[Polly] [Fortran Support] [Hack] [WIP] add option to treat function parameters as dereferencable

[bollu]

Summary

Dragonegg generates most function parameters as pointers to the actual
parameters. However, it does not mark these parameters with the
dereferencable attribute.

Polly is conservative when it comes to invariant load
hoisting, thus we refuse to invariant load hoist when we do not know
that pointers are dereferencable. This is correct behaviour, but is
inconvenient.

Add a flag that allows all pointer parameters to be dereferencable.

Please revert this in 2018.

[bollu]

@grosser - Could you please post the test case for this?

[grosser]

That's the one I sent you via email yesterday.

[bollu]

• Update code and add test case.

[bollu]

@grosser - review please.

[grosser]

Hi Siddharth,

instead of the comment "revert in 2008", can you add a TODO: contribute back to dragonegg, in case it is updated to a recent LLVM.

lib/Analysis/ScopInfo.cpp

3852 ↗

(On Diff #111948)

Debug statement

3855 ↗

(On Diff #111948)

OK, that looks good to me.

3881 ↗

(On Diff #111948)

Unrelated.

[bollu]

• [NFC] fix whitespace and debug print

[grosser]

It seems this diff is corrupted in some way. However, your changes look good.

lib/CodeGen/PPCGCodeGeneration.cpp

1384 ↗

(On Diff #111950)

Unrelated?

1409 ↗

(On Diff #111950)

Unrelated.

[bollu]

• [NFC] fix whitespace and debug print (Fix screwed up diff due to rebase).

polly/trunk/lib/Analysis/ScopInfo.cpp

PollyIgnoreInbounds("polly-ignore-inbounds",
                         cl::Hidden, cl::init(false), cl::cat(PollyCategory));
 
 static cl::opt<bool> PollyIgnoreParamBounds(
     "polly-ignore-parameter-bounds",
     cl::desc(
         "Do not add parameter bounds and do no gist simplify sets accordingly"),
     cl::Hidden, cl::init(false), cl::cat(PollyCategory));
 
+static cl::opt<bool> PollyAllowDereferenceOfAllFunctionParams(
+    "polly-allow-dereference-of-all-function-parameters",
+    cl::desc(
+        "Treat all parameters to functions that are pointers as dereferencible."
+        " This is useful for invariant load hoisting, since we can generate"
+        " less runtime checks. This is only valid if all pointers to functions"
+        " are always initialized, so that Polly can choose to hoist"
+        " their loads. "),
+    cl::Hidden, cl::init(false), cl::cat(PollyCategory));
+
 static cl::opt<bool> PollyPreciseFoldAccesses(
     "polly-precise-fold-accesses",
     cl::desc("Fold memory accesses to model more possible delinearizations "
              "(does not scale well)"),
     cl::Hidden, cl::init(false), cl::cat(PollyCategory));
 
 bool polly::UseInstructionNames;
 static cl::opt<bool, true> XUseInstructionNames(
for (auto &IAClass : InvariantEquivClasses) {
     for (auto *MA : MAs)
       if (MA->getAccessInstruction() == Val)
         return &IAClass;
   }
 
   return nullptr;
 }
 
+bool isAParameter(llvm::Value *maybeParam, const Function &F) {
+  for (const llvm::Argument &Arg : F.args())
+    if (&Arg == maybeParam)
+      return true;
+
+  return false;
+};
+
 bool Scop::canAlwaysBeHoisted(MemoryAccess *MA, bool StmtInvalidCtxIsEmpty,
                               bool MAInvalidCtxIsEmpty,
                               bool NonHoistableCtxIsEmpty) {
   LoadInst *LInst = cast<LoadInst>(MA->getAccessInstruction());
   const DataLayout &DL = LInst->getParent()->getModule()->getDataLayout();
+  if (PollyAllowDereferenceOfAllFunctionParams &&
+      isAParameter(LInst->getPointerOperand(), getFunction()))
+    return true;
+
   // TODO: We can provide more information for better but more expensive
   //       results.
   if (!isDereferenceableAndAlignedPointer(LInst->getPointerOperand(),
                                           LInst->getAlignment(), DL))
     return false;
 
   // If the location might be overwritten we do not hoist it unconditionally.
   //

polly/trunk/test/ScopInfo/allow-all-parameters-dereferencable.ll

+; RUN: opt %loadPolly -analyze -polly-invariant-load-hoisting \
+; RUN: -polly-allow-dereference-of-all-function-parameters \
+; RUN: -polly-scops < %s | FileCheck %s --check-prefix=SCOP
+
+; RUN: opt %loadPolly -S -polly-invariant-load-hoisting \
+; RUN: -polly-codegen < %s | FileCheck %s --check-prefix=CODE-RTC
+
+
+; RUN: opt %loadPolly -S -polly-invariant-load-hoisting \
+; RUN: -polly-allow-dereference-of-all-function-parameters \
+; RUN: -polly-codegen < %s | FileCheck %s --check-prefix=CODE
+
+; SCOP:      Function: hoge
+; SCOP-NEXT: Region: %bb15---%bb37
+; SCOP-NEXT: Max Loop Depth:  2
+; SCOP-NEXT: Invariant Accesses: {
+; SCOP-NEXT:         ReadAccess :=  [Reduction Type: NONE] [Scalar: 0]
+; SCOP-NEXT:             [tmp, tmp17, tmp28] -> { Stmt_bb29[i0] -> MemRef_arg1[0] };
+; SCOP-NEXT:         Execution Context: [tmp, tmp17, tmp28] -> {  :  }
+; SCOP-NEXT:         ReadAccess :=  [Reduction Type: NONE] [Scalar: 0]
+; SCOP-NEXT:             [tmp, tmp17, tmp28] -> { Stmt_bb27[] -> MemRef_arg[0] };
+; SCOP-NEXT:         Execution Context: [tmp, tmp17, tmp28] -> {  :  }
+; SCOP-NEXT: }
+
+; Check that without the option `-pollt-allow-dereference-of-all-function-parameters`
+; we do generate the runtime check.
+; CODE-RTC: polly.preload.cond:                               ; preds = %polly.preload.begin
+; CODE-RTC-NEXT: br i1 %{{[a-zA-Z0-9]*}}, label %polly.preload.exec, label %polly.preload.merge
+
+; Check that we don't generate a runtime check because we treat all
+; parameters as dereferencable.
+; CODE-NOT: polly.preload.cond:                               ; preds = %polly.preload.begin
+; CODE-NOT: br i1 %{{r1:[a-zA-Z0-9]*}}, label %polly.preload.exec, label %polly.preload.merge
+
+target datalayout = "e-p:64:64:64-S128-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f16:16:16-f32:32:32-f64:64:64-f128:128:128-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
+target triple = "x86_64-unknown-linux-gnu"
+
+@global = external global i32
+
+; Function Attrs: nounwind uwtable
+define void @hoge(i32* noalias %arg, i32* noalias %arg1, [0 x double]* noalias %arg2, float* %A) #0 {
+bb:
+  %tmp = load i32, i32* @global, align 4
+  %tmp3 = icmp sgt i32 %tmp, 1
+  br label %bb14
+
+bb14:                                             ; preds = %bb
+  br label %bb15
+
+bb15:                                             ; preds = %bb14
+  br i1 %tmp3, label %bb16, label %bb27
+
+bb16:                                             ; preds = %bb15
+  %tmp17 = load i32, i32* %arg1, align 4
+  br label %bb18
+
+bb18:                                             ; preds = %bb18, %bb16
+  %tmp19 = phi i32 [ %tmp25, %bb18 ], [ 1, %bb16 ]
+  %tmp20 = sext i32 %tmp19 to i64
+  %tmp21 = add nsw i64 %tmp20, -1
+  %tmp22 = getelementptr [0 x double], [0 x double]* %arg2, i64 0, i64 %tmp21
+  %tmp23 = bitcast double* %tmp22 to i64*
+  store i64 undef, i64* %tmp23, align 8
+  %tmp24 = icmp eq i32 %tmp19, %tmp17
+  %tmp25 = add i32 %tmp19, 1
+  br i1 %tmp24, label %bb26, label %bb18
+
+bb26:                                             ; preds = %bb18
+  br label %bb27
+
+bb27:                                             ; preds = %bb26, %bb15
+  %tmp28 = load i32, i32* %arg, align 4
+  store float 42.0, float* %A
+  br label %bb29
+
+bb29:                                             ; preds = %bb35, %bb27
+  %tmp30 = load i32, i32* %arg1, align 4
+  store float 42.0, float* %A
+  br label %bb31
+
+bb31:                                             ; preds = %bb31, %bb29
+  %tmp32 = phi i32 [ undef, %bb31 ], [ 1, %bb29 ]
+  store float 42.0, float* %A
+  %tmp33 = icmp eq i32 %tmp32, %tmp30
+  br i1 %tmp33, label %bb34, label %bb31
+
+bb34:                                             ; preds = %bb31
+  br label %bb35
+
+bb35:                                             ; preds = %bb34
+  %tmp36 = icmp eq i32 0, %tmp28
+  br i1 %tmp36, label %bb37, label %bb29
+
+bb37:                                             ; preds = %bb35
+  ret void
+}
+
+attributes #0 = { nounwind uwtable }