[Polly] [NFC] [PPCGCodeGeneration] Handle failing of invariant loads gracefully.

[bollu]

Summary

To do this, we replicate what CodeGeneration does. We expose
markNodeUnreachable from CodeGeneration to PPCGCodeGeneration.

[grosser]

LGTM

polly/trunk/include/polly/CodeGen/CodeGeneration.h

 //===------ polly/CodeGeneration.h - The Polly code generator *- C++ -*-===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef POLLY_CODEGENERATION_H
 #define POLLY_CODEGENERATION_H
 
+#include "IRBuilder.h"
 #include "polly/Config/config.h"
 #include "polly/ScopPass.h"
 #include "isl/map.h"
 #include "isl/set.h"
 
 namespace polly {
 enum VectorizerChoice {
   VECTORIZER_NONE,
   VECTORIZER_STRIPMINE,
   VECTORIZER_POLLY,
 };
 extern VectorizerChoice PollyVectorizerChoice;
 
+/// Mark a basic block unreachable.
+///
+/// Marks the basic block @p Block unreachable by equipping it with an
+/// UnreachableInst.
+void markBlockUnreachable(BasicBlock &Block, PollyIRBuilder &Builder);
+
 struct CodeGenerationPass : public PassInfoMixin<CodeGenerationPass> {
   PreservedAnalyses run(Scop &S, ScopAnalysisManager &SAM,
                         ScopStandardAnalysisResults &AR, SPMUpdater &U);
 };
 } // namespace polly
 
 #endif // POLLY_CODEGENERATION_H

polly/trunk/lib/CodeGen/CodeGeneration.cpp

static cl::opt<bool> Verify("polly-codegen-verify",
                             cl::Hidden, cl::init(false), cl::ZeroOrMore,
                             cl::cat(PollyCategory));
 
 static cl::opt<bool>
     PerfMonitoring("polly-codegen-perf-monitoring",
                    cl::desc("Add run-time performance monitoring"), cl::Hidden,
                    cl::init(false), cl::ZeroOrMore, cl::cat(PollyCategory));
 
+namespace polly {
+/// Mark a basic block unreachable.
+///
+/// Marks the basic block @p Block unreachable by equipping it with an
+/// UnreachableInst.
+void markBlockUnreachable(BasicBlock &Block, PollyIRBuilder &Builder) {
+  auto *OrigTerminator = Block.getTerminator();
+  Builder.SetInsertPoint(OrigTerminator);
+  Builder.CreateUnreachable();
+  OrigTerminator->eraseFromParent();
+}
+
+} // namespace polly
+
 namespace {
 
 static void verifyGeneratedFunction(Scop &S, Function &F, IslAstInfo &AI) {
   if (!Verify || !verifyFunction(F, &errs()))
     return;
 
   DEBUG({
     errs() << "== ISL Codegen created an invalid function ==\n\n== The "
static void fixRegionInfo(Function &F, Region &ParentRegion, RegionInfo &RI) {
   for (BasicBlock &BB : F) {
     if (RI.getRegionFor(&BB))
       continue;
 
     RI.setRegionFor(&BB, &ParentRegion);
   }
 }
 
-/// Mark a basic block unreachable.
-///
-/// Marks the basic block @p Block unreachable by equipping it with an
-/// UnreachableInst.
-static void markBlockUnreachable(BasicBlock &Block, PollyIRBuilder &Builder) {
-  auto *OrigTerminator = Block.getTerminator();
-  Builder.SetInsertPoint(OrigTerminator);
-  Builder.CreateUnreachable();
-  OrigTerminator->eraseFromParent();
-}
-
 /// Remove all lifetime markers (llvm.lifetime.start, llvm.lifetime.end) from
 /// @R.
 ///
 /// CodeGeneration does not copy lifetime markers into the optimized SCoP,
 /// which would leave the them only in the original path. This can transform
 /// code such as
 ///
 ///     llvm.lifetime.start(%p)

polly/trunk/lib/CodeGen/PPCGCodeGeneration.cpp

 //===------ PPCGCodeGeneration.cpp - Polly Accelerator Code Generation. ---===//
 //
 //                     The LLVM Compiler Infrastructure
 //
 // This file is distributed under the University of Illinois Open Source
 // License. See LICENSE.TXT for details.
 //
 //===----------------------------------------------------------------------===//
 //
 // Take a scop created by ScopInfo and map it to GPU code using the ppcg
 // GPU mapping strategy.
 //
 //===----------------------------------------------------------------------===//
 
 #include "polly/CodeGen/PPCGCodeGeneration.h"
+#include "polly/CodeGen/CodeGeneration.h"
 #include "polly/CodeGen/IslAst.h"
 #include "polly/CodeGen/IslNodeBuilder.h"
 #include "polly/CodeGen/Utils.h"
 #include "polly/DependenceInfo.h"
 #include "polly/LinkAllPasses.h"
 #include "polly/Options.h"
 #include "polly/ScopDetection.h"
 #include "polly/ScopInfo.h"
void generateCode(__isl_take isl_ast_node *Root, gpu_prog *Prog) {
     assert(CondBr && "CondBr not initialized by executeScopConditionally");
 
     GPUNodeBuilder NodeBuilder(Builder, Annotator, *DL, *LI, *SE, *DT, *S,
                                StartBlock, Prog, Runtime, Architecture);
 
     // TODO: Handle LICM
     auto SplitBlock = StartBlock->getSinglePredecessor();
     Builder.SetInsertPoint(SplitBlock->getTerminator());
-    NodeBuilder.addParameters(S->getContext().release());
 
     isl_ast_build *Build = isl_ast_build_alloc(S->getIslCtx());
     isl_ast_expr *Condition = IslAst::buildRunCondition(*S, Build);
     isl_ast_expr *SufficientCompute = createSufficientComputeCheck(*S, Build);
     Condition = isl_ast_expr_and(Condition, SufficientCompute);
     isl_ast_build_free(Build);
 
     // preload invariant loads. Note: This should happen before the RTC
     // because the RTC may depend on values that are invariant load hoisted.
-    if (!NodeBuilder.preloadInvariantLoads())
-      report_fatal_error("preloading invariant loads failed in function: " +
+    if (!NodeBuilder.preloadInvariantLoads()) {
+      DEBUG(dbgs() << "preloading invariant loads failed in function: " +
                           S->getFunction().getName() +
                           " | Scop Region: " + S->getNameStr());
+      // adjust the dominator tree accordingly.
+      auto *ExitingBlock = StartBlock->getUniqueSuccessor();
+      assert(ExitingBlock);
+      auto *MergeBlock = ExitingBlock->getUniqueSuccessor();
+      assert(MergeBlock);
+      polly::markBlockUnreachable(*StartBlock, Builder);
+      polly::markBlockUnreachable(*ExitingBlock, Builder);
+      auto *ExitingBB = S->getExitingBlock();
+      assert(ExitingBB);
+
+      DT->changeImmediateDominator(MergeBlock, ExitingBB);
+      DT->eraseNode(ExitingBlock);
+      isl_ast_expr_free(Condition);
+      isl_ast_node_free(Root);
+    } else {
 
+      NodeBuilder.addParameters(S->getContext().release());
       Value *RTC = NodeBuilder.createRTC(Condition);
       Builder.GetInsertBlock()->getTerminator()->setOperand(0, RTC);
 
       Builder.SetInsertPoint(&*StartBlock->begin());
 
       NodeBuilder.create(Root);
+    }
 
     /// In case a sequential kernel has more surrounding loops as any parallel
     /// kernel, the SCoP is probably mostly sequential. Hence, there is no
     /// point in running it on a GPU.
     if (NodeBuilder.DeepestSequential > NodeBuilder.DeepestParallel)
       CondBr->setOperand(0, Builder.getFalse());
 
     if (!NodeBuilder.BuildSuccessful)

polly/trunk/test/GPGPU/failing-invariant-load-handling.ll

+; RUN: opt %loadPolly < %s -analyze -polly-scops -polly-process-unprofitable -polly-invariant-load-hoisting | FileCheck %s -check-prefix=SCOPS
+; RUN: opt %loadPolly -S < %s -polly-codegen-ppcg -polly-process-unprofitable -polly-invariant-load-hoisting | FileCheck %s -check-prefix=CODEGEN
+
+target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n8:16:32-S64"
+
+%S = type { i32, i32, [12 x %L] }
+%L = type { i32, i32, double, i32, i32, i32, i32, i32 }
+
+define void @test(%S* %cpi, i1 %b) {
+; SCOPS-LABEL: Region: %if.then14---%exit
+; SCOPS:         Invariant Accesses: {
+; SCOPS-NEXT:            ReadAccess :=       [Reduction Type: NONE] [Scalar: 0]
+; SCOPS-NEXT:                [l2, l1] -> { Stmt_for_body_i[i0] -> MemRef_cpi[0, 0] };
+; SCOPS-NEXT:            Execution Context: [l2, l1] -> {  :  }
+; SCOPS-NEXT:            ReadAccess :=       [Reduction Type: NONE] [Scalar: 0]
+; SCOPS-NEXT:                [l2, l1] -> { Stmt_for_body_lr_ph_i[] -> MemRef_cpi[0, 1] };
+; SCOPS-NEXT:            Execution Context: [l2, l1] -> {  : l2 > 0 }
+; SCOPS-NEXT:    }
+; SCOPS:         Arrays {
+; SCOPS-NEXT:        i32 MemRef_cpi[*][(10 * %l1)]; // Element size 4
+; SCOPS-NEXT:    }
+
+; Check that we gracefully handle failing invariant loads.
+; This test case is taken from:
+; test/Isl/CodeGen/invariant-load-dimension.ll
+
+; FIXME: Figure out how to actually generate code for this loop.
+; CODEGEN-NOT: LLVM ERROR: preloading invariant loads failed in function
+
+entry:
+  %nt = getelementptr inbounds %S, %S* %cpi, i32 0, i32 1
+  br i1 %b, label %if.then14, label %exit
+
+if.then14:
+  %ns = getelementptr inbounds %S, %S* %cpi, i32 0, i32 0
+  %l0 = load i32, i32* %ns, align 8
+  %cmp12.i = icmp sgt i32 %l0, 0
+  br i1 %cmp12.i, label %for.body.lr.ph.i, label %exit
+
+for.body.lr.ph.i:
+  %l1 = load i32, i32* %nt, align 4
+  br label %for.body.i
+
+for.body.i:
+  %phi = phi i32 [ 0, %for.body.lr.ph.i ], [ %inc, %for.body.i ]
+  %mul.i163 = mul nsw i32 %phi, %l1
+  %cv = getelementptr inbounds %S, %S* %cpi, i32 0, i32 2, i32 %mul.i163, i32 0
+  store i32 0, i32* %cv, align 8
+  %inc = add nuw nsw i32 %phi, 1
+  %l2 = load i32, i32* %ns, align 8
+  %cmp.i164 = icmp slt i32 %inc, %l2
+  br i1 %cmp.i164, label %for.body.i, label %exit
+
+exit:
+  ret void
+}