[mips][sanitizer_common] Don't use `ld` in internal_clone() on 32-bit MIPS.

[dsanders]

Summary

On a 32-bit MIPS, the ld instruction does not exist. However, GAS has an ld
macro that expands to a pair of lw instructions which load to a pair of
registers (reg, and reg+1). This macro is not available in the Integrated
Assembler and its use causes -fintegrated-as builds to fail. Even if it were
available, the behaviour on 32-bit MIPS would be incorrect since the current
usage of ld causes the code to clobber $5 (which is supposed to hold
child_stack). It also clobbers $k0 which is reserved for kernel use.

Aside from enabling builds with the integrated assembler, there is no functional
change since internal_clone() is only used by StopTheWorld() which is only used
by 64-bit sanitizers.

[dsanders]

lib/sanitizer_common/sanitizer_linux.cc

943–947

It's not directly related to this patch but: The comment says 'store' but the instruction is a load from the compiler-generated stack frame.

960–971

It's not directly related to this patch but I also noticed that some of the load instructions are loading from the compiler-generated stack frame ($29 is $sp). This is likely to be sensitive to compiler changes and the offsets are likely to differ between ABI's.

[slthakur]

lib/sanitizer_common/sanitizer_linux.cc

943–947

Yes, thanks for catching that, the instruction should have been a store. None of the 32-bit sanitizers use this code as of now but we will still fix it for the sake of completeness.

960–971

Yes. For the 64-bit part the offsets will differ for N32 and N64. We should probably code it like this:

#if _MIPS_SIM == _MIPS_SIM_NABI32
  "lw $25,0($29);\n"
  "lw $4,4($29);\n"
#else
  "ld $25,0($29);\n"
  "ld $4,8($29);\n"
#endif

Thanks again. If this looks good I will submit a patch for fixing these bugs.

[dsanders]

lib/sanitizer_common/sanitizer_linux.cc

960–971

It's also sensitive to compiler version and optimizations at the moment. I think it would be better to let the compiler figure out the address by making the relevant variable (fn?) one of the arguments to the inline assembly statement.

[dsanders]

Could you confirm that this is ok to commit? I think it probably is based on the comments but I don't have an LGTM yet.

[slthakur]

LGTM

[slthakur]

LGTM

lib/sanitizer_common/sanitizer_linux.cc

__asm__ __volatile__(
                         * $a3 = new_tls,
                         * $a4 = child_tidptr)
                         */
                        ".cprestore 16;\n"
                        "move $4,%1;\n"
                        "move $5,%2;\n"
                        "move $6,%3;\n"
                        "move $7,%4;\n"
                        /* Store the fifth argument on stack
                         * if we are using 32-bit abi.
                         */
 #if SANITIZER_WORDSIZE == 32
                        "lw %5,16($29);\n"
 #else
                        "move $8,%5;\n"
 #endif
                        "li $2,%6;\n"
                        "syscall;\n"
 
                        /* if ($v0 != 0)
                         * return;
                         */
                        "bnez $2,1f;\n"
 
                        /* Call "fn(arg)". */
+#if SANITIZER_WORDSIZE == 32
+#ifdef __BIG_ENDIAN__
+                       "lw $25,4($29);\n"
+                       "lw $4,12($29);\n"
+#else
+                       "lw $25,0($29);\n"
+                       "lw $4,8($29);\n"
+#endif
+#else
                        "ld $25,0($29);\n"
                        "ld $4,8($29);\n"
+#endif
                        "jal $25;\n"
 
                        /* Call _exit($v0). */
                        "move $4,$2;\n"
                        "li $2,%7;\n"
                        "syscall;\n"
 
                        /* Return to parent. */