AMDGPU: Annotate amdgpu.noclobber for global loads only

Summary:
  This is to avoid unnecessary analysis since amdgpu.noclobber is only used for globals.

Reviewers:
  arsenm

Fixes:
   SWDEV-239161

Differential Revision:
  https://reviews.llvm.org/D94107

llvm/lib/Target/AMDGPU/AMDGPUAnnotateUniformValues.cpp

@@ -142,10 +142,11 @@ void AMDGPUAnnotateUniformValues::visitLoadInst(LoadInst &I) {
   }
 
   bool NotClobbered = false;
+  bool GlobalLoad = isGlobalLoad(I);
   if (PtrI)
-    NotClobbered = !isClobberedInFunction(&I);
+    NotClobbered = GlobalLoad && !isClobberedInFunction(&I);
   else if (isa<Argument>(Ptr) || isa<GlobalValue>(Ptr)) {
-    if (isGlobalLoad(I) && !isClobberedInFunction(&I)) {
+    if (GlobalLoad && !isClobberedInFunction(&I)) {
       NotClobbered = true;
       // Lookup for the existing GEP
       if (noClobberClones.count(Ptr)) {

llvm/test/CodeGen/AMDGPU/annotate-noclobber.ll

@@ -0,0 +1,47 @@
+; RUN: opt -S --amdgpu-annotate-uniform < %s | FileCheck -check-prefix=OPT %s
+target datalayout = "A5"
+
+
+; OPT-LABEL: @amdgpu_noclobber_global(
+; OPT:      %addr = getelementptr i32, i32 addrspace(1)* %in, i64 0, !amdgpu.uniform !0, !amdgpu.noclobber !0
+; OPT-NEXT: %load = load i32, i32 addrspace(1)* %addr, align 4
+define amdgpu_kernel void @amdgpu_noclobber_global( i32 addrspace(1)* %in,  i32 addrspace(1)* %out) {
+entry:
+  %addr = getelementptr i32, i32 addrspace(1)* %in, i64 0
+  %load = load i32, i32 addrspace(1)* %addr, align 4
+  store i32 %load, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; OPT-LABEL: @amdgpu_noclobber_local(
+; OPT:      %addr = getelementptr i32, i32 addrspace(3)* %in, i64 0, !amdgpu.uniform !0
+; OPT-NEXT: %load = load i32, i32 addrspace(3)* %addr, align 4
+define amdgpu_kernel void @amdgpu_noclobber_local( i32 addrspace(3)* %in,  i32 addrspace(1)* %out) {
+entry:
+  %addr = getelementptr i32, i32 addrspace(3)* %in, i64 0
+  %load = load i32, i32 addrspace(3)* %addr, align 4
+  store i32 %load, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; OPT-LABEL: @amdgpu_noclobber_private(
+; OPT:      %addr = getelementptr i32, i32 addrspace(5)* %in, i64 0, !amdgpu.uniform !0
+; OPT-NEXT: %load = load i32, i32 addrspace(5)* %addr, align 4
+define amdgpu_kernel void @amdgpu_noclobber_private( i32 addrspace(5)* %in,  i32 addrspace(1)* %out) {
+entry:
+  %addr = getelementptr i32, i32 addrspace(5)* %in, i64 0
+  %load = load i32, i32 addrspace(5)* %addr, align 4
+  store i32 %load, i32 addrspace(1)* %out, align 4
+  ret void
+}
+
+; OPT-LABEL: @amdgpu_noclobber_flat(
+; OPT:      %addr = getelementptr i32, i32 addrspace(4)* %in, i64 0, !amdgpu.uniform !0
+; OPT-NEXT: %load = load i32, i32 addrspace(4)* %addr, align 4
+define amdgpu_kernel void @amdgpu_noclobber_flat( i32 addrspace(4)* %in,  i32 addrspace(1)* %out) {
+entry:
+  %addr = getelementptr i32, i32 addrspace(4)* %in, i64 0
+  %load = load i32, i32 addrspace(4)* %addr, align 4
+  store i32 %load, i32 addrspace(1)* %out, align 4
+  ret void
+}

llvm/test/CodeGen/AMDGPU/store-clobbers-load.ll

@@ -4,24 +4,24 @@ target datalayout = "A5"
 ; "load vaddr" depends on the store, so we should not mark vaddr as amdgpu.noclobber.
 
 ; OPT-LABEL: @store_clobbers_load(
-; OPT:      %vaddr = bitcast [4 x i32] addrspace(5)* %alloca to <4 x i32> addrspace(5)*, !amdgpu.uniform !0
-; OPT-NEXT: %zero = load <4 x i32>, <4 x i32> addrspace(5)* %vaddr, align 16
-define amdgpu_kernel void @store_clobbers_load(i32 addrspace(1)* %out, i32 %index) {
+; OPT:      %vaddr = getelementptr <4 x i32>, <4 x i32> addrspace(1)* %input, i64 0, !amdgpu.uniform !0
+; OPT-NEXT: %zero = load <4 x i32>, <4 x i32> addrspace(1)* %vaddr, align 16
+define amdgpu_kernel void @store_clobbers_load( < 4 x i32> addrspace(1)* %input,  i32 addrspace(1)* %out, i32 %index) {
 entry:
-  %alloca = alloca [4 x i32], addrspace(5)
-  %addr0 = bitcast [4 x i32] addrspace(5)* %alloca to i32 addrspace(5)*
-  store i32 0, i32 addrspace(5)* %addr0
-  %vaddr = bitcast [4 x i32] addrspace(5)* %alloca to <4 x i32> addrspace(5)*
-  %zero = load <4 x i32>, <4 x i32> addrspace(5)* %vaddr, align 16
+  %addr0 = bitcast <4 x i32> addrspace(1)* %input to i32 addrspace(1)*
+  store i32 0, i32 addrspace(1)* %addr0
+  %vaddr = getelementptr <4 x i32>, <4 x i32> addrspace(1)* %input, i64 0
+  %zero = load <4 x i32>, <4 x i32> addrspace(1)* %vaddr, align 16
   %one = insertelement <4 x i32> %zero, i32 1, i32 1
   %two = insertelement <4 x i32> %one, i32 2, i32 2
   %three = insertelement <4 x i32> %two, i32 3, i32 3
-  store <4 x i32> %three, <4 x i32> addrspace(5)* %vaddr, align 16
+  store <4 x i32> %three, <4 x i32> addrspace(1)* %input, align 16
   %rslt = extractelement <4 x i32> %three, i32 %index
   store i32 %rslt, i32 addrspace(1)* %out, align 4
   ret void
 }
 
+
 declare i32 @llvm.amdgcn.workitem.id.x()
 @lds0 = addrspace(3) global [512 x i32] undef, align 4