[PATCH] D51592: AMDGPU: Fix DAG divergence not reporting flat loads

[Matt Arsenault via Phabricator via llvm-commits]

arsenm created this revision.
arsenm added a reviewer: scott.linder.
Herald added subscribers: t-tye, tpr, dstuttard, yaxunl, nhaehnle, wdng, jvesely, kzhuravl.

Match behavior in DAG of r340343


https://reviews.llvm.org/D51592

Files:
  lib/Target/AMDGPU/SIISelLowering.cpp
  test/CodeGen/AMDGPU/dag-divergence.ll


Index: test/CodeGen/AMDGPU/dag-divergence.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/dag-divergence.ll
@@ -0,0 +1,30 @@
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=fiji -verify-machineinstrs < %s | FileCheck -check-prefix=GCN %s
+
+; GCN-LABEL: {{^}}private_load_maybe_divergent:
+; GCN: buffer_load_dword
+; GCN-NOT: s_load s
+; GCN: flat_load_dword
+; GCN-NOT: s_load s
+define amdgpu_kernel void @private_load_maybe_divergent(i32 addrspace(4)* %k, i32* %flat) {
+  %load = load volatile i32, i32 addrspace(5)* undef, align 4
+  %gep = getelementptr inbounds i32, i32 addrspace(4)* %k, i32 %load
+  %maybe.not.uniform.load = load i32, i32 addrspace(4)* %gep, align 4
+  store i32 %maybe.not.uniform.load, i32 addrspace(1)* undef
+  ret void
+}
+
+; GCN-LABEL: {{^}}flat_load_maybe_divergent:
+; GCN: s_load_dwordx4
+; GCN-NOT: s_load
+; GCN: flat_load_dword
+; GCN-NOT: s_load
+; GCN: flat_load_dword
+; GCN-NOT: s_load
+; GCN: flat_store_dword
+define amdgpu_kernel void @flat_load_maybe_divergent(i32 addrspace(4)* %k, i32* %flat) {
+  %load = load i32, i32* %flat, align 4
+  %gep = getelementptr inbounds i32, i32 addrspace(4)* %k, i32 %load
+  %maybe.not.uniform.load = load i32, i32 addrspace(4)* %gep, align 4
+  store i32 %maybe.not.uniform.load, i32 addrspace(1)* undef
+  ret void
+}
Index: lib/Target/AMDGPU/SIISelLowering.cpp
===================================================================
--- lib/Target/AMDGPU/SIISelLowering.cpp
+++ lib/Target/AMDGPU/SIISelLowering.cpp
@@ -9270,10 +9270,10 @@
     }
     break;
     case ISD::LOAD: {
-      const LoadSDNode *L = dyn_cast<LoadSDNode>(N);
-      // FIXME: Also needs to handle flat.
-      if (L->getMemOperand()->getAddrSpace() == AMDGPUAS::PRIVATE_ADDRESS)
-        return true;
+      const LoadSDNode *L = cast<LoadSDNode>(N);
+      unsigned AS = L->getAddressSpace();
+      // A flat load may access private memory.
+      return AS == AMDGPUAS::PRIVATE_ADDRESS || AS == AMDGPUAS::FLAT_ADDRESS;
     } break;
     case ISD::CALLSEQ_END:
     return true;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D51592.163703.patch
Type: text/x-patch
Size: 2100 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20180903/7ddba6ec/attachment.bin>