[Qemu-devel] [RFC PATCH] exec-all: Translate TCI return addresses backwards too

[Peter Crosthwaite]

I'm trying to debug why TCI doesn't work for the Microblaze example at
http://wiki.qemu.org/Testing. My debug led me to the return addresses
for exceptions being too far forward and this adjustment looked
related so I gave this change a shot, and it works!

I did some further-afield image testing, and it turns out that this
patch fixes TCI for at least several arches. Here are my test results:

ARCH        TCI-WITH-PATCH      TCI-WITHOUT-PATCH       HOST-TCG
microblaze  Y                   N (blank)               Y
cris        Y                   N (K panic)             Y
arm         Y                   N (halfway)             Y
aarch64     N (abort)           N (abort)               Y
SH4         Y                   N (K panic)             Y

The patch gets you to a linux login prompt on MB,CRIS,ARM and SH4
whereas before the patch various crashes occured.

AArch64 has what appears to be a separate issue with TCI. I am use this
for testing:

http://www.bennee.com/~alex/blog/2014/05/09/running-linux-in-qemus-aarch64-system-emulation-mode/

I don't understand this TCI code fully yet, so I doubt my change is
correct, but RFCing incase someone has some theories to help me debug,
or justify the change.

My debug strategy is to run QEMU both with and without TCI and pass:

-d op,exec,in_asm,mmu

then filter logs with:

s/\(exit_tb\).*$/\1/
s/^\(Trace \)[0-9a-fx]*/\1/

The two logs then give you a nice diff between to TCI and host-TCG
run. The first diff I saw was a bad exception return address in the TCI
case:

 IN: PC=20
 rmsr=4300 resr=412 rear=c7fffffc debug=0 imm=fffffffc iflags=2100 fsr=0
 btaken=1 btarget=c0291d28 mode=kernel(saved=kernel) eip=512 ie=0
 r00=00000000 r01=c026ff90 r02=c026de90 r03=deadbeef
 r04=c7ffe38c r05=00001c74 r06=97ffe38c r07=00001c74
 r08=00000000 r09=91fe0000 r10=00000010 r11=c1c8bd94
 r12=c1c8bdc0 r13=c027f680 r14=00000000 r15=c0291d20
-r16=00000000 r17=c0291d3c r18=07ffe38c r19=00000000
+r16=00000000 r17=c0291d44 r18=07ffe38c r19=00000000
 r20=00000000 r21=00000000 r22=00001c70 r23=c026ffac
 r24=00000000 r25=00000000 r26=00000000 r27=00000000
 r28=00000000 r29=01000000 r30=00000380 r31=c02722f8

In Microblaze, PC=20 is the MMU fault exception vector and R17 is
the exception return address. The faulting instruction is at
0xc0291d2c as shown by the host-TCG run (-) but TCI has it at
0xc0291d44 (+). It is a store:

 0xc0291d38:  andi  r22, r22, -4
 0xc0291d3c:  sw    r3, r4, r22
 0xc0291d40:  imm   -15928
 0xc0291d44:  lwi   r8, r0, -10452

Signed-off-by: Peter Crosthwaite <address@hidden>
---
 include/exec/exec-all.h | 4 ----
 1 file changed, 4 deletions(-)

diff --git a/include/exec/exec-all.h b/include/exec/exec-all.h
index a6fce04..31c2405 100644
--- a/include/exec/exec-all.h
+++ b/include/exec/exec-all.h
@@ -308,11 +308,7 @@ extern uintptr_t tci_tb_ptr;
    to indicate the compressed mode; subtracting two works around that.  It
    is also the case that there are no host isas that contain a call insn
    smaller than 4 bytes, so we don't worry about special-casing this.  */
-#if defined(CONFIG_TCG_INTERPRETER)
-# define GETPC_ADJ   0
-#else
 # define GETPC_ADJ   2
-#endif
 
 #define GETPC()  (GETRA() - GETPC_ADJ)
 
-- 
1.9.1