[Polly] [ManagedMemoryRewrite] Rewrite malloc, free correctly so it works inside bitcasts as well.

[bollu]

Summary

Reuse the machinery built for replacing global arrays to replace malloc/free as well. Example replacement that was missed earlier:

ll
call void bitcast (void (i8*)* @free to void (%struct.fson_string*)*)
    (%struct.fson_string* %13) #2

[bollu]

lib/CodeGen/ManagedMemoryRewrite.cpp

324

Suggest a better name please :)

[grosser]

Hi Siddharth,

some first comment. Also, do you have a test case here?

[bollu]

• [NFC] add testcase

[bollu]

@grosser Dont, added tests.

[grosser]

lib/CodeGen/ManagedMemoryRewrite.cpp

324

I thought I sent this out, but I apparently did not. "Really" does not add any information. Can you instead explain what the difference is here.

Maybe name it:

replaceAllUsesAndConstantUses

And describe that constants are normally ignored by replaceAllUses, and how this function handles them differently.

[bollu]

• [NFC] rename function as what Tobias suggested and add comment explaining why

[grosser]

LGTM.

lib/CodeGen/ManagedMemoryRewrite.cpp

320

Newline in between here.

322

actually

330

Two whitespaces?

lib/CodeGen/ManagedMemoryRewrite.cpp

if (!Return)
       continue;
     Builder.SetInsertPoint(Return);
 
     Value *FreeManagedFn = getOrCreatePollyFreeManaged(*M);
     Builder.CreateCall(FreeManagedFn, {RawManagedMem});
   }
 }
 
+// Replace all uses of `Old` with `New`, even inside `ConstantExpr`.
+// `replaceAllUsesWith` does replace values in `ConstantExpr`. This function
+// actuall does replace it in `ConstantExpr`. The caveat is that if there is
+// a use that is *outside* a function (say, at global declarations), we fail.
+// So, this is meant to be used on values which we know will only be used
+// within functions.
+//
+// This process works by looking through the uses of `Old`. If it finds a
+// `ConstantExpr`, it recursively looks for the owning instruction.
+//  Then, it expands all the `ConstantExpr` to instructions and replaces
+//  `Old` with `New` in the expanded instructions.
+static void replaceAllUsesAndConstantUses(Value *Old, Value *New,
+                                          PollyIRBuilder &Builder) {
+  SmallVector<Instruction *, 4> UserInstructions;
+  // Get all instructions that use array. We need to do this weird thing
+  // because `Constant`s that contain this array neeed to be expanded into
+  // instructions so that we can replace their parameters. `Constant`s cannot
+  // be edited easily, so we choose to convert all `Constant`s to
+  // `Instruction`s and handle all of the uses of `Array` uniformly.
+  for (Use &ArrayUse : Old->uses())
+    getInstructionUsersOfValue(ArrayUse.getUser(), UserInstructions);
+
+  for (Instruction *I : UserInstructions)
+    rewriteOldValToNew(I, Old, New, Builder);
+}
+
 class ManagedMemoryRewritePass : public ModulePass {
 public:
   static char ID;
   GPUArch Architecture;
   GPURuntime Runtime;
 
   ManagedMemoryRewritePass() : ModulePass(ID) {}
   virtual bool runOnModule(Module &M) {
     const DataLayout &DL = M.getDataLayout();
 
     Function *Malloc = M.getFunction("malloc");
 
     if (Malloc) {
+      PollyIRBuilder Builder(M.getContext());
       Function *PollyMallocManaged = getOrCreatePollyMallocManaged(M);
       assert(PollyMallocManaged && "unable to create polly_mallocManaged");
-      Malloc->replaceAllUsesWith(PollyMallocManaged);
+
+      replaceAllUsesAndConstantUses(Malloc, PollyMallocManaged, Builder);
       Malloc->eraseFromParent();
     }
 
     Function *Free = M.getFunction("free");
 
     if (Free) {
+      PollyIRBuilder Builder(M.getContext());
       Function *PollyFreeManaged = getOrCreatePollyFreeManaged(M);
       assert(PollyFreeManaged && "unable to create polly_freeManaged");
-      Free->replaceAllUsesWith(PollyFreeManaged);
+
+      replaceAllUsesAndConstantUses(Free, PollyFreeManaged, Builder);
       Free->eraseFromParent();
     }
 
     SmallPtrSet<GlobalVariable *, 4> GlobalsToErase;
     for (GlobalVariable &Global : M.globals())
       replaceGlobalArray(M, DL, Global, GlobalsToErase);
     for (GlobalVariable *G : GlobalsToErase)
       G->eraseFromParent();

test/GPGPU/managed-memory-rewrite-malloc-free-inside-constexpr.ll

+; RUN: opt %loadPolly -polly-scops \
+; RUN: -analyze < %s | FileCheck %s --check-prefix=SCOP
+
+; RUN: opt %loadPolly -polly-codegen-ppcg \
+; RUN: -S -polly-acc-codegen-managed-memory \
+; RUN: -polly-acc-rewrite-managed-memory < %s | FileCheck %s --check-prefix=HOST-IR
+;
+; REQUIRES: pollyacc
+;
+; Check that we can correctly rewrite `malloc` to `polly_mallocManaged`, and
+; `free` to `polly_freeManaged` with the `polly-acc-rewrite-managed-memory`
+; pass, even inside `constantExpr`. This is necessary because a cookie cutter
+; Inst->replaceUsesOfWith(...) call does not actually work, because this does
+; not replace the instruction within a ConstantExpr.
+;
+; #include <memory.h>
+;
+; static const int N = 100;
+; int* f(int *ToFree) {
+;     free(ToFree);
+;     int *A = (int *)malloc(sizeof(int) * N);
+;     for(int i = 0; i < N; i++) {
+;         A[i] = 42;
+;     }
+;     return A;
+;
+; }
+
+; SCOP:      Function: f
+; SCOP-NEXT: Region: %for.body---%for.end
+; SCOP-NEXT: Max Loop Depth:  1
+
+; SCOP:      Arrays {
+; SCOP-NEXT:     i32 MemRef_tmp[*]; // Element size 4
+; SCOP-NEXT: }
+
+; // Check that polly_mallocManaged is declared and used correctly.
+; HOST-IR: %1 = bitcast i8* (i64)* @polly_mallocManaged to i32* (i64)*
+; HOST-IR: declare i8* @polly_mallocManaged(i64)
+
+; // Check that polly_freeManaged is declared and used correctly.
+; HOST-IR  call void @polly_freeManaged(i8* %toFree)
+; HOST-IR: declare void @polly_freeManaged(i8*)
+
+; // Check that we remove the original malloc,free
+; HOST-IR-NOT: declare i8* @malloc(i64)
+; HOST-IR-NOT: declare void @free(i8*)
+
+target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-macosx10.12.0"
+
+define i32* @f(i32 *%toFree) {
+entry:
+  ; Free inside bitcast
+  call void bitcast (void (i8*)* @free to void (i32 *)*) (i32 * %toFree)
+  br label %entry.split
+
+entry.split:                                      ; preds = %entry
+  ; malloc inside bitcast.
+  %tmp = call i32* bitcast (i8* (i64)* @malloc to i32* (i64)*) (i64 400)
+  br label %for.body
+
+for.body:                                         ; preds = %entry.split, %for.body
+  %indvars.iv1 = phi i64 [ 0, %entry.split ], [ %indvars.iv.next, %for.body ]
+  %arrayidx = getelementptr inbounds i32, i32* %tmp, i64 %indvars.iv1
+  store i32 42, i32* %arrayidx, align 4, !tbaa !3
+  %indvars.iv.next = add nuw nsw i64 %indvars.iv1, 1
+  %exitcond = icmp eq i64 %indvars.iv.next, 100
+  br i1 %exitcond, label %for.end, label %for.body
+
+for.end:                                          ; preds = %for.body
+  ret i32* %tmp
+}
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.start.p0i8(i64, i8* nocapture) #0
+
+declare i8* @malloc(i64)
+declare void @free(i8*)
+
+; Function Attrs: argmemonly nounwind
+declare void @llvm.lifetime.end.p0i8(i64, i8* nocapture) #0
+
+attributes #0 = { argmemonly nounwind }
+
+!llvm.module.flags = !{!0, !1}
+!llvm.ident = !{!2}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 7, !"PIC Level", i32 2}
+!2 = !{!"clang version 6.0.0 (http://llvm.org/git/clang.git 6660f0d30ef23b3142a6b08f9f41aad3d47c084f) (http://llvm.org/git/llvm.git 052dd78cb30f77a05dc8bb06b851402c4b6c6587)"}
+!3 = !{!4, !4, i64 0}
+!4 = !{!"int", !5, i64 0}
+!5 = !{!"omnipotent char", !6, i64 0}
+!6 = !{!"Simple C/C++ TBAA"}