[lldb-dev] r260768 (Removed many JIT workarounds from IRForTarget) broke expression parser with function on Hexagon

Wed Feb 24 16:56:55 PST 2016

I added this to ResolveConstantValue():

        case Value::FunctionVal:

            if (const Function *constant_func = dyn_cast<Function>(constant))

            {

                value = 0;

                return true;

            }

            break;

value is an APInt &, so it won’t take nullptr.

“p main” returned the correct signature, but an address of 0.

“p factorial(5)” alternates between returning the parameter (5, in this case) and an error, with the program restarting when it gets the error. _start on Hexagon is 0, so my guess is lldb thought &factorial() was 0, so it set the pc to there and continued. The error is “error: Supposed to interpret, but failed: ThreadPlanCallFunctionUsingABI failed”. By half the time

Something interesting – constant_func->getName().str().c_str() gives “”. I’m pretty sure I want to return the address of the function, but I’m not sure how to get it.

--

Qualcomm Innovation Center, Inc.

The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project

From: scallanan at apple.com [mailto:scallanan at apple.com] 
Sent: Wednesday, February 24, 2016 5:10 PM
To: Ted Woodward
Subject: Re: r260768 (Removed many JIT workarounds from IRForTarget) broke expression parser with function on Hexagon

Yeah, I think we’ll have to return true in CanResolveConstant() for FunctionVal, like you did as an experiment earlier.

Sean

On Feb 24, 2016, at 3:08 PM, Ted Woodward <ted.woodward at codeaurora.org <mailto:ted.woodward at codeaurora.org> > wrote:

It never gets to IRInterpreter::Interpret(); it errors out in IRInterpreter::CanInterpret(), in the call to CanResolveConstant() at line 656, so it never calls getCalledValue() or EvaluateValue().

--

Qualcomm Innovation Center, Inc.

The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project

From: scallanan at apple.com <mailto:scallanan at apple.com>  [mailto:scallanan at apple.com] 
Sent: Wednesday, February 24, 2016 4:36 PM
To: Ted Woodward
Cc: LLDB
Subject: Re: r260768 (Removed many JIT workarounds from IRForTarget) broke expression parser with function on Hexagon

I think I understand what’s going on.  The IR interpreter does this [IRInterpreter.cpp:1573]:

–

                // Find the address of the callee function

                lldb_private::Scalar I;

                const llvm::Value *val = call_inst->getCalledValue();

                if (!frame.EvaluateValue(I, val, module))

                {

                    error.SetErrorToGenericError();

                    error.SetErrorString("unable to get address of function");

                    return false;

                }

–

It assumes that getCalledValue returns something we can evaluate – which used to be a function pointer so everything was fine.

We have to modify EvaluateValue to, when it encounters a FunctionVal, look up the corresponding symbol (using IRExecutionUnit::FindSymbol(), ideally) and return a Scalar containing the function pointer.

As a first step, I’d suggest trying to make EvaluateValue return true but just put nullptr into the Scalar when it gets a FunctionVal – does that result (as I’d expect) in a bad function call, or do we still get some kind of “doesn’t handle” error?

Sean

On Feb 24, 2016, at 2:17 PM, Ted Woodward < <mailto:ted.woodward at codeaurora.org> ted.woodward at codeaurora.org> wrote:

I did some digging and found where the ID is being changed from 0x5 to 0xa in the original code.

IRForTarget::runOnModule() calls ResolveFunctionPointers(), which gets a Constant * from BuildFunctionPointer(). This Value has an ID of 0xa, and runOnModule() then calls the original Function’s replaceAllUsesWith(), passing in the new Value, changing ID from 0x5 to 0xa.

If I comment out the call to ResolveFunctionPointer() in the original code, I get the same error as the current code: “error: Can't run the expression locally: Interpreter doesn't handle one of the expression's operands”

So I went to r260768 and added back in the call to ResolveFunctionPointer(), as well as functions that it depended on:

ClangExpressionDeclMap::FindCodeSymbolInContext()

ClangExpressionDeclMap::FindBestAlternateMangledName()

ClangExpressionDeclMap::GetFunctionAddress()

IRForTarget::GetFunctionAddress()

IRForTarget::BuildFunctionPointer()

I commented out the call to RegisterFunctionMetadata() since it didn’t seem to matter.

After those changes, I was able to print out main’s info and call functions with the expression parser.

Is there anything in these functions that I can get rid of, or is handled by newer code?

--

Qualcomm Innovation Center, Inc.

The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project

From:  <mailto:scallanan at apple.com> scallanan at apple.com [ <mailto:scallanan at apple.com> mailto:scallanan at apple.com] 
Sent: Tuesday, February 23, 2016 6:38 PM
To: Ted Woodward
Cc: LLDB
Subject: Re: r260768 (Removed many JIT workarounds from IRForTarget) broke expression parser with function on Hexagon

At that point, I’d set a watchpoint and see what is setting it. I would expect that

  %call = call i32 @factorial(i32 5)

would put a normal value in call, which would then be the value stored by the store instruction

  store i32 %call, i32* @"_ZZ12$__lldb_exprPvE19$__lldb_expr_result", align 4

The store instruction should not have an operand of function type… right?

Sean

On Feb 23, 2016, at 4:21 PM, Ted Woodward < <mailto:ted.woodward at codeaurora.org> ted.woodward at codeaurora.org> wrote:

Unfortunately, that leads to another error, in the Instruction::Store case. IRInterpreter::ResolveConstantValue() returns an error because it doesn’t like the value id of FunctionVal. “Interpreter couldn't resolve a value during execution”.

If I go back 1 commit from r260768 (r260767), it works. The value id is 0xa, ConstantIntVal. So something in r260768 is either setting the value id to 0x5, or keeping the value id from being set to 0xa.

Ted

--

Qualcomm Innovation Center, Inc.

The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project

From:  <mailto:scallanan at apple.com> scallanan at apple.com [ <mailto:scallanan at apple.com> mailto:scallanan at apple.com] 
Sent: Tuesday, February 23, 2016 5:41 PM
To: Ted Woodward
Cc: LLDB
Subject: Re: r260768 (Removed many JIT workarounds from IRForTarget) broke expression parser with function on Hexagon

Ted,

I’m not sure who inside Clang actually sets the value ID – it’s the code generator’s job to make IR, we don’t construct it.

I would be fine with adding FunctionVal to the switch in CanResolveConstant, returning true.

Sean

On Feb 23, 2016, at 3:28 PM, Ted Woodward < <mailto:ted.woodward at codeaurora.org> ted.woodward at codeaurora.org> wrote:

Background: Hexagon clang doesn’t have JIT support, so lldb for Hexagon only uses the IR Interpreter (Codeplay wrote it for us).

Sean, r260768 broke the expression parser with functions.

Without connecting to a target, I can’t get the info for main:

(lldb) e main

error: Can't run the expression locally: Interpreter doesn't handle one of the expression's operands

Connected to a target, I can’t run a function:

(lldb) e factorial(5)

error: Can't run the expression locally: Interpreter doesn't handle one of the expression's operands

I’ve traced the failure to the call to CanResolveConstant() in IRInterpreter::CanIntepret(). The failure happens on the 2nd operand. In the working case, the Value ID is 0xa – Value::ConstantExprVal. In the failing case, it is 0x5. Since it’s defined in a .def file, I can’t be sure, but my guess is its Value::FunctionVal.

Where is the Value ID set?

Some info from the expr log:

; Function Attrs: nounwind

define void @"_Z12$__lldb_exprPv"(i8* %"$__lldb_arg") #0 {

entry:

  %"$__lldb_arg.addr" = alloca i8*, align 4, !clang.decl.ptr !4

  store i8* %"$__lldb_arg", i8** %"$__lldb_arg.addr", align 4

  %0 = load i8, i8* @"_ZGVZ12$__lldb_exprPvE19$__lldb_expr_result", align 1

  %guard.uninitialized = icmp eq i8 %0, 0

  br i1 %guard.uninitialized, label %init.check, label %init.end

init.check:                                       ; preds = %entry

  %call = call i32 @factorial(i32 5)

  store i32 %call, i32* @"_ZZ12$__lldb_exprPvE19$__lldb_expr_result", align 4

  store i8 1, i8* @"_ZGVZ12$__lldb_exprPvE19$__lldb_expr_result", align 1

  br label %init.end

init.end:                                         ; preds = %init.check, %entry

  ret void

}

Unsupported constant: declare i32 @factorial(i32) #1

Ted

--

Qualcomm Innovation Center, Inc.

The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum, a Linux Foundation Collaborative Project

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/lldb-dev/attachments/20160224/bdd42d4c/attachment-0001.html>