update test cases to LLVM 9.0

[mewmew]

Some of the previously valid test cases (taken from the 7.0 release of LLVM) are no longer valid.

Error reported by opt of the LLVM 9.0 release when parsing testdata/llvm/test/Analysis/DominanceFrontier/new_pm_test.ll:

$ opt -S -o foo.ll new_pm_test.ll
opt: new_pm_test.ll:29:1: error: label expected to be numbered '12'
13:
^

Error reported by go test for llir/llvm once updated to handle explicitly named unnamed basic blocks (as done in LLVM 9.0); see #104.

From https://travis-ci.org/llir/llvm/jobs/605020808#L708:

--- FAIL: TestParseFile (40.39s)

    asm_test.go:533: unable to parse "../testdata/llvm/test/Analysis/DominanceFrontier/new_pm_test.ll" into AST;
    invalid local ID in function "@a_linear_impl_fig_1", expected %12, got %13

[mewmew]

Note, to update the test cases from LLVM 7.0 to LLVM 9.0, the following steps were taken.

Update VER in testdata/llvm/Makefile:

 # LLVM version.
-VER=7.0.0
+VER=9.0.0

Run make to download the LLVM 9.0 source release and extract *.ll test cases from it.

From the testdata/llvm directory of the testdata repository.

make

Run skip.sh to skip a few "bad" test cases.

./skip.sh

Commit new test cases to the Git repo.

git add llvm/test
git commit -m 'llvm: update test cases to LLVM 9.0 release'

The remaining step is to update the testdata submodule of the llir/llvm repository, and then run go test github.com/llir/llvm/... to see if any of the golden output changed.

Each test case (e.g. llvm/test/Analysis/DominanceFrontier/new_pm_test.ll) has a corresponding golden output (e.g. llvm/test/Analysis/DominanceFrontier/new_pm_test.ll.golden), since the input may differ from the expected output (in the golden output comments are removed and whitespace may change as compared to the input).

Sometimes the expected output changes when the grammar of LLVM IR is updated in between LLVM releases. Otherwise the expected output should remain stable.

Once the test cases are commited, we enter an iterative process of updating the LLVM IR grammar to incorporate new changes made to the LLVM IR language as part of the release, and then if new concepts are added to the language, then add the respective concepts to llir/llvm/ir. For the LLVM 9.0 release, this corresponds to the tracking issue #101.

[mewmew]

The only test case currently failing is testdata/llvm/test/Assembler/diexpression.ll (syntax error at line 24), because the parser generated from the [ll.tm](!7 = !DIExpression(DW_OP_LLVM_convert, 16, DW_ATE_unsigned, DW_OP_LLVM_convert, 32, DW_ATE_signed)) grammar fails to parse this line:

!7 = !DIExpression(DW_OP_LLVM_convert, 16, DW_ATE_unsigned, DW_OP_LLVM_convert, 32, DW_ATE_signed)

Results from go test ./...

--- FAIL: TestParseFile (34.42s)
    asm_test.go:533: unable to parse "../testdata/llvm/test/Assembler/diexpression.ll" into AST; syntax error at line 24
        unable to parse "../testdata/llvm/test/Assembler/diexpression.ll" into an AST
        github.com/llir/llvm/asm.ParseString
        	/home/u/Desktop/go/src/github.com/llir/llvm/asm/asm.go:58
        github.com/llir/llvm/asm.ParseBytes
        	/home/u/Desktop/go/src/github.com/llir/llvm/asm/asm.go:48
        github.com/llir/llvm/asm.ParseFile
        	/home/u/Desktop/go/src/github.com/llir/llvm/asm/asm.go:29
        github.com/llir/llvm/asm.TestParseFile
        	/home/u/Desktop/go/src/github.com/llir/llvm/asm/asm_test.go:531
        testing.tRunner
        	/home/u/go/src/testing/testing.go:954
        runtime.goexit
        	/home/u/go/src/runtime/asm_amd64.s:1375

[mewmew]

@dannypsnl, feel like taking a look at adding the enums needed to help the parser handle this line? (I think it is DW_OP_LLVM_xxx)

!7 = !DIExpression(DW_OP_LLVM_convert, 16, DW_ATE_unsigned, DW_OP_LLVM_convert, 32, DW_ATE_signed)

[dannypsnl]

I dig into this case, the parser returns error: unable to locate DwarfOp enum corresponding to "DW_OP_LLVM_convert", but it seems like there was a rule for it?

https://github.com/llir/grammar/blob/f1eb80e86b542bd8a463371e8a99e116c61d2364/ll.tm#L107

Update

I think I got the point, but how to generate enum?

[dannypsnl]

By the way, would you like to use official IR Parser by submodule LLVM-IR-parser and mapping ast only? textmapper didn't provide a reasonable result for parsing error. I didn't know which step it stuck.

All I got was: syntax error at line 1

Here is the test code:

package asm

import (
	"testing"

	"github.com/llir/ll/ast"
)

func TestParse(t *testing.T) {
	testCode := `!7 = !DIExpression(DW_OP_LLVM_convert, 16, DW_ATE_unsigned, DW_OP_LLVM_convert, 32, DW_ATE_signed)`
	_, err := ast.Parse("", testCode)
	if err != nil {
		t.Error(err)
	}
}

[mewmew]

Hi @dannypsnl!

Great that you started working on this!

The test case you wrote is good. Unfortunately Textmapper does not show column, but only line offset of errors. I'll submit an issue upstream.

One thing you could do for now to get more useful errors is to split the line into multiple lines (since whitespace tokens are ignored):

!7
=
!DIExpression(DW_OP_LLVM_convert,
16,
DW_ATE_unsigned,
DW_OP_LLVM_convert,
32,
DW_ATE_signed
)

Now, we get the error at "line 5".

$ go test -v
=== RUN   TestParse
    TestParse: aaa_test.go:18: syntax error at line 5

Which corresponds to this line:

DW_ATE_unsigned,

So, we now know it fails to parse Dwarf Attribute Encodings in DIExpressions. Lets look at the update to the C++ parser in LLVM 9.0 to see what changed.

An update was made to ParseDIExpression in LLVM 9.0. It now includes DwarfAttEncoding as a valid field. Previously only DwarfOp and Int were valid.

To see the diff, run:

diff -u llvm-7.0.0.src/lib/AsmParser/ llvm-9.0.0.src/lib/AsmParser/

The relevant part for this change is in lib/AsmParser/LLParser.cpp, more specifically in LLParser::ParseDIExpression, where the following code was added:

+      if (Lex.getKind() == lltok::DwarfAttEncoding) {
+        if (unsigned Op = dwarf::getAttributeEncoding(Lex.getStrVal())) {
+          Lex.Lex();
+          Elements.push_back(Op);
+          continue;
+        }
+        return TokError(Twine("invalid DWARF attribute encoding '") + Lex.getStrVal() + "'");
+      }
+

So, the ll.tm grammar has to be updated accordingly.

From:

DIExpressionField -> DIExpressionField
	: UintLit
	| DwarfOp
;

To:

DIExpressionField -> DIExpressionField
	: UintLit
	| DwarfAttEncoding
	| DwarfOp
;

[mewmew]

The grammar has now been updated llir/grammar@7cadd4c

Edit: so now the test case you provided works, the input LLVM IR is successfully parsed into its AST representation:

$ go test -v
=== RUN   TestParse
--- PASS: TestParse (0.00s)

[mewmew]

Fixed by @dannypsnl in #110.

[mewmew]

The test case you wrote is good. Unfortunately Textmapper does not show column, but only line offset of errors. I'll submit an issue upstream.

Reported upstream at inspirer/textmapper#36