[AMDGPU] Run always inliner early in opt

[rampitec]

Summary

We can mark functions to always inline early in the opt. Since we do not have
call support this early inlining creates opportunities for inter-procedural
optimizations which would not occur otherwise.

[vpykhtin]

LGTM.

[arsenm]

I'm not sure we should immediately inline. Some level of simplification should occur on the individual functions first. Otherwise the code size is growing and increasing compile time everywhere. AlwaysInline doesn't immediately inline the function though. When after is the inliner actually run?

[rampitec]

In D31016#702818, @arsenm wrote:

I'm not sure we should immediately inline. Some level of simplification should occur on the individual functions first. Otherwise the code size is growing and increasing compile time everywhere. AlwaysInline doesn't immediately inline the function though. When after is the inliner actually run?

Right, we do not immediately inline, we just immediately mark all functions for inlining. The inliner pass runs in its regular place in opt. Also module passes work after function passes, so function simplifications happen.

llvm/trunk/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp

[Internalize](const PassManagerBuilder &, legacy::PassManagerBase &PM) {
             case CallingConv::SPIR_KERNEL:
               return true;
             }
           }
           return !GV.use_empty();
         }));
         PM.add(createGlobalDCEPass());
       }
+      PM.add(createAMDGPUAlwaysInlinePass());
   });
 }
 
 //===----------------------------------------------------------------------===//
 // R600 Target Machine (R600 -> Cayman)
 //===----------------------------------------------------------------------===//
 
 R600TargetMachine::R600TargetMachine(const Target &T, const Triple &TT,

llvm/trunk/test/CodeGen/AMDGPU/early-inline.ll

+; RUN: opt -mtriple=amdgcn-- -O1 -S -inline-threshold=1 %s | FileCheck %s
+
+define i32 @callee(i32 %x) {
+entry:
+  %mul1 = mul i32 %x, %x
+  %mul2 = mul i32 %mul1, %x
+  %mul3 = mul i32 %mul1, %mul2
+  %mul4 = mul i32 %mul3, %mul2
+  %mul5 = mul i32 %mul4, %mul3
+  ret i32 %mul5
+}
+
+; CHECK-LABEL: @caller
+; CHECK: mul i32
+; CHECK-NOT: call i32
+
+define i32 @caller(i32 %x) {
+entry:
+  %res = call i32 @callee(i32 %x)
+  ret i32 %res
+}