[llvm] r256959 - [X86] Correctly model TLS calls w.r.t. frame requirements.

Quentin Colombet via llvm-commits llvm-commits at lists.llvm.org
Wed Jan 6 11:23:04 PST 2016


Hi Manman,

Just a heads-up to tell you that the fix for PR25820 landed and as discussed before the holidays, this implies poorer codegen for the cxx fast TLS calling convention on x86 since shrink-wrapping cannot help you anymore. I.e., you need to apply the same tricks that you did for AArch64 on x86.

Cheers,
-Quentin 
> On Jan 6, 2016, at 11:09 AM, Quentin Colombet via llvm-commits <llvm-commits at lists.llvm.org> wrote:
> 
> Author: qcolombet
> Date: Wed Jan  6 13:09:26 2016
> New Revision: 256959
> 
> URL: http://llvm.org/viewvc/llvm-project?rev=256959&view=rev
> Log:
> [X86] Correctly model TLS calls w.r.t. frame requirements.
> TLS calls need the stack frame to be properly set up and this
> implies that such calls need ADJUSTSTACK_xxx markers.
> 
> Fixes PR25820.
> 
> Modified:
>    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
>    llvm/trunk/test/CodeGen/X86/cxx_tlscc64.ll
>    llvm/trunk/test/CodeGen/X86/x86-shrink-wrapping.ll
> 
> Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=256959&r1=256958&r2=256959&view=diff
> ==============================================================================
> --- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
> +++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Wed Jan  6 13:09:26 2016
> @@ -12474,8 +12474,12 @@ X86TargetLowering::LowerGlobalTLSAddress
>     // location.
>     SDValue Chain = DAG.getEntryNode();
>     SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
> +    Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, DL, true), DL);
>     SDValue Args[] = { Chain, Offset };
>     Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
> +    Chain =
> +        DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, DL, true),
> +                           DAG.getIntPtrConstant(0, DL, true), SDValue(), DL);
> 
>     // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
>     MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
> 
> Modified: llvm/trunk/test/CodeGen/X86/cxx_tlscc64.ll
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/cxx_tlscc64.ll?rev=256959&r1=256958&r2=256959&view=diff
> ==============================================================================
> --- llvm/trunk/test/CodeGen/X86/cxx_tlscc64.ll (original)
> +++ llvm/trunk/test/CodeGen/X86/cxx_tlscc64.ll Wed Jan  6 13:09:26 2016
> @@ -1,5 +1,9 @@
> ; RUN: llc < %s -mtriple=x86_64-apple-darwin | FileCheck %s
> -; RUN: llc < %s -mtriple=x86_64-apple-darwin -enable-shrink-wrap=true | FileCheck --check-prefix=SHRINK %s
> +; TLS function were wrongly model and after fixing that, shrink-wrapping
> +; cannot help here. To achieve the expected lowering, we need to playing
> +; tricks similar to AArch64 fast TLS calling convention (r255821).
> +; Re-enable the following run line when 
> +; _RUN_: llc < %s -mtriple=x86_64-apple-darwin -enable-shrink-wrap=true | FileCheck --check-prefix=SHRINK %s
> %struct.S = type { i8 }
> 
> @sg = internal thread_local global %struct.S zeroinitializer, align 1
> 
> Modified: llvm/trunk/test/CodeGen/X86/x86-shrink-wrapping.ll
> URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/x86-shrink-wrapping.ll?rev=256959&r1=256958&r2=256959&view=diff
> ==============================================================================
> --- llvm/trunk/test/CodeGen/X86/x86-shrink-wrapping.ll (original)
> +++ llvm/trunk/test/CodeGen/X86/x86-shrink-wrapping.ll Wed Jan  6 13:09:26 2016
> @@ -878,4 +878,48 @@ for.end:
> 
> declare i32 @varfunc(i8* nocapture readonly)
> 
> + at sum1 = external hidden thread_local global i32, align 4
> +
> +
> +; Function Attrs: nounwind
> +; Make sure the TLS call used to access @sum1 happens after the prologue
> +; and before the epilogue.
> +; TLS calls used to be wrongly model and shrink-wrapping would have inserted
> +; the prologue and epilogue just around the call to doSomething.
> +; PR25820.
> +;
> +; CHECK-LABEL: tlsCall:
> +; CHECK: pushq
> +; CHECK: testb $1, %dil
> +; CHECK: je [[ELSE_LABEL:LBB[0-9_]+]]
> +;
> +; master bb
> +; CHECK: movq _sum1 at TLVP(%rip), %rdi
> +; CHECK-NEXT: callq *(%rdi)
> +; CHECK: jmp [[EXIT_LABEL:LBB[0-9_]+]]
> +;
> +; [[ELSE_LABEL]]:
> +; CHECK: callq _doSomething
> +;
> +; [[EXIT_LABEL]]:
> +; CHECK: popq
> +; CHECK-NEXT: retq
> +define i32 @tlsCall(i1 %bool1, i32 %arg, i32* readonly dereferenceable(4) %sum1) #3 {
> +entry:
> +  br i1 %bool1, label %master, label %else
> +
> +master:
> +  %tmp1 = load i32, i32* %sum1, align 4
> +  store i32 %tmp1, i32* @sum1, align 4
> +  br label %exit
> +
> +else:
> +  %call = call i32 @doSomething(i32 0, i32* null)
> +  br label %exit
> +
> +exit:
> +  %res = phi i32 [ %arg, %master], [ %call, %else ]
> +  ret i32 %res
> +}
> +
> attributes #3 = { nounwind }
> 
> 
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits



More information about the llvm-commits mailing list