[EraVM] Only generate GEPs that dominate latch BB in EraVMIndexedMemO…

…psPrepare

In case we generate GEP that doesn't dominate latch BB,
we will run into an issue and verifier will complain:
```
Instruction does not dominate all uses!
  %3 = getelementptr inbounds i256, ptr addrspace(1) %0, i256 1
  %0 = phi ptr addrspace(1) [ %arg3, %bb1 ], [ %3, %bb5 ]
```
This patch fixes this issue.

Signed-off-by: Vladimir Radosavljevic <vr@matterlabs.dev>

llvm/lib/Target/EraVM/EraVMIndexedMemOpsPrepare.cpp

@@ -71,6 +71,7 @@
 #include "llvm/Analysis/LoopPass.h"
 #include "llvm/Analysis/ScalarEvolution.h"
 #include "llvm/Analysis/ScalarEvolutionExpressions.h"
+#include "llvm/IR/Dominators.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Support/CommandLine.h"
@@ -85,7 +86,7 @@ using namespace llvm;
 namespace {
 
 class EraVMIndexedMemOpsPrepare : public LoopPass {
-
+  DominatorTree *DT = nullptr;
   ScalarEvolution *SE = nullptr;
   LLVMContext *Ctx = nullptr;
   Loop *CurrentLoop = nullptr;
@@ -102,8 +103,10 @@ class EraVMIndexedMemOpsPrepare : public LoopPass {
   }
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.addRequired<DominatorTreeWrapperPass>();
     AU.addRequired<ScalarEvolutionWrapperPass>();
     AU.addRequired<LoopInfoWrapperPass>();
+    AU.addPreserved<DominatorTreeWrapperPass>();
     AU.addPreserved<LoopInfoWrapperPass>();
     AU.setPreservesCFG();
   }
@@ -208,6 +211,12 @@ bool EraVMIndexedMemOpsPrepare::isValidGEPAndIncByOneCell(
   if (!PHI)
     return false;
 
+  // The PHI node must be in the loop header, and BasePtr BB must dominate
+  // the latch BB.
+  if (CurrentLoop->getHeader() != PHI->getParent() ||
+      !DT->dominates(BasePtr->getParent(), CurrentLoop->getLoopLatch()))
+    return false;
+
   // In rewriteToFavorIndexedMemOps, new PHI is created with incoming values
   // from preheader and latch. Check that this PHI has the same incoming BBs.
   return PHI->getNumIncomingValues() == 2 &&
@@ -225,6 +234,7 @@ bool EraVMIndexedMemOpsPrepare::runOnLoop(Loop *L, LPPassManager &) {
     return false;
 
   auto &LI = getAnalysis<LoopInfoWrapperPass>().getLoopInfo();
+  DT = &getAnalysis<DominatorTreeWrapperPass>().getDomTree();
   SE = &getAnalysis<ScalarEvolutionWrapperPass>().getSE();
   Ctx = &L->getLoopPreheader()->getContext();
   CurrentLoop = L;
@@ -268,6 +278,7 @@ char EraVMIndexedMemOpsPrepare::ID = 0;
 
 INITIALIZE_PASS_BEGIN(EraVMIndexedMemOpsPrepare, DEBUG_TYPE,
                       ERAVM_PREPARE_INDEXED_MEMOPS_NAME, false, false)
+INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
 INITIALIZE_PASS_DEPENDENCY(ScalarEvolutionWrapperPass)
 INITIALIZE_PASS_DEPENDENCY(LoopInfoWrapperPass)
 INITIALIZE_PASS_END(EraVMIndexedMemOpsPrepare, DEBUG_TYPE,

llvm/test/CodeGen/EraVM/indexed-memops-gep-dominate-bug.ll

@@ -0,0 +1,43 @@
+; RUN: llc -O3 -stop-after=verify -compile-twice=false < %s | FileCheck %s
+
+target datalayout = "E-p:256:256-i256:256:256-S32-a:256:256"
+target triple = "eravm"
+
+define i256 @test(i256 %arg, ptr addrspace(1) %arg1, ptr addrspace(1) %arg2) {
+; CHECK-LABEL: @test(
+entry:
+  %icmp1 = icmp eq i256 %arg, 0
+  br i1 %icmp1, label %bb7, label %bb1
+
+bb1:
+  br label %bb2
+
+bb2:
+  %phi1 = phi i256 [ %add, %bb5 ], [ 0, %bb1 ]
+  %icmp2 = icmp ugt i256 %phi1, 100
+  br i1 %icmp2, label %bb3, label %bb4
+
+bb3:
+  ; CHECK: getelementptr inbounds i256, ptr addrspace(1) %arg1, i256 %phi1
+  %gep1 = getelementptr inbounds i256, ptr addrspace(1) %arg1, i256 %phi1
+  store i256 5, ptr addrspace(1) %gep1, align 32
+  br label %bb5
+
+bb4:
+  ; CHECK: getelementptr inbounds i256, ptr addrspace(1) %arg2, i256 %phi1
+  %gep2 = getelementptr inbounds i256, ptr addrspace(1) %arg2, i256 %phi1
+  store i256 10, ptr addrspace(1) %gep2, align 32
+  br label %bb5
+
+bb5:
+  %add = add i256 %phi1, 1
+  %cmp3 = icmp eq i256 %add, 0
+  br i1 %cmp3, label %bb6, label %bb2
+
+bb6:
+  br label %bb7
+
+bb7:
+  %phi2 = phi i256 [ 0, %entry ], [ %add, %bb6 ]
+  ret i256 %phi2
+}