SIbuffer_atomic_cond_sub_u32>;
def : GINodeEquiv<G_AMDGPU_S_BUFFER_LOAD, SIsbuffer_load>;
def : GINodeEquiv<G_FPTRUNC_ROUND_UPWARD, SIfptrunc_round_upward>;
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 4bf4707553e5fe..609c70a60d6463 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -5438,6 +5438,7 @@ const char* AMDGPUTargetLowering::getTargetNodeName(unsigned Opcode) const {
NODE_NAME_CASE(BUFFER_ATOMIC_FADD)
NODE_NAME_CASE(BUFFER_ATOMIC_FMIN)
NODE_NAME_CASE(BUFFER_ATOMIC_FMAX)
+ NODE_NAME_CASE(BUFFER_ATOMIC_COND_SUB_U32)
case AMDGPUISD::LAST_AMDGPU_ISD_NUMBER: break;
}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index 827fb106b55199..6828db6e0220d5 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -589,6 +589,7 @@ enum NodeType : unsigned {
BUFFER_ATOMIC_FADD,
BUFFER_ATOMIC_FMIN,
BUFFER_ATOMIC_FMAX,
+ BUFFER_ATOMIC_COND_SUB_U32,
LAST_AMDGPU_ISD_NUMBER
};
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
index eaf72d7157ee2d..e3c6f46c758e89 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructions.td
@@ -642,10 +642,14 @@ defm int_amdgcn_global_atomic_fmax : noret_op;
defm int_amdgcn_global_atomic_csub : noret_op;
defm int_amdgcn_flat_atomic_fadd : local_addr_space_atomic_op;
defm int_amdgcn_ds_fadd_v2bf16 : noret_op;
+defm int_amdgcn_flat_atomic_cond_sub_u32 : noret_op;
+defm int_amdgcn_global_atomic_cond_sub_u32 : noret_op;
defm int_amdgcn_flat_atomic_fmin_num : noret_op;
defm int_amdgcn_flat_atomic_fmax_num : noret_op;
defm int_amdgcn_global_atomic_fmin_num : noret_op;
defm int_amdgcn_global_atomic_fmax_num : noret_op;
+defm int_amdgcn_ds_cond_sub_u32 : noret_op;
+defm int_amdgcn_ds_cond_sub_u32 : local_addr_space_atomic_op;
multiclass noret_binary_atomic_op<SDNode atomic_op, bit IsInt = 1> {
let HasNoUse = true in
diff --git a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
index fbee2888945185..e605986564f2fc 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPULegalizerInfo.cpp
@@ -5882,6 +5882,9 @@ static unsigned getBufferAtomicPseudo(Intrinsic::ID IntrID) {
case Intrinsic::amdgcn_struct_buffer_atomic_fmax:
case Intrinsic::amdgcn_struct_ptr_buffer_atomic_fmax:
return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_FMAX;
+ case Intrinsic::amdgcn_raw_buffer_atomic_cond_sub_u32:
+ case Intrinsic::amdgcn_struct_buffer_atomic_cond_sub_u32:
+ return AMDGPU::G_AMDGPU_BUFFER_ATOMIC_COND_SUB_U32;
default:
llvm_unreachable("unhandled atomic opcode");
}
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
index c9412f720c62ec..51d2bb130774f9 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURegisterBankInfo.cpp
@@ -4690,6 +4690,9 @@ AMDGPURegisterBankInfo::getInstrMapping(const MachineInstr &MI) const {
case Intrinsic::amdgcn_flat_atomic_fmax_num:
case Intrinsic::amdgcn_global_atomic_fadd_v2bf16:
case Intrinsic::amdgcn_flat_atomic_fadd_v2bf16:
+ case Intrinsic::amdgcn_flat_atomic_cond_sub_u32:
+ case Intrinsic::amdgcn_global_atomic_cond_sub_u32:
+ case Intrinsic::amdgcn_ds_cond_sub_u32:
return getDefaultMappingAllVGPR(MI);
case Intrinsic::amdgcn_ds_ordered_add:
case Intrinsic::amdgcn_ds_ordered_swap:
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td b/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td
index beb670669581f1..1a28197fd90265 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPUSearchableTables.td
@@ -243,17 +243,20 @@ def : SourceOfDivergence<int_amdgcn_global_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_global_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_global_atomic_fmin_num>;
def : SourceOfDivergence<int_amdgcn_global_atomic_fmax_num>;
+def : SourceOfDivergence<int_amdgcn_global_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fadd>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fmin_num>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fmax_num>;
+def : SourceOfDivergence<int_amdgcn_flat_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_global_atomic_fadd_v2bf16>;
def : SourceOfDivergence<int_amdgcn_flat_atomic_fadd_v2bf16>;
def : SourceOfDivergence<int_amdgcn_ds_fadd>;
def : SourceOfDivergence<int_amdgcn_ds_fmin>;
def : SourceOfDivergence<int_amdgcn_ds_fmax>;
def : SourceOfDivergence<int_amdgcn_ds_fadd_v2bf16>;
+def : SourceOfDivergence<int_amdgcn_ds_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_buffer_atomic_swap>;
def : SourceOfDivergence<int_amdgcn_buffer_atomic_add>;
def : SourceOfDivergence<int_amdgcn_buffer_atomic_sub>;
@@ -281,6 +284,7 @@ def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_fadd>;
def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_cmpswap>;
+def : SourceOfDivergence<int_amdgcn_raw_buffer_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_swap>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_add>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_sub>;
@@ -297,6 +301,7 @@ def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_fadd>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_cmpswap>;
+def : SourceOfDivergence<int_amdgcn_raw_ptr_buffer_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_swap>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_add>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_sub>;
@@ -313,6 +318,7 @@ def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_fadd>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_cmpswap>;
+def : SourceOfDivergence<int_amdgcn_struct_buffer_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_swap>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_add>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_sub>;
@@ -329,6 +335,7 @@ def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_fadd>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_fmin>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_fmax>;
def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_cmpswap>;
+def : SourceOfDivergence<int_amdgcn_struct_ptr_buffer_atomic_cond_sub_u32>;
def : SourceOfDivergence<int_amdgcn_buffer_atomic_csub>;
def : SourceOfDivergence<int_amdgcn_ps_live>;
def : SourceOfDivergence<int_amdgcn_live_mask>;
diff --git a/llvm/lib/Target/AMDGPU/BUFInstructions.td b/llvm/lib/Target/AMDGPU/BUFInstructions.td
index 43d35fa5291ca0..fb4ef8620b795d 100644
--- a/llvm/lib/Target/AMDGPU/BUFInstructions.td
+++ b/llvm/lib/Target/AMDGPU/BUFInstructions.td
@@ -1245,6 +1245,12 @@ defm BUFFER_ATOMIC_PK_ADD_F16 : MUBUF_Pseudo_Atomics_RTN <
"buffer_atomic_pk_add_f16", VGPR_32, v2f16, null_frag
>;
+let SubtargetPredicate = isGFX12Plus in {
+defm BUFFER_ATOMIC_COND_SUB_U32 : MUBUF_Pseudo_Atomics <
+ "buffer_atomic_cond_sub_u32", VGPR_32, i32
+>;
+}
+
//===----------------------------------------------------------------------===//
// MTBUF Instructions
//===----------------------------------------------------------------------===//
@@ -1708,6 +1714,13 @@ defm : SIBufferAtomicPat<"SIbuffer_atomic_dec", i64, "BUFFER_ATOMIC_DEC_X2">;
let SubtargetPredicate = HasAtomicCSubNoRtnInsts in
defm : SIBufferAtomicPat<"SIbuffer_atomic_csub", i32, "BUFFER_ATOMIC_CSUB", ["noret"]>;
+let SubtargetPredicate = isGFX12Plus in {
+ defm : SIBufferAtomicPat_Common<"SIbuffer_atomic_cond_sub_u32", i32, "BUFFER_ATOMIC_COND_SUB_U32_VBUFFER", ["ret"]>;
+
+ let OtherPredicates = [HasAtomicCSubNoRtnInsts] in
+ defm : SIBufferAtomicPat_Common<"SIbuffer_atomic_cond_sub_u32", i32, "BUFFER_ATOMIC_COND_SUB_U32_VBUFFER", ["noret"]>;
+}
+
let SubtargetPredicate = isGFX6GFX7GFX10Plus in {
defm : SIBufferAtomicPat<"SIbuffer_atomic_fmin", f32, "BUFFER_ATOMIC_FMIN">;
defm : SIBufferAtomicPat<"SIbuffer_atomic_fmax", f32, "BUFFER_ATOMIC_FMAX">;
@@ -2610,6 +2623,7 @@ defm BUFFER_ATOMIC_AND_X2 : MUBUF_Real_Atomic_gfx11_gfx12_Renamed<0x049,
defm BUFFER_ATOMIC_CMPSWAP : MUBUF_Real_Atomic_gfx11_gfx12_Renamed<0x034, "buffer_atomic_cmpswap_b32">;
defm BUFFER_ATOMIC_CMPSWAP_X2 : MUBUF_Real_Atomic_gfx11_gfx12_Renamed<0x042, "buffer_atomic_cmpswap_b64">;
defm BUFFER_ATOMIC_FCMPSWAP : MUBUF_Real_Atomic_gfx11_Renamed<0x050, "buffer_atomic_cmpswap_f32">;
+defm BUFFER_ATOMIC_COND_SUB_U32 : MUBUF_Real_Atomic_gfx12<0x050>;
defm BUFFER_ATOMIC_CSUB : MUBUF_Real_Atomic_gfx11_gfx12_Renamed_gfx12_Renamed<0x037, "buffer_atomic_sub_clamp_u32", "buffer_atomic_csub_u32">;
def : Mnem_gfx11_gfx12<"buffer_atomic_csub", "buffer_atomic_csub_u32">;
defm BUFFER_ATOMIC_DEC : MUBUF_Real_Atomic_gfx11_gfx12_Renamed<0x040, "buffer_atomic_dec_u32">;
diff --git a/llvm/lib/Target/AMDGPU/DSInstructions.td b/llvm/lib/Target/AMDGPU/DSInstructions.td
index bc9049b4ef33c0..fb79d9264a3583 100644
--- a/llvm/lib/Target/AMDGPU/DSInstructions.td
+++ b/llvm/lib/Target/AMDGPU/DSInstructions.td
@@ -438,6 +438,12 @@ class DS_1A1D_PERMUTE <string opName, SDPatternOperator node = null_frag,
let has_gds = 0;
}
+class DSAtomicRetPat<DS_Pseudo inst, ValueType vt, PatFrag frag, int complexity = 0,
+ bit gds=0> : GCNPat <(frag (DS1Addr1Offset i32:$ptr, i32:$offset), vt:$value),
+ (inst $ptr, getVregSrcForVT<vt>.ret:$value, offset:$offset, (i1 gds))> {
+ let AddedComplexity = complexity;
+}
+
defm DS_ADD_U32 : DS_1A1D_NORET_mc<"ds_add_u32">;
defm DS_SUB_U32 : DS_1A1D_NORET_mc<"ds_sub_u32">;
defm DS_RSUB_U32 : DS_1A1D_NORET_mc<"ds_rsub_u32">;
@@ -733,9 +739,22 @@ def DS_BVH_STACK_RTN_B32 : DS_BVH_STACK<"ds_bvh_stack_rtn_b32">;
let SubtargetPredicate = isGFX12Plus in {
+defm DS_COND_SUB_U32 : DS_1A1D_NORET_mc<"ds_cond_sub_u32">;
+defm DS_COND_SUB_RTN_U32 : DS_1A1D_RET_mc<"ds_cond_sub_rtn_u32", VGPR_32, "ds_cond_sub_u32">;
defm DS_SUB_CLAMP_U32 : DS_1A1D_NORET_mc<"ds_sub_clamp_u32">;
defm DS_SUB_CLAMP_RTN_U32 : DS_1A1D_RET_mc<"ds_sub_clamp_rtn_u32", VGPR_32, "ds_sub_clamp_u32">;
+multiclass DSAtomicRetNoRetPatIntrinsic_mc<DS_Pseudo inst, DS_Pseudo noRetInst,
+ ValueType vt, string frag> {
+ def : DSAtomicRetPat<inst, vt,
+ !cast<PatFrag>(frag#"_local_addrspace")>;
+
+ let OtherPredicates = [HasAtomicCSubNoRtnInsts] in
+ def : DSAtomicRetPat<noRetInst, vt,
+ !cast<PatFrag>(frag#"_noret_local_addrspace"), /* complexity */ 1>;
+}
+
+defm : DSAtomicRetNoRetPatIntrinsic_mc<DS_COND_SUB_RTN_U32, DS_COND_SUB_U32, i32, "int_amdgcn_ds_cond_sub_u32">;
} // let SubtargetPredicate = isGFX12Plus
//===----------------------------------------------------------------------===//
@@ -955,12 +974,6 @@ defm : DSWritePat_mc <DS_WRITE_B128, vt, "store_align_less_than_4_local">;
} // End AddedComplexity = 100
-class DSAtomicRetPat<DS_Pseudo inst, ValueType vt, PatFrag frag, int complexity = 0,
- bit gds=0> : GCNPat <(frag (DS1Addr1Offset i32:$ptr, i32:$offset), vt:$value),
- (inst $ptr, getVregSrcForVT<vt>.ret:$value, offset:$offset, (i1 gds))> {
- let AddedComplexity = complexity;
-}
-
multiclass DSAtomicRetPat_mc<DS_Pseudo inst, ValueType vt, string frag> {
let OtherPredicates = [LDSRequiresM0Init] in {
def : DSAtomicRetPat<inst, vt, !cast<PatFrag>(frag#"_local_m0_"#vt.Size)>;
@@ -1238,7 +1251,9 @@ defm DS_MIN_NUM_F64 : DS_Real_Renamed_gfx12<0x052, DS_MIN_F64, "ds_min_num
defm DS_MAX_NUM_F64 : DS_Real_Renamed_gfx12<0x053, DS_MAX_F64, "ds_max_num_f64">;
defm DS_MIN_NUM_RTN_F64 : DS_Real_Renamed_gfx12<0x072, DS_MIN_RTN_F64, "ds_min_num_rtn_f64">;
defm DS_MAX_NUM_RTN_F64 : DS_Real_Renamed_gfx12<0x073, DS_MAX_RTN_F64, "ds_max_num_rtn_f64">;
+defm DS_COND_SUB_U32 : DS_Real_gfx12<0x098>;
defm DS_SUB_CLAMP_U32 : DS_Real_gfx12<0x099>;
+defm DS_COND_SUB_RTN_U32 : DS_Real_gfx12<0x0a8>;
defm DS_SUB_CLAMP_RTN_U32 : DS_Real_gfx12<0x0a9>;
//===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/AMDGPU/FLATInstructions.td b/llvm/lib/Target/AMDGPU/FLATInstructions.td
index 0dd2b3f5c2c912..4d5ebf11d7232f 100644
--- a/llvm/lib/Target/AMDGPU/FLATInstructions.td
+++ b/llvm/lib/Target/AMDGPU/FLATInstructions.td
@@ -801,6 +801,7 @@ let SubtargetPredicate = HasFlatAtomicFaddF32Inst in {
let SubtargetPredicate = isGFX12Plus in {
defm FLAT_ATOMIC_CSUB_U32 : FLAT_Atomic_Pseudo <"flat_atomic_csub_u32", VGPR_32, i32>;
+ defm FLAT_ATOMIC_COND_SUB_U32 : FLAT_Atomic_Pseudo <"flat_atomic_cond_sub_u32", VGPR_32, i32>;
} // End SubtargetPredicate = isGFX12Plus
defm GLOBAL_LOAD_UBYTE : FLAT_Global_Load_Pseudo <"global_load_ubyte", VGPR_32>;
@@ -926,6 +927,10 @@ defm GLOBAL_LOAD_LDS_USHORT : FLAT_Global_Load_LDS_Pseudo <"global_load_lds_usho
defm GLOBAL_LOAD_LDS_SSHORT : FLAT_Global_Load_LDS_Pseudo <"global_load_lds_sshort">;
defm GLOBAL_LOAD_LDS_DWORD : FLAT_Global_Load_LDS_Pseudo <"global_load_lds_dword">;
+let SubtargetPredicate = isGFX12Plus in {
+ defm GLOBAL_ATOMIC_COND_SUB_U32 : FLAT_Global_Atomic_Pseudo <"global_atomic_cond_sub_u32", VGPR_32, i32>;
+} // End SubtargetPredicate = isGFX12Plus
+
} // End is_flat_global = 1
@@ -1268,6 +1273,13 @@ defm : FlatAtomicPat <"FLAT_ATOMIC_CMPSWAP_X2", "AMDGPUatomic_cmp_swap_"#as, i64
defm : FlatAtomicPat <"FLAT_ATOMIC_XOR_X2", "atomic_load_xor_"#as, i64>;
} // end foreach as
+let SubtargetPredicate = isGFX12Plus in {
+ defm : FlatAtomicIntrRtnPat<"FLAT_ATOMIC_COND_SUB_U32", "int_amdgcn_flat_atomic_cond_sub_u32", i32>;
+
+ let OtherPredicates = [HasAtomicCSubNoRtnInsts] in
+ defm : FlatAtomicIntrNoRtnPat<"FLAT_ATOMIC_COND_SUB_U32", "int_amdgcn_flat_atomic_cond_sub_u32", i32>;
+}
+
def : FlatStorePat <FLAT_STORE_BYTE, truncstorei8_flat, i16>;
def : FlatStorePat <FLAT_STORE_SHORT, store_flat, i16>;
@@ -1377,6 +1389,16 @@ multiclass GlobalFLATAtomicIntrPats<string inst, string node, ValueType vt,
defm : GlobalFLATAtomicPats<inst, node, vt, data_vt, /* isIntr */ 1>;
}
+multiclass GlobalFLATAtomicIntrPatsNoRtn<string inst, string node, ValueType vt,
+ ValueType data_vt = vt> {
+ defm : GlobalFLATAtomicPatsNoRtn<inst, node, vt, data_vt, /* isIntr */ 1>;
+}
+
+multiclass GlobalFLATAtomicIntrPatsRtn<string inst, string node, ValueType vt,
+ ValueType data_vt = vt> {
+ defm : GlobalFLATAtomicPatsRtn<inst, node, vt, data_vt, /* isIntr */ 1>;
+}
+
multiclass ScratchFLATLoadPats<FLAT_Pseudo inst, SDPatternOperator node, ValueType vt> {
def : ScratchLoadSignedPat <inst, node, vt> {
let AddedComplexity = 25;
@@ -1529,6 +1551,13 @@ defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_SWAP_X2", "atomic_swap_global", i64>
defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_CMPSWAP_X2", "AMDGPUatomic_cmp_swap_global", i64, v2i64>;
defm : GlobalFLATAtomicPats <"GLOBAL_ATOMIC_XOR_X2", "atomic_load_xor_global", i64>;
+let SubtargetPredicate = isGFX12Plus in {
+ defm : GlobalFLATAtomicIntrPatsRtn <"GLOBAL_ATOMIC_COND_SUB_U32", "int_amdgcn_global_atomi...
[truncated]
|
arsenm
reviewed
Dec 22, 2023
Comment on lines
2502
to
2505
| def int_amdgcn_flat_atomic_cond_sub_u32 : AMDGPUAtomicRtn<llvm_i32_ty, flat_ptr_ty>; | ||
| def int_amdgcn_global_atomic_cond_sub_u32 : AMDGPUAtomicRtn<llvm_i32_ty, global_ptr_ty>; | ||
|
|
||
| def int_amdgcn_ds_cond_sub_u32 : AMDGPUAtomicRtn<llvm_i32_ty, local_ptr_ty>; |
There was a problem hiding this comment.
The other memory operations we just have one intrinsic mangled by the pointer type. We should also consider just adding it to atomicrmw
|
Adding support in atomicrmw. This will require to add new operation to aromicrmw "cond_sub" or you had something else in mind @arsenm ? |
|
ping |
Yes, and we have (Matt has) done this in the past, but it will require a wider consensus. I think it's fine to add AMDGPU intrinsics for this in the mean time. |
arsenm
reviewed
Jan 15, 2024
Done |
arsenm
approved these changes
Jan 17, 2024
arsenm
reviewed
Jan 17, 2024
arsenm
approved these changes
Jan 17, 2024
ampandey-1995
pushed a commit
to ampandey-1995/llvm-project
that referenced
this pull request
Jan 19, 2024
Co-authored-by: Vang Thao <Vang.Thao@amd.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.