Message ID | 20181210152850.435-1-peter.maydell@linaro.org |
---|---|
State | Superseded |
Headers | show |
Series | target/arm: Emit barriers for A32/T32 load-acquire/store-release insns | expand |
Peter Maydell <peter.maydell@linaro.org> writes: > Now that MTTCG is here, the comment in the 32-bit Arm decoder that > "Since the emulation does not have barriers, the acquire/release > semantics need no special handling" is no longer true. Emit the > correct barriers for the load-acquire/store-release insns, as > we already do in the A64 decoder. > > Signed-off-by: Peter Maydell <peter.maydell@linaro.org> > --- > I've not run into this in practice, and there's very little > aarch32 code out there that is built to use the new-in-v8 > instructions as far as I'm aware, but since it would be very > painful to track down this bug if we ran into it in the > wild it seems worth fixing... There should be a patch in reply to this that ports my barrier tests to a linux-user tcg test case. However I've been unable to get the "mp" test to fail under translation, even on aarch64 hosts (ThunderX/qemu-test) which are nominally out of order. x86 hosts are pretty forgiving anyway. The SynQuacer being A53 based is stubbornly in-order so won't fail anyway. If you have access to any fancy OoO AArch32 hardware to confirm the test is good please let me know. Anyway I can at least confirm that when running the mp_acqrel test case that the correct barrier ops are emitted. All the barrier tests pass so: Tested-by: Alex Bennée <alex.bennee@linaro.org> [by inspection] Reviewed-by: Alex Bennée <alex.bennee@linaro.org> > --- > target/arm/translate.c | 33 ++++++++++++++++++++++++++------- > 1 file changed, 26 insertions(+), 7 deletions(-) > > diff --git a/target/arm/translate.c b/target/arm/translate.c > index 7c4675ffd8a..e8fd824f3f5 100644 > --- a/target/arm/translate.c > +++ b/target/arm/translate.c > @@ -9733,6 +9733,8 @@ static void disas_arm_insn(DisasContext *s, unsigned int insn) > rd = (insn >> 12) & 0xf; > if (insn & (1 << 23)) { > /* load/store exclusive */ > + bool is_ld = extract32(insn, 20, 1); > + bool is_lasr = !extract32(insn, 8, 1); > int op2 = (insn >> 8) & 3; > op1 = (insn >> 21) & 0x3; > > @@ -9760,11 +9762,12 @@ static void disas_arm_insn(DisasContext *s, unsigned int insn) > addr = tcg_temp_local_new_i32(); > load_reg_var(s, addr, rn); > > - /* Since the emulation does not have barriers, > - the acquire/release semantics need no special > - handling */ > + if (is_lasr && !is_ld) { > + tcg_gen_mb(TCG_MO_ALL | TCG_BAR_STRL); > + } > + > if (op2 == 0) { > - if (insn & (1 << 20)) { > + if (is_ld) { > tmp = tcg_temp_new_i32(); > switch (op1) { > case 0: /* lda */ > @@ -9810,7 +9813,7 @@ static void disas_arm_insn(DisasContext *s, unsigned int insn) > } > tcg_temp_free_i32(tmp); > } > - } else if (insn & (1 << 20)) { > + } else if (is_ld) { > switch (op1) { > case 0: /* ldrex */ > gen_load_exclusive(s, rd, 15, addr, 2); > @@ -9847,6 +9850,10 @@ static void disas_arm_insn(DisasContext *s, unsigned int insn) > } > } > tcg_temp_free_i32(addr); > + > + if (is_lasr && is_ld) { > + tcg_gen_mb(TCG_MO_ALL | TCG_BAR_LDAQ); > + } > } else if ((insn & 0x00300f00) == 0) { > /* 0bcccc_0001_0x00_xxxx_xxxx_0000_1001_xxxx > * - SWP, SWPB > @@ -10862,6 +10869,8 @@ static void disas_thumb2_insn(DisasContext *s, uint32_t insn) > tcg_gen_addi_i32(tmp, tmp, s->pc); > store_reg(s, 15, tmp); > } else { > + bool is_lasr = false; > + bool is_ld = extract32(insn, 20, 1); > int op2 = (insn >> 6) & 0x3; > op = (insn >> 4) & 0x3; > switch (op2) { > @@ -10883,12 +10892,18 @@ static void disas_thumb2_insn(DisasContext *s, uint32_t insn) > case 3: > /* Load-acquire/store-release exclusive */ > ARCH(8); > + is_lasr = true; > break; > } > + > + if (is_lasr && !is_ld) { > + tcg_gen_mb(TCG_MO_ALL | TCG_BAR_STRL); > + } > + > addr = tcg_temp_local_new_i32(); > load_reg_var(s, addr, rn); > if (!(op2 & 1)) { > - if (insn & (1 << 20)) { > + if (is_ld) { > tmp = tcg_temp_new_i32(); > switch (op) { > case 0: /* ldab */ > @@ -10927,12 +10942,16 @@ static void disas_thumb2_insn(DisasContext *s, uint32_t insn) > } > tcg_temp_free_i32(tmp); > } > - } else if (insn & (1 << 20)) { > + } else if (is_ld) { > gen_load_exclusive(s, rs, rd, addr, op); > } else { > gen_store_exclusive(s, rm, rs, rd, addr, op); > } > tcg_temp_free_i32(addr); > + > + if (is_lasr && is_ld) { > + tcg_gen_mb(TCG_MO_ALL | TCG_BAR_LDAQ); > + } > } > } else { > /* Load/store multiple, RFE, SRS. */ -- Alex Bennée
diff --git a/target/arm/translate.c b/target/arm/translate.c index 7c4675ffd8a..e8fd824f3f5 100644 --- a/target/arm/translate.c +++ b/target/arm/translate.c @@ -9733,6 +9733,8 @@ static void disas_arm_insn(DisasContext *s, unsigned int insn) rd = (insn >> 12) & 0xf; if (insn & (1 << 23)) { /* load/store exclusive */ + bool is_ld = extract32(insn, 20, 1); + bool is_lasr = !extract32(insn, 8, 1); int op2 = (insn >> 8) & 3; op1 = (insn >> 21) & 0x3; @@ -9760,11 +9762,12 @@ static void disas_arm_insn(DisasContext *s, unsigned int insn) addr = tcg_temp_local_new_i32(); load_reg_var(s, addr, rn); - /* Since the emulation does not have barriers, - the acquire/release semantics need no special - handling */ + if (is_lasr && !is_ld) { + tcg_gen_mb(TCG_MO_ALL | TCG_BAR_STRL); + } + if (op2 == 0) { - if (insn & (1 << 20)) { + if (is_ld) { tmp = tcg_temp_new_i32(); switch (op1) { case 0: /* lda */ @@ -9810,7 +9813,7 @@ static void disas_arm_insn(DisasContext *s, unsigned int insn) } tcg_temp_free_i32(tmp); } - } else if (insn & (1 << 20)) { + } else if (is_ld) { switch (op1) { case 0: /* ldrex */ gen_load_exclusive(s, rd, 15, addr, 2); @@ -9847,6 +9850,10 @@ static void disas_arm_insn(DisasContext *s, unsigned int insn) } } tcg_temp_free_i32(addr); + + if (is_lasr && is_ld) { + tcg_gen_mb(TCG_MO_ALL | TCG_BAR_LDAQ); + } } else if ((insn & 0x00300f00) == 0) { /* 0bcccc_0001_0x00_xxxx_xxxx_0000_1001_xxxx * - SWP, SWPB @@ -10862,6 +10869,8 @@ static void disas_thumb2_insn(DisasContext *s, uint32_t insn) tcg_gen_addi_i32(tmp, tmp, s->pc); store_reg(s, 15, tmp); } else { + bool is_lasr = false; + bool is_ld = extract32(insn, 20, 1); int op2 = (insn >> 6) & 0x3; op = (insn >> 4) & 0x3; switch (op2) { @@ -10883,12 +10892,18 @@ static void disas_thumb2_insn(DisasContext *s, uint32_t insn) case 3: /* Load-acquire/store-release exclusive */ ARCH(8); + is_lasr = true; break; } + + if (is_lasr && !is_ld) { + tcg_gen_mb(TCG_MO_ALL | TCG_BAR_STRL); + } + addr = tcg_temp_local_new_i32(); load_reg_var(s, addr, rn); if (!(op2 & 1)) { - if (insn & (1 << 20)) { + if (is_ld) { tmp = tcg_temp_new_i32(); switch (op) { case 0: /* ldab */ @@ -10927,12 +10942,16 @@ static void disas_thumb2_insn(DisasContext *s, uint32_t insn) } tcg_temp_free_i32(tmp); } - } else if (insn & (1 << 20)) { + } else if (is_ld) { gen_load_exclusive(s, rs, rd, addr, op); } else { gen_store_exclusive(s, rm, rs, rd, addr, op); } tcg_temp_free_i32(addr); + + if (is_lasr && is_ld) { + tcg_gen_mb(TCG_MO_ALL | TCG_BAR_LDAQ); + } } } else { /* Load/store multiple, RFE, SRS. */
Now that MTTCG is here, the comment in the 32-bit Arm decoder that "Since the emulation does not have barriers, the acquire/release semantics need no special handling" is no longer true. Emit the correct barriers for the load-acquire/store-release insns, as we already do in the A64 decoder. Signed-off-by: Peter Maydell <peter.maydell@linaro.org> --- I've not run into this in practice, and there's very little aarch32 code out there that is built to use the new-in-v8 instructions as far as I'm aware, but since it would be very painful to track down this bug if we ran into it in the wild it seems worth fixing... --- target/arm/translate.c | 33 ++++++++++++++++++++++++++------- 1 file changed, 26 insertions(+), 7 deletions(-) -- 2.19.2