AMDGPU/SI: Fix visit order assumption in SIFixSGPRCopies

[• tstellarAMD]

Summary

This pass was assuming that when a PHI instruction defined a register
used by another PHI instruction that the defining insstruction would
be legalized before the using instruction.

This assumption was causing the pass to not legalize some PHI nodes
within divergent flow-control.

This fixes a bug that was uncovered by r285762.

[nhaehnle]

LGTM, one minor comment.

test/CodeGen/AMDGPU/salu-to-valu.ll

460

GCN-LABEL

[arsenm]

LGTM with nits

lib/Target/AMDGPU/SIFixSGPRCopies.cpp

242

Spaces around +=

253

Ditto

262

Space before (

263

break on new line

lib/Target/AMDGPU/SIFixSGPRCopies.cpp

for (unsigned I = 1, N = MI.getNumOperands(); I != N; I += 2) {
 
     MI.getOperand(I).setReg(TmpReg);
   }
 
   CopyUse.eraseFromParent();
   return true;
 }
 
+static bool phiHasVGPROperands(const MachineInstr &PHI,
+                               const MachineRegisterInfo &MRI,
+                               const SIRegisterInfo *TRI,
+                               const SIInstrInfo *TII) {
+
+  for (unsigned i = 1; i < PHI.getNumOperands(); i+=2) {
+    unsigned Reg = PHI.getOperand(i).getReg();
+    if (TRI->hasVGPRs(MRI.getRegClass(Reg)))
+      return true;
+  }
+  return false;
+}
+static bool phiHasBreakDef(const MachineInstr &PHI,
+                           const MachineRegisterInfo &MRI,
+                           SmallSet<unsigned, 8> &Visited) {
+
+  for (unsigned i = 1; i < PHI.getNumOperands(); i+=2) {
+    unsigned Reg = PHI.getOperand(i).getReg();
+    if (Visited.count(Reg))
+      continue;
+
+    Visited.insert(Reg);
+
+    MachineInstr *DefInstr = MRI.getUniqueVRegDef(Reg);
+    assert(DefInstr);
+    switch(DefInstr->getOpcode()) {
+    default: break;
+    case AMDGPU::SI_BREAK:
+    case AMDGPU::SI_IF_BREAK:
+    case AMDGPU::SI_ELSE_BREAK:
+      return true;
+    case AMDGPU::PHI:
+      if (phiHasBreakDef(*DefInstr, MRI, Visited))
+        return true;
+    }
+  }
+  return false;
+}
+
 bool SIFixSGPRCopies::runOnMachineFunction(MachineFunction &MF) {
   const SISubtarget &ST = MF.getSubtarget<SISubtarget>();
   MachineRegisterInfo &MRI = MF.getRegInfo();
   const SIRegisterInfo *TRI = ST.getRegisterInfo();
   const SIInstrInfo *TII = ST.getInstrInfo();
 
   SmallVector<MachineInstr *, 16> Worklist;
 
for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
 
         // The one exception to this rule is when one of the operands
         // is defined by a SI_BREAK, SI_IF_BREAK, or SI_ELSE_BREAK
         // instruction.  In this case, there we know the program will
         // never enter the second block (the loop) without entering
         // the first block (where the condition is computed), so there
         // is no chance for values to be over-written.
 
-        bool HasBreakDef = false;
-        for (unsigned i = 1; i < MI.getNumOperands(); i+=2) {
-          unsigned Reg = MI.getOperand(i).getReg();
-          if (TRI->hasVGPRs(MRI.getRegClass(Reg))) {
-            TII->moveToVALU(MI);
-            break;
-          }
-          MachineInstr *DefInstr = MRI.getUniqueVRegDef(Reg);
-          assert(DefInstr);
-          switch(DefInstr->getOpcode()) {
-
-          case AMDGPU::SI_BREAK:
-          case AMDGPU::SI_IF_BREAK:
-          case AMDGPU::SI_ELSE_BREAK:
-          // If we see a PHI instruction that defines an SGPR, then that PHI
-          // instruction has already been considered and should have
-          // a *_BREAK as an operand.
-          case AMDGPU::PHI:
-            HasBreakDef = true;
-            break;
-          }
+        SmallSet<unsigned, 8> Visited;
+        if (phiHasVGPROperands(MI, MRI, TRI, TII) ||
+            (!SGPRBranch && !phiHasBreakDef(MI, MRI, Visited))) {
         }
-
-        if (!SGPRBranch && !HasBreakDef)
-          TII->moveToVALU(MI);
         break;
       }
       case AMDGPU::REG_SEQUENCE: {
         if (TRI->hasVGPRs(TII->getOpRegClass(MI, 0)) ||
             !hasVGPROperands(MI, TRI)) {
           foldVGPRCopyIntoRegSequence(MI, TRI, TII, MRI);
           continue;
         }

test/CodeGen/AMDGPU/salu-to-valu.ll

 bb6:
   store i32 1, i32 addrspace(1)* %out
   br label %bb7
 
 bb7:                                              ; preds = %bb3
   ret void
 }
 
+; GCN: {{^}}phi_visit_order:
+; GCN: v_add_i32_e32 v{{[0-9]+}}, vcc, 1, v{{[0-9]+}}
+define void @phi_visit_order() {
+bb:
+  br label %bb1
+
+bb1:
+  %tmp = phi i32 [ 0, %bb ], [ %tmp5, %bb4 ]
+  %tid = call i32 @llvm.amdgcn.workitem.id.x()
+  %cnd = icmp eq i32 %tid, 0
+  br i1 %cnd, label %bb4, label %bb2
+
+bb2:
+  %tmp3 = add nsw i32 %tmp, 1
+  br label %bb4
+
+bb4:
+  %tmp5 = phi i32 [ %tmp3, %bb2 ], [ %tmp, %bb1 ]
+  br label %bb1
+}
+
 attributes #0 = { nounwind readnone }
 attributes #1 = { nounwind }

test/CodeGen/MIR/AMDGPU/si-fix-sgpr-copies.mir

+# RUN: llc -march=amdgcn -run-pass si-fix-sgpr-copies %s -o - | FileCheck %s -check-prefixes=GCN
+
+--- |
+  define void @phi_visit_order() { ret void }
+
+name: phi_visit_order
+tracksRegLiveness: true
+registers:
+  - { id: 0, class: sreg_32 }
+  - { id: 1, class: sreg_64 }
+  - { id: 2, class: sreg_32 }
+  - { id: 7, class: vgpr_32 }
+  - { id: 8, class: sreg_32 }
+  - { id: 9, class: vgpr_32 }
+  - { id: 10, class: sreg_64 }
+  - { id: 11, class: sreg_32 }
+
+body: |
+  ; GCN-LABEL: name: phi_visit_order
+  ; GCN: V_ADD_I32
+  bb.0:
+    liveins: %vgpr0
+    successors: %bb.1
+    %7 = COPY %vgpr0
+    %8 = S_MOV_B32 0
+
+  bb.1:
+    successors: %bb.1, %bb.2
+    %0 = PHI %8, %bb.0, %0, %bb.1, %2, %bb.2
+    %9 = V_MOV_B32_e32 9, implicit %exec
+    %10 = V_CMP_EQ_U32_e64 %7, %9, implicit %exec
+    %1 = SI_IF %10, %bb.2, implicit-def %exec, implicit-def %scc, implicit %exec
+    S_BRANCH %bb.1
+
+  bb.2:
+    successors: %bb.1
+    SI_END_CF %1, implicit-def %exec, implicit-def %scc, implicit %exec
+    %11 = S_MOV_B32 1
+    %2 = S_ADD_I32 %0, %11, implicit-def %scc
+    S_BRANCH %bb.1
+
+...
+---