Conversation
| // Preserve an asm statement as an analysis barrier, including for empty | ||
| // blocks. Hide assembler # operands before simplecpp stringifies them. | ||
| std::unique_ptr<simplecpp::Token> open(new simplecpp::Token("(", start->location)); | ||
| std::unique_ptr<simplecpp::Token> close(new simplecpp::Token(")", start->location)); |
There was a problem hiding this comment.
This is an AI review. Take it with a grain of salt and feel free to reject it by resolving the comment.
I built the PR and compared it with main. Other asm forms are unchanged: __asm("..."), __asm volatile(...), MS __asm { ... }, one-line __asm nop __endasm;, and int x __asm("sym");. Unbraced if/do bodies with a block still parse. As a bonus, a { inside an asm ; comment no longer gives a syntax error.
One behavior change I noticed: the assembler text is now dropped. For
void f(void) {
__asm
mov ax,bx
__endasm;
}main produces asm ( "mov ax , bx" ) ; (from Tokenizer::simplifyAsm(), see testtokenize.cpp around line 1191). With this PR, the --debug output and the dump file show asm ( "" ) ;. I could not find any checker or addon that reads the asm string, and #pragma asm already produces an empty asm ( ), so this may be fine. But it means the __endasm branch in Tokenizer::simplifyAsm() is now effectively dead for preprocessed input. Is that intended? If so, maybe remove or document that branch. Otherwise, maybe keep the text here by turning the removed tokens into a string literal between ( and )?
2 participants
Addresses the remaining SDCC examples in Trac 6028.
@dptroperands currently cause a syntax error, and#(s_XINIT >> 8)is altered during preprocessing before the late assembler simplifier runs, leaving an unmatched).Normalize complete literal
__asm/__endasmregions in the raw preprocessor tokens, including newly loaded headers. Preserve an opaqueasm()analysis barrier, exactly one terminator and the locations of surrounding C code. Leave ambiguous/missing/nested regions, real preprocessing directives, macro definitions and other assembler dialects on the existing path. The existing pragma-asm handling is unchanged.Add seven preprocessor test methods covering operands, empty/adjacent blocks, comments, missing/nested markers, other dialects, configurations and unbraced control flow. Two CLI regressions cover the main-file and included-header paths and verify that a real division-by-zero diagnostic after the block remains at its original line.
Validation:
@and#(failures; 18 original/patched CLI comparisons pass, including outside-C diagnostic and GNU/MS/macro controls.git diff --checkpass.One earlier Windows full-suite attempt exited with
0xc0000005while displayingTestType::checkTooBigShift_Unix32; the full rerun above passed. The recorded fault offset resolves to the unchangedToken::Match. A bounded debugger run did not reproduce that failure but was stopped after becoming slow inTestValueFlow::valueFlowHang; its outcome is inconclusive. No root cause or baseline attribution is claimed.This is a bounded fix for the demonstrated literal marker-delimited blocks. It does not repair initial lexer errors, macro-generated markers, or ARM assembler-function syntax.
Please assign Trac 6028 to KiritoYG for this patch and confirm whether this remaining-defect fix qualifies under the published USD 10 bounty bracket. After qualifying closure, I can use the documented bounty-request process; please also confirm the available settlement channel.