[AMDGPU] Inline non-kernel functions using extern lds

[vangthao]

Summary

In https://reviews.llvm.org/D100481, forceful inline of all non-kernel
functions using lds was disabled since AMDGPULowerModuleLDS pass now handles
static lds. However that pass does not handle extern lds so non-kernel
functions using extern lds must sill be inline.

[hsmhsm]

LGTM

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUAlwaysInlinePass.cpp

125

What does the hasInitializer check have to do with this? Initializers are illegal for LDS anyway

[vangthao]

llvm/lib/Target/AMDGPU/AMDGPUAlwaysInlinePass.cpp

125

From https://github.com/llvm/llvm-project/blob/main/llvm/lib/Target/AMDGPU/Utils/AMDGPULDSUtils.cpp#L319, this implies that the variable is extern shared. Is there a better method for checking extern lds?

[arsenm]

llvm/lib/Target/AMDGPU/AMDGPUAlwaysInlinePass.cpp

125

This is just a weird helper function, I would have expected !isDeclaration to work just as well

llvm/lib/Target/AMDGPU/AMDGPUAlwaysInlinePass.cpp

static bool alwaysInlineImpl(Module &M, bool GlobalOpt) {
   // should only appear when IPO passes manages to move LDs defined in a kernel
   // into a single user function.
 
   for (GlobalVariable &GV : M.globals()) {
     // TODO: Region address
     unsigned AS = GV.getAddressSpace();
     if ((AS == AMDGPUAS::REGION_ADDRESS) ||
         (AS == AMDGPUAS::LOCAL_ADDRESS &&
-         !AMDGPUTargetMachine::EnableLowerModuleLDS))
+         (!AMDGPUTargetMachine::EnableLowerModuleLDS || !GV.hasInitializer())))
       recursivelyVisitUsers(GV, FuncsToAlwaysInline);
   }
 
   if (!AMDGPUTargetMachine::EnableFunctionCalls || StressCalls) {
     auto IncompatAttr
       = StressCalls ? Attribute::AlwaysInline : Attribute::NoInline;
 
     for (Function &F : M) {

llvm/test/CodeGen/AMDGPU/hip.extern.shared.array.ll

define amdgpu_kernel void @dynamic_shared_array_6(i32 %idx) {
   %val2 = uitofp i8 %val0 to double
   %arrayidx1 = getelementptr inbounds [0 x float], [0 x float] addrspace(3)* @dynamic_shared0, i32 0, i32 %tid.x
   store float %val1, float addrspace(3)* %arrayidx1, align 4
   %arrayidx2 = getelementptr inbounds [0 x double], [0 x double] addrspace(3)* @dynamic_shared3, i32 0, i32 %tid.x
   store double %val2, double addrspace(3)* %arrayidx2, align 4
   ret void
 }
 
+; CHECK-LABEL: dynamic_shared_array_with_call:
+; CHECK-NOT: s_swappc_b64
+define amdgpu_kernel void @dynamic_shared_array_with_call(float addrspace(1)* nocapture readnone %out) local_unnamed_addr {
+  %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x()
+  %1 = sext i32 %tid.x to i64
+  %arrayidx0 = getelementptr inbounds [512 x float], [512 x float] addrspace(3)* @lds0, i64 0, i64 %1
+  %val0 = load float, float addrspace(3)* %arrayidx0, align 4
+  tail call void @store_value(float %val0)
+  ret void
+}
+
+; CHECK-NOT: store_value
+define linkonce_odr hidden void @store_value(float %val1) local_unnamed_addr {
+entry:
+  %tid.x = tail call i32 @llvm.amdgcn.workitem.id.x()
+  %0 = sext i32 %tid.x to i64
+  %arrayidx1 = getelementptr inbounds [0 x float], [0 x float] addrspace(3)* @dynamic_shared0, i64 0, i64 %0
+  store float %val1, float addrspace(3)* %arrayidx1, align 4
+  ret void
+}
+
 declare i32 @llvm.amdgcn.workitem.id.x()