From 7a3f751cbc0d9be18d92ec1b6917ced4f0cd84b8 Mon Sep 17 00:00:00 2001
From: Stanislav Mekhanoshin <Stanislav.Mekhanoshin@amd.com>
Date: Mon, 11 Jun 2018 16:50:49 +0000
Subject: [PATCH] [AMDGPU] Do not consider indirect acces through phi for wave
 limiter

Rational: if there is indirect access that is usually an issue
because load is not ready by the use. However, if use is inside a
loop and load is outside that is potentially an issue for a first
iteration only.

Differential Revision: https://reviews.llvm.org/D47740

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@334420 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/AMDGPU/AMDGPUPerfHintAnalysis.cpp |  6 ------
 test/CodeGen/AMDGPU/perfhint.ll              | 26 ++++++++++++++++++++++++++
 2 files changed, 26 insertions(+), 6 deletions(-)

diff --git a/lib/Target/AMDGPU/AMDGPUPerfHintAnalysis.cpp b/lib/Target/AMDGPU/AMDGPUPerfHintAnalysis.cpp
index d307f818d7b..3cfdccc9fe5 100644
--- a/lib/Target/AMDGPU/AMDGPUPerfHintAnalysis.cpp
+++ b/lib/Target/AMDGPU/AMDGPUPerfHintAnalysis.cpp
@@ -198,12 +198,6 @@ bool AMDGPUPerfHint::isIndirectAccess(const Instruction *Inst) const {
       continue;
     }
 
-    if (auto Phi = dyn_cast<PHINode>(V)) {
-      for (unsigned I = 0, E = Phi->getNumIncomingValues(); I != E; ++I)
-        WorkSet.insert(Phi->getIncomingValue(I));
-      continue;
-    }
-
     LLVM_DEBUG(dbgs() << "    dropped\n");
   }
 
diff --git a/test/CodeGen/AMDGPU/perfhint.ll b/test/CodeGen/AMDGPU/perfhint.ll
index 06dc7ad4898..a8990be7f19 100644
--- a/test/CodeGen/AMDGPU/perfhint.ll
+++ b/test/CodeGen/AMDGPU/perfhint.ll
@@ -82,4 +82,30 @@ bb:
   ret void
 }
 
+; GCN-LABEL: {{^}}test_indirect_through_phi:
+; MemoryBound: 0
+; WaveLimiterHint : 0
+define amdgpu_kernel void @test_indirect_through_phi(float addrspace(1)* %arg) {
+bb:
+  %load = load float, float addrspace(1)* %arg, align 8
+  %load.f = bitcast float %load to i32
+  %n = tail call i32 @llvm.amdgcn.workitem.id.x()
+  br label %bb1
+
+bb1:                                              ; preds = %bb1, %bb
+  %phi = phi i32 [ %load.f, %bb ], [ %and2, %bb1 ]
+  %ind = phi i32 [ 0, %bb ], [ %inc2, %bb1 ]
+  %and1 = and i32 %phi, %n
+  %gep = getelementptr inbounds float, float addrspace(1)* %arg, i32 %and1
+  store float %load, float addrspace(1)* %gep, align 4
+  %inc1 = add nsw i32 %phi, 1310720
+  %and2 = and i32 %inc1, %n
+  %inc2 = add nuw nsw i32 %ind, 1
+  %cmp = icmp eq i32 %inc2, 1024
+  br i1 %cmp, label %bb2, label %bb1
+
+bb2:                                              ; preds = %bb1
+  ret void
+}
+
 declare i32 @llvm.amdgcn.workitem.id.x()
-- 
2.11.0