[Bug 295431] CURRENT compiler miscompiles lang/mit-scheme

https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=295431

            Bug ID: 295431
           Summary: CURRENT compiler miscompiles lang/mit-scheme
           Product: Base System
           Version: 16.0-CURRENT
          Hardware: amd64
                OS: Any
            Status: New
          Severity: Affects Some People
          Priority: ---
         Component: bin
          Assignee: bugs@FreeBSD.org
          Reporter: jfc@mit.edu

Created attachment 270915
  --> https://bugs.freebsd.org/bugzilla/attachment.cgi?id=270915&action=edit
llvm IR

The system cc compiler on CURRENT miscompiles lang/mit-scheme for amd64.  The
attached LLVM IR file is from src/microcode/cmpint.c.

Compile it -O2 -S -fno-strict-aliasing -march=x86-64-v3.  The assembly output
for function guarantee_interp_return contains

        movabsq $7493989779944505344, %rcx      # imm = 0x6800000000000000
        orq     %rax, %rcx
        movq    $0, stack_pointer(%rip)
        movq    %rcx, 0

The last line is a store to address 0, causing a runtime crash.  The previous
line is also incorrect, causing a store of NULL to an interpreter stack pointer
that should never be NULL.  That NULL value is used for the store address on
the last line.

Usually -fno-strict-aliasing prevents miscompilation but this time it does the
opposite.

FreeBSD clang version 21.1.8 (https://github.com/llvm/llvm-project.git
llvmorg-21.1.8-0-g2078da43e25a)

devel/llvm21 has the same behavior.  The IR is incompatible with llvm19 due to
the new "capture" keyword.  That keyword is used for alias analysis.  A
possible clue.

It's conceivable that the MIT Scheme code has undefined behavior other than
aliasing.  I have not had time to trace the optimization passes to see where
the bad code came in.

-- 
You are receiving this mail because:
You are the assignee for the bug.