[X86] for split stack, not save/restore nested arg if unused

[cherry]

Summary

For split-stack, if the nested argument (i.e. R10) is not used, no need to save/restore it in the prologue.

[thanm]

Good cleanup -- LGTM.

We can obviously test this with Go -- do we know of any other easily usable testcases/applications that use the nested argument?

[cherry]

Thanks for the review.

We can obviously test this with Go -- do we know of any other easily usable testcases/applications that use the nested argument?

Closures do. The save/restore R10 code path is used when a closure call splits the stack. For example, a recursive closure like https://play.golang.org/p/6t1xV8ThJuN .

[thanm]

SGTM

llvm/trunk/lib/Target/X86/X86FrameLowering.cpp

void X86FrameLowering::determineCalleeSaves(MachineFunction &MF,
   }
 }
 
 static bool
 HasNestArgument(const MachineFunction *MF) {
   const Function &F = MF->getFunction();
   for (Function::const_arg_iterator I = F.arg_begin(), E = F.arg_end();
        I != E; I++) {
-    if (I->hasNestAttr())
+    if (I->hasNestAttr() && !I->use_empty())
       return true;
   }
   return false;
 }
 
 /// GetScratchRegister - Get a temp register for performing work in the
 /// segmented stack and the Erlang/HiPE stack prologue. Depending on platform
 /// and the properties of the function either one or two registers will be

llvm/trunk/test/CodeGen/X86/segmented-stacks.ll

 
 define void @test_nostack() #0 {
   ret void
 
 ; X32-Linux-LABEL: test_nostack:
 ; X32-Linux-NOT:   calll __morestack
 
 ; X64-Linux-LABEL: test_nostack:
-; X32-Linux-NOT:   callq __morestack
+; X64-Linux-NOT:   callq __morestack
 
 ; X32ABI-LABEL: test_nostack:
 ; X32ABI-NOT:   callq __morestack
 
 ; X32-Darwin-LABEL: test_nostack:
 ; X32-Darwin-NOT:   calll __morestack
 
 ; X64-Darwin-LABEL: test_nostack:

 ; X32-DFlyBSD:  calll __morestack
 ; X32-DFlyBSD-NEXT:  ret
 
 ; X64-DFlyBSD-LABEL:       test_sibling_call_empty_frame:
 ; X64-DFlyBSD:  callq __morestack
 
 }
 
+; Test that unused nested argument doesn't need saving/restoring.
+
+define i32 @test_nested_unused(i32 * nest %unused) #0 {
+       %mem = alloca i32, i32 10
+       call void @dummy_use (i32* %mem, i32 10)
+       ret i32 123
+
+; X64-Linux-LABEL: test_nested_unused:
+; X64-Linux-NOT:   movq %r10, %rax
+; X64-Linux:       callq __morestack
+; X64-Linux-NOT:   movq %rax, %r10
+
+; X64-Darwin-LABEL: test_nested_unused:
+; X64-Darwin-NOT:   movq %r10, %rax
+; X64-Darwin:       callq ___morestack
+; X64-Darwin-NOT:   movq %rax, %r10
+
+; X64-FreeBSD-LABEL: test_nested_unused:
+; X64-FreeBSD-NOT:   movq %r10, %rax
+; X64-FreeBSD:       callq __morestack
+; X64-FreeBSD-NOT:   movq %rax, %r10
+
+; X64-DFlyBSD-LABEL: test_nested_unused:
+; X64-DFlyBSD-NOT:   movq %r10, %rax
+; X64-DFlyBSD:       callq __morestack
+; X64-DFlyBSD-NOT:   movq %rax, %r10
+
+; X64-MinGW-LABEL: test_nested_unused:
+; X64-MinGW-NOT:   movq %r10, %rax
+; X64-MinGW:       callq __morestack
+; X64-MinGW-NOT:   movq %rax, %r10
+
+; X32ABI-LABEL: test_nested_unused:
+; X32ABI-NOT:   movl %r10d, %eax
+; X32ABI:       callq __morestack
+; X32ABI-NOT:   movq %rax, %r10
+
+}
+
 attributes #0 = { "split-stack" }
 
 ; X64-Linux-Large: .rodata
 ; X64-Linux-Large-NEXT: __morestack_addr:
 ; X64-Linux-Large-NEXT: .quad __morestack
 
 ; X32-Linux: .section ".note.GNU-split-stack","",@progbits
 ; X32-Linux: .section ".note.GNU-no-split-stack","",@progbits