Enabling the copy-constant-to-alloca optimization in more instances

Patch by Mohammad Fawaz

This patch allows lifetime calls to be ignored (and later erased) if we
know that the copy-constant-to-alloca optimization is going to happen.
The case that is missed is when the global variable is in a different address
space than the alloca (as shown in the example added to the lit test.)

This used to work before llvm@6da31fa

Differential Revision: https://reviews.llvm.org/D106573

Author: Andy Kaylor

llvm/lib/Transforms/InstCombine/InstCombineLoadStoreAlloca.cpp

@@ -272,6 +272,8 @@ bool PointerReplacer::collectUsers(Instruction &I) {
         return false;
     } else if (isa<MemTransferInst>(Inst)) {
       Worklist.insert(Inst);
+    } else if (Inst->isLifetimeStartOrEnd()) {
+      continue;
     } else {
       LLVM_DEBUG(dbgs() << "Cannot handle pointer user: " << *U << '\n');
       return false;

llvm/test/Transforms/InstCombine/memcpy-from-global.ll

@@ -73,6 +73,7 @@ declare void @llvm.memcpy.p1i8.p1i8.i64(i8 addrspace(1)* nocapture, i8 addrspace
 
 @G = constant %T {i8 1, [123 x i8] zeroinitializer }
 @H = constant [2 x %U] zeroinitializer, align 16
+@I = internal addrspace(1) constant [4 x float] zeroinitializer , align 4
 
 define void @test2() {
 ; CHECK-LABEL: @test2(
@@ -323,4 +324,22 @@ entry:
   ret void
 }
 
+; Should replace alloca with global even when the global is in a different address space
+define float @test11(i64 %i) {
+; CHECK-LABEL: @test11(
+; CHECK-NEXT: entry:
+; CHECK-NEXT: [[GEP:%.*]] = getelementptr [4 x float], [4 x float] addrspace(1)* @I, i64 0, i64 %i
+; CHECK-NEXT: [[LD:%.*]] = load float, float addrspace(1)* [[GEP]]
+; CHECK-NEXT: ret float [[LD]]
+
+entry:
+  %a = alloca [4 x float], align 4
+  %b = bitcast [4 x float]* %a to i8*
+  call void @llvm.lifetime.start.p0i8(i64 16, i8* %b)
+  call void @llvm.memcpy.p0i8.p1i8.i64(i8* align 4 %b, i8 addrspace(1)* align 4 bitcast ([4 x float] addrspace(1)* @I to i8 addrspace(1)*), i64 16, i1 false)
+  %g = getelementptr inbounds [4 x float], [4 x float]* %a, i64 0, i64 %i
+  %r = load float, float* %g, align 4
+  ret float %r
+}
+
 attributes #0 = { null_pointer_is_valid }