[AMDGPU/MemOpsCluster] Compute `width` for `MIMG` instruction class.

Summary:
`width` computation is missing for newly added `MIMG`
instruction class. Add it.

Reviewers: foad, rampitec, arsenm

Reviewed By: foad

Subscribers: MatzeB, javed.absar, kzhuravl, jvesely, wdng, nhaehnle, yaxunl, dstuttard, tpr, t-tye, hiraditya, kerbowa, llvm-commits

Tags: #llvm

Differential Revision: https://reviews.llvm.org/D81649

diff --git a/llvm/lib/CodeGen/MachineScheduler.cpp b/llvm/lib/CodeGen/MachineScheduler.cpp
index 0f21c97..cf75d53 100644 (file)
--- a/llvm/lib/CodeGen/MachineScheduler.cpp
+++ b/llvm/lib/CodeGen/MachineScheduler.cpp
@@ -1573,8 +1573,13 @@ void BaseMemOpClusterMutation::clusterNeighboringMemOps(
     bool OffsetIsScalable;
     unsigned Width;
     if (TII->getMemOperandsWithOffsetWidth(MI, BaseOps, Offset,
-                                           OffsetIsScalable, Width, TRI))
+                                           OffsetIsScalable, Width, TRI)) {
       MemOpRecords.push_back(MemOpInfo(SU, BaseOps, Offset, Width));
+
+      LLVM_DEBUG(dbgs() << "Num BaseOps: " << BaseOps.size() << ", Offset: "
+                        << Offset << ", OffsetIsScalable: " << OffsetIsScalable
+                        << ", Width: " << Width << "\n");
+    }
 #ifndef NDEBUG
     for (auto *Op : BaseOps)
       assert(Op);
@@ -1630,6 +1635,10 @@ void BaseMemOpClusterMutation::clusterNeighboringMemOps(
                         << ")\n");
       DAG->addEdge(Succ.getSUnit(), SDep(SUb, SDep::Artificial));
     }
+
+    LLVM_DEBUG(dbgs() << "  Curr cluster length: " << ClusterLength
+                      << ", Curr cluster bytes: " << CurrentClusterBytes
+                      << "\n");
   }
 }
 

diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 0dcaecd..77bfd88 100644 (file)
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -396,6 +396,9 @@ bool SIInstrInfo::getMemOperandsWithOffsetWidth(
       BaseOps.push_back(getNamedOperand(LdSt, AMDGPU::OpName::vaddr));
     }
     Offset = 0;
+    // Get appropriate operand, and compute width accordingly.
+    DataOpIdx = AMDGPU::getNamedOperandIdx(Opc, AMDGPU::OpName::vdata);
+    Width = getOpSize(LdSt, DataOpIdx);
     return true;
   }
 

diff --git a/llvm/test/CodeGen/AMDGPU/cluster_stores.ll b/llvm/test/CodeGen/AMDGPU/cluster_stores.ll
index 67d6f68..bc3bcfe 100644 (file)
--- a/llvm/test/CodeGen/AMDGPU/cluster_stores.ll
+++ b/llvm/test/CodeGen/AMDGPU/cluster_stores.ll
@@ -5,6 +5,14 @@
 ; CHECK-LABEL: {{^}}cluster_load_cluster_store:
 define amdgpu_kernel void @cluster_load_cluster_store(i32* noalias %lb, i32* noalias %sb) {
 bb:
+
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 8
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 8
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 4
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 4
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 4
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 4
+
 ; DBG: Cluster ld/st SU(1) - SU(2)
 
 ; DBG: Cluster ld/st SU([[L1:[0-9]+]]) - SU([[L2:[0-9]+]])
@@ -45,6 +53,13 @@ bb:
 ; CHECK-LABEL: {{^}}cluster_load_valu_cluster_store:
 define amdgpu_kernel void @cluster_load_valu_cluster_store(i32* noalias %lb, i32* noalias %sb) {
 bb:
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 8
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 8
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 4
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 4
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 4
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 4
+
 ; DBG: Cluster ld/st SU(1) - SU(2)
 
 ; DBG: Cluster ld/st SU([[L1:[0-9]+]]) - SU([[L2:[0-9]+]])
@@ -86,6 +101,8 @@ bb:
 
 ; Cluster loads from the same texture with different coordinates
 ; CHECK-LABEL: {{^}}cluster_image_load:
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 16
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 16
 ; DBG: {{^}}Cluster ld/st [[SU1:SU\([0-9]+\)]] - [[SU2:SU\([0-9]+\)]]
 ; DBG: {{^}}[[SU1]]: {{.*}} IMAGE_LOAD
 ; DBG: {{^}}[[SU2]]: {{.*}} IMAGE_LOAD
@@ -106,6 +123,9 @@ entry:
 
 ; Don't cluster loads from different textures
 ; CHECK-LABEL: {{^}}no_cluster_image_load:
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 16
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 16
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 16
 ; DBG-NOT: {{^}}Cluster ld/st
 define amdgpu_ps void @no_cluster_image_load(<8 x i32> inreg %src1, <8 x i32> inreg %src2, <8 x i32> inreg %dst, i32 %x, i32 %y) {
 entry:
@@ -118,6 +138,8 @@ entry:
 
 ; Cluster loads from the same texture and sampler with different coordinates
 ; CHECK-LABEL: {{^}}cluster_image_sample:
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 16
+; DBG: Num BaseOps: {{[1-9]+}}, Offset: {{[0-9]+}}, OffsetIsScalable: {{[01]}}, Width: 16
 ; DBG: {{^}}Cluster ld/st [[SU1:SU\([0-9]+\)]] - [[SU2:SU\([0-9]+\)]]
 ; DBG: {{^}}[[SU1]]: {{.*}} IMAGE_SAMPLE
 ; DBG: {{^}}[[SU2]]: {{.*}} IMAGE_SAMPLE