[clang][integrated-as] "expected absolute expression" diffing symbols

[nickdesaulniers]

.text
start:
  nop
end:
.if (end - start) == 0
.err
.endif

$ clang test-label.s -c
$

.text
.set start, .;
  nop
.set end, .;
.if (end - start) == 0
.err
.endif

$ clang test-symbol.s
test-symbol.s:6:5: error: expected absolute expression
.if (end - start) == 0
    ^
test-symbol.s:7:1: error: .err encountered
.err
^
$ as test-symbol.s 
$

reported via lkml

[llvmbot]

@llvm/issue-subscribers-bug

[nickdesaulniers]

slightly reduced case:

.text
.set start, .
  nop
.set end, .
.if end - start
.err
.endif

From what I can tell, it looks like when recursive evaluating the MCBinaryExpr::Sub in evaluateAsRelocatableImpl, the call to EvaluateSymbolicAdd returns true. MCExpr::evaluateAsAbsolute thinks the MCExpr is relocatable but not absolute.

I would think if . is in the same section, we might be able to compute the difference.

Even

.text

.set start, .
  nop
.set end, .

.ifeq end == start
.err
.endif

[nickdesaulniers]

AttemptToFoldSymbolOffsetDifference in llvm/lib/MC/MCExpr.cpp looks like it might be helpful.

I feel like if the two symbols are in the same section (or at least the same fragment), we should be able to verify that their offsets are the same or different.

See also:

llvm-project/llvm/lib/MC/MCObjectWriter.cpp

Line 52 in ceb5dc5

// On ELF and COFF A - B is absolute if A and B are in the same section.

llvm-project/llvm/lib/MC/MCExpr.cpp

Lines 649 to 655 in ceb5dc5

	// When layout is not finalized, our ability to resolve differences between
	// symbols is limited to specific cases where the fragments between two
	// symbols (including the fragments the symbols are defined in) are
	// fixed-size fragments so the difference can be calculated. For example,
	// this is important when the Subtarget is changed and a new MCDataFragment
	// is created in the case of foo: instr; .arch_extension ext; instr .if . -
	// foo.

[pskocik]

Hi. I seem to be experiencing a relevant issue in that while code such as your test-label.s works from an assembly file
(on both gcc and clang):

.text
start:
  nop
end:
.if (end - start) == 0
.err
.endif

a straightforward re-rendering of it in a C file fails on clang (not gcc) with "expected absolute expression":

__asm(
".text\n"
"start:\n"
"  nop\n"
"end:\n"
".if (end - start) == 0\n"
".err\n"
".endif\n"
);

[musamaanjum]

I've hit the same bug. Clang gives an error with the following extended asm:

<inline asm>:7:8: error: expected absolute expression
    7 |         .ifne . - test_page - 4096
      |               ^
<inline asm>:8:2: error: test page is not one page long
    8 |         .error "test page is not one page long"
      |         ^

asm (
	".pushsection \".text\", \"ax\"\n\t"
	".balign 4096\n\t"
	"test_page: .globl test_page\n\t"
	".fill 4094,1,0xcc\n\t"
	"test_syscall_insn:\n\t"
	"syscall\n\t"
	".ifne . - test_page - 4096\n\t"
	".error \"test page is not one page long\"\n\t"
	".endif\n\t"
	".popsection"
    );

But clang builds fine if the asm is moved to the assembly file, or by generating the asm by GCC and then building the asm to binary by clang.

[MaskRay]

The original report of expression folding in .if for MCObjectStreamer has been fixed https://reviews.llvm.org/D69411 and my followups.

The last two comments are about getUseAssemblerInfoForParsing() and should be discussed at #62520