target/microblaze/translate: fix reversed load/store translation
PATCH 1/1 target/microblaze/translate: fix reversed load/store translation
Since 36a9529e60e0 the TCGv returned by compute_ldst_addr_type{a,b} for the address can be a target register or a constant. However do_{load,store} might modify it in the case of a reversed load/store.
Fix this by factoring out the reversed access logic and copying the address in a new TCGv_i32 if it needs to be modified.
--- target/microblaze/translate.c | 56 ++++++++++++++++++++--------------- 1 file changed, 32 insertions(+), 24 deletions(-)
target/microblaze/translate.c +32 -24
@@ -701,30 +701,47 @@ static void gen_alignment_check_ea(DisasContext *dc, TCGv_i64 ea, int rb,static inline MemOp mo_endian (DisasContext *dc) {return dc->cfg->endi ?MO_LE :MO_BE ; } +static void handle_reversed_access (MemOp *mop,TCGv_i32 *addr) +{ +MemOp size = *mop &MO_SIZE ; + +/* +* When doing reverse accesses we need to do two things. +* +* 1. Reverse the address wrt endianness. +* 2. Byteswap the data lanes on the way back into the CPU core. +* +* Use a new TCGv_i32 to modify the address because the current one might be +* a register or a constant. +*/ + +if (size >MO_8 ) { + *mop ^=MO_BSWAP ; + } + +if (size <MO_32 ) { +TCGv_i32 new_addr =tcg_temp_new_i32 (); + +tcg_gen_xori_i32 (new_addr, *addr,3 - size); + *addr = new_addr; + } +} +static bool do_load (DisasContext *dc,int rd,TCGv_i32 addr,MemOp mop,int mem_index,bool rev) { +#ifndef CONFIG_USER_ONLY MemOp size = mop &MO_SIZE ; +#endif mop |=mo_endian (dc); -/* -* When doing reverse accesses we need to do two things. -* -* 1. Reverse the address wrt endianness. -* 2. Byteswap the data lanes on the way back into the CPU core. -*/ if (rev) { -if (size >MO_8 ) { - mop ^=MO_BSWAP ; - } -if (size <MO_32 ) { -tcg_gen_xori_i32 (addr, addr,3 - size); - } +handle_reversed_access (&mop, &addr); }/* * For system mode, enforce alignment if the cpu configuration * requires it. For user-mode, the Linux kernel will have fixed up
@@ -864,27 +881,18 @@ static bool trans_lwx(DisasContext *dc, arg_typea *arg) }static bool do_store (DisasContext *dc,int rd,TCGv_i32 addr,MemOp mop,int mem_index,bool rev) { +#ifndef CONFIG_USER_ONLY MemOp size = mop &MO_SIZE ; +#endif mop |=mo_endian (dc); -/* -* When doing reverse accesses we need to do two things. -* -* 1. Reverse the address wrt endianness. -* 2. Byteswap the data lanes on the way back into the CPU core. -*/ if (rev) { -if (size >MO_8 ) { - mop ^=MO_BSWAP ; - } -if (size <MO_32 ) { -tcg_gen_xori_i32 (addr, addr,3 - size); - } +handle_reversed_access (&mop, &addr); }/* * For system mode, enforce alignment if the cpu configuration * requires it. For user-mode, the Linux kernel will have fixed up
Signature · 1 line
2.53.0
Hello,
Gentle ping. The fix just consist in allocating a new TCG temp for the address if we need to modify it.
Thanks
Luc
Quoted 111 lines · Luc Michel
Since
36a9529e60e0the TCGv returned by compute_ldst_addr_type{a,b} for the address can be a target register or a constant. However do_{load,store} might modify it in the case of a reversed load/store.Fix this by factoring out the reversed access logic and copying the address in a new
TCGv_i32if it needs to be modified.Fixes 36a9529e60e0 (target/microblaze: Simplify compute_ldst_addr_type{a,b})Signed-off-by Luc Michel <[email protected]>--- target/microblaze/translate.c | 56 ++++++++++++++++++++--------------- 1 file changed, 32 insertions(+), 24 deletions(-)target/microblaze/translate.c +32
-24@@ -701,30 +701,47 @@ static void gen_alignment_check_ea(DisasContext *dc, TCGv_i64 ea, int rb,static inline MemOp mo_endian (DisasContext *dc) {return dc->cfg->endi ?MO_LE :MO_BE ; } +static void handle_reversed_access (MemOp *mop,TCGv_i32 *addr) +{ +MemOp size = *mop &MO_SIZE ; + +/* +* When doing reverse accesses we need to do two things. +* +* 1. Reverse the address wrt endianness. +* 2. Byteswap the data lanes on the way back into the CPU core. +* +* Use a new TCGv_i32 to modify the address because the current one might be +* a register or a constant. +*/ + +if (size >MO_8 ) { + *mop ^=MO_BSWAP ; + } + +if (size <MO_32 ) { +TCGv_i32 new_addr =tcg_temp_new_i32 (); + +tcg_gen_xori_i32 (new_addr, *addr,3 - size); + *addr = new_addr; + } +} +static bool do_load (DisasContext *dc,int rd,TCGv_i32 addr,MemOp mop,int mem_index,bool rev) { +#ifndef CONFIG_USER_ONLY MemOp size = mop &MO_SIZE ; +#endif mop |=mo_endian (dc); -/* -* When doing reverse accesses we need to do two things. -* -* 1. Reverse the address wrt endianness. -* 2. Byteswap the data lanes on the way back into the CPU core. -*/ if (rev) { -if (size >MO_8 ) { - mop ^=MO_BSWAP ; - } -if (size <MO_32 ) { -tcg_gen_xori_i32 (addr, addr,3 - size); - } +handle_reversed_access (&mop, &addr); }/* * For system mode, enforce alignment if the cpu configuration * requires it. For user-mode, the Linux kernel will have fixed up @@ -864,27 +881,18 @@ static bool trans_lwx(DisasContext *dc, arg_typea *arg) }static bool do_store (DisasContext *dc,int rd,TCGv_i32 addr,MemOp mop,int mem_index,bool rev) { +#ifndef CONFIG_USER_ONLY MemOp size = mop &MO_SIZE ; +#endif mop |=mo_endian (dc); -/* -* When doing reverse accesses we need to do two things. -* -* 1. Reverse the address wrt endianness. -* 2. Byteswap the data lanes on the way back into the CPU core. -*/ if (rev) { -if (size >MO_8 ) { - mop ^=MO_BSWAP ; - } -if (size <MO_32 ) { -tcg_gen_xori_i32 (addr, addr,3 - size); - } +handle_reversed_access (&mop, &addr); }/* * For system mode, enforce alignment if the cpu configuration * requires it. For user-mode, the Linux kernel will have fixed up Signature · 1 line
2.53.0
Hi Luc,
Quoted 3 lines · Luc Michel
Since
36a9529e60e0the TCGv returned by compute_ldst_addr_type{a,b} for the address can be a target register or a constant. However do_{load,store} might modify it in the case of a reversed load/store.
Register are modifiable, aren't they? IIUC compute_ldst_addr_typeb() is not used by the reversed opcodes and the only issue is the 0 constant in compute_ldst_addr_typea(). Is that analysis correct?
Quoted 8 lines
Fix this by factoring out the reversed access logic and copying the address in a new
TCGv_i32if it needs to be modified.Fixes 36a9529e60e0 (target/microblaze: Simplify compute_ldst_addr_type{a,b})Signed-off-by Luc Michel <[email protected]>--- target/microblaze/translate.c | 56 ++++++++++++++++++++--------------- 1 file changed, 32 insertions(+), 24 deletions(-)
Quoted 10 lines · Philippe Mathieu-Daudé
Hi Luc,
Quoted 3 lines · Luc Michel
Since
36a9529e60e0the TCGv returned by compute_ldst_addr_type{a,b} for the address can be a target register or a constant. However do_{load,store} might modify it in the case of a reversed load/store.Register are modifiable, aren't they? IIUC
compute_ldst_addr_typeb()is not used by the reversed opcodes and the only issue is the 0 constant incompute_ldst_addr_typea(). Is that analysis correct?
Hi Phil,
The address registers are never modified by a load/store instruction. (See e.g., https://docs.amd.com/r/en-US/ug984-vivado-microblaze-ref/lw )
36a9529e60e0 introduced an optimization that can return the global TCGv used for rA or rB storage instead of copying it.
The bug appears in the reversed load/store case, where the do_load/do_store functions treat the `addr' TCGv_i32 as a temp and modify it. Since 36a9529e60e0, addr can alias rA, rB, or a 0 TCG constant. Those 3 cases are buggy.
Thanks
Luc
Quoted 10 lines
Quoted 9 lines
Fix this by factoring out the reversed access logic and copying the address in a new
TCGv_i32if it needs to be modified.Fixes 36a9529e60e0 (target/microblaze: Simplify compute_ldst_addr_type{a,b})Signed-off-by Luc Michel <[email protected]>--- target/microblaze/translate.c | 56 ++++++++++++++++++++--------------- 1 file changed, 32 insertions(+), 24 deletions(-)
Quoted 12 lines · Luc Michel
Quoted 10 lines · Philippe Mathieu-Daudé
Hi Luc,
Quoted 3 lines · Luc Michel
Since
36a9529e60e0the TCGv returned by compute_ldst_addr_type{a,b} for the address can be a target register or a constant. However do_{load,store} might modify it in the case of a reversed load/store.Register are modifiable, aren't they? IIUC
compute_ldst_addr_typeb()is not used by the reversed opcodes and the only issue is the 0 constant incompute_ldst_addr_typea(). Is that analysis correct?
Quoted 7 lines
Hi Phil,
The address registers are never modified by a load/store instruction. (See e.g., https://docs.amd.com/r/en-US/ug984-vivado-microblaze-ref/lw )
36a9529e60e0introduced an optimization that can return the global TCGv used for rA or rB storage instead of copying it.
Doh indeed we don't want to update these.
Quoted 21 lines
The bug appears in the reversed load/store case, where the do_load/do_store functions treat the `addr'
TCGv_i32as a temp and modify it. Since36a9529e60e0, addr can alias rA, rB, or a 0 TCG constant. Those 3 cases are buggy.Thanks
Luc
Quoted 10 lines
Quoted 9 lines
Fix this by factoring out the reversed access logic and copying the address in a new
TCGv_i32if it needs to be modified.Fixes 36a9529e60e0 (target/microblaze: Simplify compute_ldst_addr_type{a,b})Signed-off-by Luc Michel <[email protected]>--- target/microblaze/translate.c | 56 ++++++++++++++++++++--------------- 1 file changed, 32 insertions(+), 24 deletions(-)
Quoted 9 lines · Luc Michel
Since
36a9529e60e0the TCGv returned by compute_ldst_addr_type{a,b} for the address can be a target register or a constant. However do_{load,store} might modify it in the case of a reversed load/store.Fix this by factoring out the reversed access logic and copying the address in a new
TCGv_i32if it needs to be modified.Fixes 36a9529e60e0 (target/microblaze: Simplify compute_ldst_addr_type{a,b})Signed-off-by Luc Michel <[email protected]>
LGTM!
Quoted 102 lines
--- target/microblaze/translate.c | 56 ++++++++++++++++++++--------------- 1 file changed, 32 insertions(+), 24 deletions(-)target/microblaze/translate.c +32
-24@@ -701,30 +701,47 @@ static void gen_alignment_check_ea(DisasContext *dc, TCGv_i64 ea, int rb,static inline MemOp mo_endian (DisasContext *dc) {return dc->cfg->endi ?MO_LE :MO_BE ; } +static void handle_reversed_access (MemOp *mop,TCGv_i32 *addr) +{ +MemOp size = *mop &MO_SIZE ; + +/* +* When doing reverse accesses we need to do two things. +* +* 1. Reverse the address wrt endianness. +* 2. Byteswap the data lanes on the way back into the CPU core. +* +* Use a new TCGv_i32 to modify the address because the current one might be +* a register or a constant. +*/ + +if (size >MO_8 ) { + *mop ^=MO_BSWAP ; + } + +if (size <MO_32 ) { +TCGv_i32 new_addr =tcg_temp_new_i32 (); + +tcg_gen_xori_i32 (new_addr, *addr,3 - size); + *addr = new_addr; + } +} +static bool do_load (DisasContext *dc,int rd,TCGv_i32 addr,MemOp mop,int mem_index,bool rev) { +#ifndef CONFIG_USER_ONLY MemOp size = mop &MO_SIZE ; +#endif mop |=mo_endian (dc); -/* -* When doing reverse accesses we need to do two things. -* -* 1. Reverse the address wrt endianness. -* 2. Byteswap the data lanes on the way back into the CPU core. -*/ if (rev) { -if (size >MO_8 ) { - mop ^=MO_BSWAP ; - } -if (size <MO_32 ) { -tcg_gen_xori_i32 (addr, addr,3 - size); - } +handle_reversed_access (&mop, &addr); }/* * For system mode, enforce alignment if the cpu configuration * requires it. For user-mode, the Linux kernel will have fixed up @@ -864,27 +881,18 @@ static bool trans_lwx(DisasContext *dc, arg_typea *arg) }static bool do_store (DisasContext *dc,int rd,TCGv_i32 addr,MemOp mop,int mem_index,bool rev) { +#ifndef CONFIG_USER_ONLY MemOp size = mop &MO_SIZE ; +#endif mop |=mo_endian (dc); -/* -* When doing reverse accesses we need to do two things. -* -* 1. Reverse the address wrt endianness. -* 2. Byteswap the data lanes on the way back into the CPU core. -*/ if (rev) { -if (size >MO_8 ) { - mop ^=MO_BSWAP ; - } -if (size <MO_32 ) { -tcg_gen_xori_i32 (addr, addr,3 - size); - } +handle_reversed_access (&mop, &addr); }/* * For system mode, enforce alignment if the cpu configuration * requires it. For user-mode, the Linux kernel will have fixed up Signature · 1 line
2.53.0
Hi Luc,
Quoted 4 lines · Luc Michel
Since
36a9529e60e0the TCGv returned by compute_ldst_addr_type{a,b} for the address can be a target register or a constant. However do_{load,store} might modify it in the case of a reversed load/store.
Quoted 8 lines
Fix this by factoring out the reversed access logic and copying the address in a new
TCGv_i32if it needs to be modified.Fixes 36a9529e60e0 (target/microblaze: Simplify compute_ldst_addr_type{a,b})Signed-off-by Luc Michel <[email protected]>--- target/microblaze/translate.c | 56 ++++++++++++++++++++--------------- 1 file changed, 32 insertions(+), 24 deletions(-)
The tests in our CI didn't catch this problem.
Do you have an image I can use to test your fix?