4.0.0-rc3 crashes with tcg/tcg.c:3952: tcg_gen_code: Assertion `s->gen_insn_end_off[num_insns] == off' failed

I tried to bootstrap and regtested gcc trunk (gcc svn rev 270278, datestamp 20190411) inside my arm64-gentoo installation under qemu-system-aarch64.

Qemu version was 4.0.0-rc3 and -cpu cortex-a57. Qemu configured with only --target-list=aarch64-softmmu,aarch64-linux-user and compiled using gcc "version 5.5.0 20171010 (Ubuntu 5.5.0-12ubuntu1~16.04)".

Executable created from gcc/testsuite/gcc.target/aarch64/advsimd-intrinsics/vldX.c compiled with -O2 crashed the whole qemu-system.

To investigate a bit I also manually run
~/gcc/inst/trunk/bin/gcc ~/gcc/src/trunk/gcc/testsuite/gcc.target/aarch64/advsimd-intrinsics/vldX.c
with different options like:
-O0 -lm -o d0.exe
-O1 -lm -o d1.exe
-O2 -lm -o d2.exe
-O0 -static -lm -o s0.exe
-O1 -static -lm -o s1.exe
-O2 -static -lm -o s2.exe

So, now I have 6 different arm64 executables created with different optimization levels. O0 and O1 versions run ok.
Three sN.exe static executables I've also tried in qemu user mode (with same -cpu), no issue in user mode.

And inside qemu-system I can see that
running "d2.exe" (attached) gives:
tcg/tcg.c:3952: tcg_gen_code: Assertion `s->gen_insn_end_off[num_insns] == off' failed.

And running "s2.exe" gives:
tcg/tcg.c:320: set_jmp_reset_offset: Assertion `s->tb_jmp_reset_offset[which] == off' failed.

It seems like this test is an counter-example for logic that "tcg_ctx->nb_ops < 4000" implies tcg will fit into 16-bit signed size (see tcg_op_buf_full comments).

Richard's changes in abebf92597186 and 9f754620651d were not enough, translation block must be smaller, or we have to find some proper way to bail out when buffer overflows.
I don't know why this situation is not caught by code_gen_highwater logic in tcg.c

I've also tried this "bail out" patch

diff --git a/tcg/tcg.c b/tcg/tcg.c
--- a/tcg/tcg.c
+++ b/tcg/tcg.c
@@ -3949,7 +3949,8 @@ int tcg_gen_code(TCGContext *s, TranslationBlock *tb)
                 size_t off = tcg_current_code_size(s);
                 s->gen_insn_end_off[num_insns] = off;
                 /* Assert that we do not overflow our stored offset.  */
-                assert(s->gen_insn_end_off[num_insns] == off);
+                if (s->gen_insn_end_off[num_insns] != off)
+                  return -1;
             }
             num_insns++;
             for (i = 0; i < TARGET_INSN_START_WORDS; ++i) {

But then running "d2.exe" just hangs the whole qemu-system. It seems that when tcg_gen_code return -1 (like in highwater logic mentioned before), we just re-call it again and again.



Also attaching static-compiled executable "s2.exe".

Returning -1 does not help because all that signals that the buffer is full.
We then flush the buffer and try again, assuming the at the buffer will not fill.
Given that the buffer is usually many megabytes, this is reasonable.

We need something different to signal that the buffer is not full, but that
another offset has overflowed.

Patch set posted:
https://patchwork.ozlabs.org/project/qemu-devel/list/?series=102978



Richard, thank you for solving this so fast!
I certainly can confirm attached executables work fine for me on patched version.

I'll also re-run full gcc regtest a bit later, but it runs for a rather long time, not sure this result will be important next week.

Hopefully, patchset will be included into 4 release.

Unfortunately the fix is too big for this point in the 4.0 release cycle; it'll go into 4.1.


The fix should now be in git master (commits 8b86d6d25807e13a6 and 6e6c4efed995d9ec), so it will be in the 4.1 release.