R600: Workaround for cayman loop bug

llvm-svn: 196121

Author: Vincent Lejeune

llvm/lib/Target/R600/R600ControlFlowFinalizer.cpp

@@ -332,6 +332,7 @@ class R600ControlFlowFinalizer : public MachineFunctionPass {
 
     unsigned MaxStack = 0;
     unsigned CurrentStack = 0;
+    unsigned CurrentLoopDepth = 0;
     bool HasPush = false;
     for (MachineFunction::iterator MB = MF.begin(), ME = MF.end(); MB != ME;
         ++MB) {
@@ -370,6 +371,13 @@ class R600ControlFlowFinalizer : public MachineFunctionPass {
           CurrentStack++;
           MaxStack = std::max(MaxStack, CurrentStack);
           HasPush = true;
+          if (ST.hasCaymanISA() && CurrentLoopDepth > 1) {
+            BuildMI(MBB, MI, MBB.findDebugLoc(MI), TII->get(AMDGPU::CF_PUSH_CM))
+                .addImm(CfCount + 1)
+                .addImm(1);
+            MI->setDesc(TII->get(AMDGPU::CF_ALU));
+            CfCount++;
+          }
         case AMDGPU::CF_ALU:
           I = MI;
           AluClauses.push_back(MakeALUClause(MBB, I));
@@ -378,6 +386,7 @@ class R600ControlFlowFinalizer : public MachineFunctionPass {
           break;
         case AMDGPU::WHILELOOP: {
           CurrentStack+=4;
+          CurrentLoopDepth++;
           MaxStack = std::max(MaxStack, CurrentStack);
           MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
               getHWInstrDesc(CF_WHILE_LOOP))
@@ -392,6 +401,7 @@ class R600ControlFlowFinalizer : public MachineFunctionPass {
         }
         case AMDGPU::ENDLOOP: {
           CurrentStack-=4;
+          CurrentLoopDepth--;
           std::pair<unsigned, std::set<MachineInstr *> > Pair =
               LoopStack.back();
           LoopStack.pop_back();

llvm/lib/Target/R600/R600Instructions.td

@@ -1867,6 +1867,10 @@ def : Pat <
     let COUNT = 0;
   }
 
+  def CF_PUSH_CM : CF_CLAUSE_EG<11, (ins i32imm:$ADDR, i32imm:$POP_COUNT), "PUSH @$ADDR POP:$POP_COUNT"> {
+    let COUNT = 0;
+  }
+
 def : Pat<(fsqrt f32:$src), (MUL R600_Reg32:$src, (RECIPSQRT_CLAMPED_cm $src))>;
 
 class RAT_STORE_DWORD <RegisterClass rc, ValueType vt, bits<4> mask> :

llvm/test/CodeGen/R600/cayman-loop-bug.ll

@@ -0,0 +1,32 @@
+; RUN: llc < %s -march=r600 -mcpu=cayman | FileCheck %s
+
+; CHECK-LABEL: @main
+; CHECK: LOOP_START_DX10
+; CHECK: ALU_PUSH_BEFORE
+; CHECK: LOOP_START_DX10
+; CHECK: PUSH
+; CHECK-NOT: ALU_PUSH_BEFORE
+; CHECK: END_LOOP
+; CHECK: END_LOOP
+define void @main (<4 x float> inreg %reg0) #0 {
+entry:
+  br label %outer_loop
+outer_loop:
+  %cnt = phi i32 [0, %entry], [%cnt_incr, %inner_loop]
+  %cond = icmp eq i32 %cnt, 16
+  br i1 %cond, label %outer_loop_body, label %exit
+outer_loop_body:
+  %cnt_incr = add i32 %cnt, 1
+  br label %inner_loop
+inner_loop:
+  %cnt2 = phi i32 [0, %outer_loop_body], [%cnt2_incr, %inner_loop_body]
+  %cond2 = icmp eq i32 %cnt2, 16
+  br i1 %cond, label %inner_loop_body, label %outer_loop
+inner_loop_body:
+  %cnt2_incr = add i32 %cnt2, 1
+  br label %inner_loop
+exit:
+  ret void
+}
+
+attributes #0 = { "ShaderType"="0" }