[GlobalOpt] Extend logic in SRA heuristic to skip stores of initializer.

[fhahn]

Summary

If all stores only store the initializer value of a global, consider it
as not stored in the heuristic. GlobalOpt will remove such stores later
on.

Depends on D129857.

[fhahn]

Rebase and ping :)

[nikic]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

401

Rebase over https://github.com/llvm/llvm-project/commit/0ecef88cb167dbd19023a932e7abac09059c3e4e?

409

Shouldn't we be comparing with the initializer at the specific offset?

[fhahn]

Update to use ConstantFoldLoadFromConst to check the initializer value in the specifi field.

[fhahn]

Rebase on top of 0ecef88.

[fhahn]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

401

Done!

409

Yeah, I updated the code to use ConstantFoldLoadFromConst

[nikic]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

409

Why can't this reuse the offset calculation from above?

[fhahn]

Use existing offset directly.

[fhahn]

llvm/lib/Transforms/IPO/GlobalOpt.cpp

409

Missed that, should be updated now, thanks!

[nikic]

LGTM, though I think it might be cleaner if we collected the initializer part during the construction of GlobalPart, rather than calculating it twice, here and during the SRA split.

[fhahn]

In D144476#4171586, @nikic wrote:

LGTM, though I think it might be cleaner if we collected the initializer part during the construction of GlobalPart, rather than calculating it twice, here and during the SRA split.

Posted D145490 as follow-up

llvm/lib/Transforms/IPO/GlobalOpt.cpp

if (Value *Ptr = getLoadStorePointerOperand(V)) {
       auto It = Parts.try_emplace(Offset.getZExtValue(), GlobalPart{Ty}).first;
       if (Ty != It->second.Ty)
         return false;
 
       // Scalable types not currently supported.
       if (isa<ScalableVectorType>(Ty))
         return false;
 
+      auto IsStored = [GV](Value *V) {
+        auto *SI = dyn_cast<StoreInst>(V);
+        if (!SI)
+          return false;
+
+        auto *GlobalVar = dyn_cast<GlobalVariable>(GV);
+        Constant *StoredConst = dyn_cast<Constant>(SI->getOperand(0));
+        if (!GlobalVar || !StoredConst)
+          return true;
+
+        // Don't consider stores that only write the initializer value.
+        return StoredConst != GlobalVar->getInitializer() &&
+               !(StoredConst->isNullValue() &&
+                 GlobalVar->getInitializer()->isNullValue());
+      };
+
       It->second.IsLoaded |= isa<LoadInst>(V);
-      It->second.IsStored |= isa<StoreInst>(V);
+      It->second.IsStored |= IsStored(V);
       continue;
     }
 
     // Ignore dead constant users.
     if (auto *C = dyn_cast<Constant>(V)) {
       if (!isSafeToDestroyConstant(C))
         return false;
       continue;

llvm/test/Transforms/GlobalOpt/sra-many-stores.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals
 ; RUN: opt -passes=globalopt -S %s | FileCheck %s
 
 %struct.widget = type { ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr, ptr }
 
 @global = internal global %struct.widget zeroinitializer
 
 
 ;.
 ; CHECK: @[[A:[a-zA-Z0-9_$"\\.-]+]] = global i8 0, align 4
+; CHECK: @[[C:[a-zA-Z0-9_$"\\.-]+]] = global ptr null
 ;.
 define internal void @read_struct() {
 ; CHECK-LABEL: @read_struct(
 ; CHECK-NEXT:    ret void
 ;
   %tmp = load ptr, ptr getelementptr inbounds (%struct.widget, ptr @global, i64 0, i32 16), align 8
   ret void
 }

 
 @global.20ptr = internal global %struct.20ptr zeroinitializer
 
 @c = global ptr null
 
 define void @store_initializer() {
 ; CHECK-LABEL: @store_initializer(
 ; CHECK-NEXT:  entry:
-; CHECK-NEXT:    store ptr null, ptr @global.20ptr, align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR:%.*]], ptr @global.20ptr, i64 0, i32 1), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 2), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 3), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 4), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 5), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 6), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 7), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 8), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 9), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 10), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 11), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 12), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 13), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 14), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 15), align 8
-; CHECK-NEXT:    store ptr null, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 16), align 8
-; CHECK-NEXT:    [[L0:%.*]] = load ptr, ptr @global.20ptr, align 8
-; CHECK-NEXT:    store volatile ptr [[L0]], ptr @c, align 8
-; CHECK-NEXT:    [[L1:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 1), align 8
-; CHECK-NEXT:    store volatile ptr [[L1]], ptr @c, align 8
-; CHECK-NEXT:    [[L2:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 2), align 8
-; CHECK-NEXT:    store volatile ptr [[L2]], ptr @c, align 8
-; CHECK-NEXT:    [[L3:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 3), align 8
-; CHECK-NEXT:    store volatile ptr [[L3]], ptr @c, align 8
-; CHECK-NEXT:    [[L4:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 4), align 8
-; CHECK-NEXT:    store volatile ptr [[L4]], ptr @c, align 8
-; CHECK-NEXT:    [[L5:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 5), align 8
-; CHECK-NEXT:    store volatile ptr [[L5]], ptr @c, align 8
-; CHECK-NEXT:    [[L6:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 6), align 8
-; CHECK-NEXT:    store volatile ptr [[L6]], ptr @c, align 8
-; CHECK-NEXT:    [[L7:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 7), align 8
-; CHECK-NEXT:    store volatile ptr [[L7]], ptr @c, align 8
-; CHECK-NEXT:    [[L8:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 8), align 8
-; CHECK-NEXT:    store volatile ptr [[L8]], ptr @c, align 8
-; CHECK-NEXT:    [[L9:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 9), align 8
-; CHECK-NEXT:    store volatile ptr [[L9]], ptr @c, align 8
-; CHECK-NEXT:    [[L10:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 10), align 8
-; CHECK-NEXT:    store volatile ptr [[L10]], ptr @c, align 8
-; CHECK-NEXT:    [[L11:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 11), align 8
-; CHECK-NEXT:    store volatile ptr [[L11]], ptr @c, align 8
-; CHECK-NEXT:    [[L12:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 12), align 8
-; CHECK-NEXT:    store volatile ptr [[L12]], ptr @c, align 8
-; CHECK-NEXT:    [[L13:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 13), align 8
-; CHECK-NEXT:    store volatile ptr [[L13]], ptr @c, align 8
-; CHECK-NEXT:    [[L14:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 14), align 8
-; CHECK-NEXT:    store volatile ptr [[L14]], ptr @c, align 8
-; CHECK-NEXT:    [[L15:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 15), align 8
-; CHECK-NEXT:    store volatile ptr [[L15]], ptr @c, align 8
-; CHECK-NEXT:    [[L16:%.*]] = load ptr, ptr getelementptr inbounds ([[STRUCT_20PTR]], ptr @global.20ptr, i64 0, i32 16), align 8
-; CHECK-NEXT:    store volatile ptr [[L16]], ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
+; CHECK-NEXT:    store volatile ptr null, ptr @c, align 8
 ; CHECK-NEXT:    ret void
 ;
 entry:
   store ptr null, ptr getelementptr inbounds (%struct.20ptr, ptr @global.20ptr, i64 0, i32 0), align 8
   store ptr null, ptr getelementptr inbounds (%struct.20ptr, ptr @global.20ptr, i64 0, i32 1), align 8
   store ptr null, ptr getelementptr inbounds (%struct.20ptr, ptr @global.20ptr, i64 0, i32 2), align 8
   store ptr null, ptr getelementptr inbounds (%struct.20ptr, ptr @global.20ptr, i64 0, i32 3), align 8
   store ptr null, ptr getelementptr inbounds (%struct.20ptr, ptr @global.20ptr, i64 0, i32 4), align 8