| Message ID | 20170829063313.10237-4-bobby.prani@gmail.com |
|---|---|
| State | New |
| Headers | show |
| Series | None | expand |
On 08/28/2017 11:33 PM, Pranith Kumar wrote: > Currently, we cannot use mttcg for running strong memory model guests > on weak memory model hosts due to missing ordering semantics. > > We implicitly generate fence instructions for stronger guests if an > ordering mismatch is detected. We generate fences only for the orders > for which fence instructions are necessary, for example a fence is not > necessary between a store and a subsequent load on x86 since its > absence in the guest binary tells that ordering need not be > ensured. Also note that if we find multiple subsequent fence > instructions in the generated IR, we combine them in the TCG > optimization pass. > > This patch allows us to boot an x86 guest on ARM64 hosts using mttcg. > > Signed-off-by: Pranith Kumar <bobby.prani@gmail.com> > --- > tcg/tcg-op.c | 16 ++++++++++++++++ > 1 file changed, 16 insertions(+) Reviewed-by: Richard Henderson <richard.henderson@linaro.org> r~
On Tue, Aug 29, 2017 at 02:33:12 -0400, Pranith Kumar wrote: > Currently, we cannot use mttcg for running strong memory model guests > on weak memory model hosts due to missing ordering semantics. > > We implicitly generate fence instructions for stronger guests if an This confused me. By "We implicitly" are we still talking about the current state (as per the "currently" above?). If not, I'd rephrase as: "We cannot use [...]. To fix it, generate fences [...]" Also, I think you meant s/stronger/weaker/ in the last sentence. > ordering mismatch is detected. We generate fences only for the orders > for which fence instructions are necessary, for example a fence is not > necessary between a store and a subsequent load on x86 since its > absence in the guest binary tells that ordering need not be > ensured. Also note that if we find multiple subsequent fence > instructions in the generated IR, we combine them in the TCG > optimization pass. A before/after example of -d out_asm would be great to have here. > > This patch allows us to boot an x86 guest on ARM64 hosts using mttcg. A test with a simple program that *cannot* work without this patch would be even better. Thanks, Emilio
diff --git a/tcg/tcg-op.c b/tcg/tcg-op.c index 87f673ef49..688d91755b 100644 --- a/tcg/tcg-op.c +++ b/tcg/tcg-op.c @@ -28,6 +28,7 @@ #include "exec/exec-all.h" #include "tcg.h" #include "tcg-op.h" +#include "tcg-mo.h" #include "trace-tcg.h" #include "trace/mem.h" @@ -2662,8 +2663,20 @@ static void gen_ldst_i64(TCGOpcode opc, TCGv_i64 val, TCGv addr, #endif } +static void tcg_gen_req_mo(TCGBar type) +{ +#ifdef TCG_GUEST_DEFAULT_MO + type &= TCG_GUEST_DEFAULT_MO; +#endif + type &= ~TCG_TARGET_DEFAULT_MO; + if (type) { + tcg_gen_mb(type | TCG_BAR_SC); + } +} + void tcg_gen_qemu_ld_i32(TCGv_i32 val, TCGv addr, TCGArg idx, TCGMemOp memop) { + tcg_gen_req_mo(TCG_MO_LD_LD | TCG_MO_ST_LD); memop = tcg_canonicalize_memop(memop, 0, 0); trace_guest_mem_before_tcg(tcg_ctx.cpu, tcg_ctx.tcg_env, addr, trace_mem_get_info(memop, 0)); @@ -2672,6 +2685,7 @@ void tcg_gen_qemu_ld_i32(TCGv_i32 val, TCGv addr, TCGArg idx, TCGMemOp memop) void tcg_gen_qemu_st_i32(TCGv_i32 val, TCGv addr, TCGArg idx, TCGMemOp memop) { + tcg_gen_req_mo(TCG_MO_LD_ST | TCG_MO_ST_ST); memop = tcg_canonicalize_memop(memop, 0, 1); trace_guest_mem_before_tcg(tcg_ctx.cpu, tcg_ctx.tcg_env, addr, trace_mem_get_info(memop, 1)); @@ -2680,6 +2694,7 @@ void tcg_gen_qemu_st_i32(TCGv_i32 val, TCGv addr, TCGArg idx, TCGMemOp memop) void tcg_gen_qemu_ld_i64(TCGv_i64 val, TCGv addr, TCGArg idx, TCGMemOp memop) { + tcg_gen_req_mo(TCG_MO_LD_LD | TCG_MO_ST_LD); if (TCG_TARGET_REG_BITS == 32 && (memop & MO_SIZE) < MO_64) { tcg_gen_qemu_ld_i32(TCGV_LOW(val), addr, idx, memop); if (memop & MO_SIGN) { @@ -2698,6 +2713,7 @@ void tcg_gen_qemu_ld_i64(TCGv_i64 val, TCGv addr, TCGArg idx, TCGMemOp memop) void tcg_gen_qemu_st_i64(TCGv_i64 val, TCGv addr, TCGArg idx, TCGMemOp memop) { + tcg_gen_req_mo(TCG_MO_LD_ST | TCG_MO_ST_ST); if (TCG_TARGET_REG_BITS == 32 && (memop & MO_SIZE) < MO_64) { tcg_gen_qemu_st_i32(TCGV_LOW(val), addr, idx, memop); return;
Currently, we cannot use mttcg for running strong memory model guests on weak memory model hosts due to missing ordering semantics. We implicitly generate fence instructions for stronger guests if an ordering mismatch is detected. We generate fences only for the orders for which fence instructions are necessary, for example a fence is not necessary between a store and a subsequent load on x86 since its absence in the guest binary tells that ordering need not be ensured. Also note that if we find multiple subsequent fence instructions in the generated IR, we combine them in the TCG optimization pass. This patch allows us to boot an x86 guest on ARM64 hosts using mttcg. Signed-off-by: Pranith Kumar <bobby.prani@gmail.com> --- tcg/tcg-op.c | 16 ++++++++++++++++ 1 file changed, 16 insertions(+)