Support CST_CODE_INLINEASM constant codes added in LLVM 13 and 14

[RyanGlScott]

Prior to LLVM 13, there were two constant codes (18 and 23) for inline assembly statements. There are now two more such codes as of LLVM 14:

• Code 28 (now named CST_CODE_INLINEASM_OLD3), added in LLVM 13 in llvm/llvm-project@8ec9fd4
• Code 30 (now named CST_CODE_INLINEASM), added in LLVM 14 in llvm/llvm-project@62b1682

Currently, llvm-pretty-bc-parser does not support either of these. Here is a test case:

int main() {
  asm("nop");
  return 0;
}

When compiled with ~/Software/clang+llvm-13.0.1/bin/clang test.c -c -emit-llvm, this produces the following parse error:

λ> parseBitCodeFromFile "test.bc"
Left (Error {errContext = ["CONSTANTS_BLOCK","@main","FUNCTION_BLOCK","FUNCTION_BLOCK_ID","value symbol table","MODULE_BLOCK","Bitstream"], errMessage = "Unknown constant record code 28\nAre you sure you're using a supported compiler?\nCheck here: https://github.com/GaloisInc/llvm-pretty-bc-parser\n"})

The error is much the same when compiled with LLVM 14, except that the code will be 30 instead of 28.

With some effort, we should be able to adapt the existing llvm-pretty-bc-parser code for code 18 (now named CST_CODE_INLINEASM_OLD as of LLVM 14):

llvm-pretty-bc-parser/src/Data/LLVM/BitCode/IR/Constants.hs

Lines 360 to 373 in daffc92

	18 -> label "CST_CODE_INLINEASM_OLD" $ do
	let field = parseField r
	ty <- getTy
	flags <- field 0 numeric
	let sideEffect = testBit (flags :: Int) 0
	alignStack = (flags `shiftR` 1) == 1
	alen <- field 1 numeric
	asm <- UTF8.decode <$> parseSlice r 2 alen char
	clen <- field (2+alen) numeric
	cst <- UTF8.decode <$> parseSlice r (3+alen) clen char
	return (getTy, Typed ty (ValAsm sideEffect alignStack asm cst):cs)

And for code 23 (now named CST_CODE_INLINEASM_OLD2 as of LLVM 14):

llvm-pretty-bc-parser/src/Data/LLVM/BitCode/IR/Constants.hs

Lines 416 to 437 in daffc92

	23 -> label "CST_CODE_INLINEASM" $ do
	let field = parseField r
	mask <- field 0 numeric
	let test = testBit (mask :: Word32)
	hasSideEffects = test 0
	isAlignStack = test 1
	_asmDialect = mask `shiftR` 2
	asmStrSize <- field 1 numeric
	Assert.recordSizeGreater r (1 + asmStrSize)
	constStrSize <- field (2 + asmStrSize) numeric
	Assert.recordSizeGreater r (2 + asmStrSize + constStrSize)
	asmStr <- fmap UTF8.decode $ parseSlice r 2 asmStrSize char
	constStr <- fmap UTF8.decode $ parseSlice r (3 + asmStrSize) constStrSize char
	ty <- getTy
	let val = ValAsm hasSideEffects isAlignStack asmStr constStr
	return (getTy, Typed ty val : cs)

[RyanGlScott]

@eddywestbrook also notes that this bug is triggered when compiling the following program with rustc-1.60.0 (bundled with LLVM 14):

/* A function that immediately panics */
pub fn get_out () -> ! {
    panic!("Uh oh!")
}

This is because the generated LLVM bitcode will use inline assembly, somewhat surprisingly:

; Function Attrs: inlinehint uwtable
define void @_ZN4core4hint9black_box17haf534fee2d513d5fE() unnamed_addr #0 {
start:
  call void asm sideeffect "", "r,~{memory}"({}* undef), !srcloc !7
  br label %bb1

bb1:                                              ; preds = %start
  ret void
}

[RyanGlScott]

Note that:

• Code 28 (CST_CODE_INLINEASM_OLD3) adds an unwind flag
• Code 30 (CST_CODE_INLINEASM) adds a function type

In order to support this new information, we will need to augment the ValAsm data constructor in llvm-pretty with additional fields.

[langston-barrett]

These constants also appear when compiling an x86_64 "tinyconfig" Linux kernel 5.17 with LLVM 13 like so. Bitcode: vmlinux-5.17-llvm-13.bc.zip.