Implement atomic bt[csr]

Sekedo664 · Mar 11, 2019 · 8308c5c · 8308c5c
1 parent 9327969
commit 8308c5c
Show file tree

Hide file tree

Showing 5 changed files with 85 additions and 1 deletion.
diff --git a/emu/decode.h b/emu/decode.h
@@ -210,7 +210,7 @@ __no_instrument DECODER_RET glue(DECODER_NAME, OP_SIZE)(DECODER_ARGS) {
                            READMODRM; MOVZX(modrm_val, modrm_reg,16,oz); break;
 
 #define GRP8(bit, val,z) \
-    switch(modrm.opcode) { \
+    switch (modrm.opcode) { \
         case 4: TRACEI("bt"); BT(bit, val,z); break; \
         case 5: TRACEI("bts"); BTS(bit, val,z); break; \
         case 6: TRACEI("btr"); BTR(bit, val,z); break; \
@@ -732,6 +732,24 @@ __no_instrument DECODER_RET glue(DECODER_NAME, OP_SIZE)(DECODER_ARGS) {
                 case 0x0f:
                     READINSN;
                     switch (insn) {
+                        case 0xab: TRACEI("lock bts reg, modrm");
+                                   READMODRM; ATOMIC_BTS(modrm_reg, modrm_val,oz); break;
+                        case 0xb3: TRACEI("lock btr reg, modrm");
+                                   READMODRM; ATOMIC_BTR(modrm_reg, modrm_val,oz); break;
+                        case 0xbb: TRACEI("lock btc reg, modrm");
+                                   READMODRM; ATOMIC_BTC(modrm_reg, modrm_val,oz); break;
+
+#define GRP8_ATOMIC(bit, val,z) \
+    switch (modrm.opcode) { \
+        case 5: TRACEI("bts"); ATOMIC_BTS(bit, val,z); break; \
+        case 6: TRACEI("btr"); ATOMIC_BTR(bit, val,z); break; \
+        case 7: TRACEI("btc"); ATOMIC_BTC(bit, val,z); break; \
+        default: UNDEFINED; \
+    }
+                        case 0xba: TRACEI("lock grp8 imm8, modrm");
+                                   READMODRM; READIMM8; GRP8_ATOMIC(imm, modrm_val,oz); break;
+#undef GRP8_ATOMIC
+
                         case 0xb1: TRACEI("lock cmpxchg reg, modrm");
                                    READMODRM_MEM; ATOMIC_CMPXCHG(modrm_reg, modrm_val,oz); break;
 

diff --git a/emu/interp.c b/emu/interp.c
@@ -559,6 +559,9 @@ static bool modrm_compute(struct cpu_state *cpu, struct tlb *tlb, addr_t *addr_o
 #define ATOMIC_DEC DEC
 #define ATOMIC_CMPXCHG CMPXCHG
 #define ATOMIC_XADD XADD
+#define ATOMIC_BTS BTS
+#define ATOMIC_BTR BTR
+#define ATOMIC_BTC BTC
 #include "emu/interp/sse.h"
 #include "emu/interp/fpu.h"
 

diff --git a/jit/gadgets-aarch64/bits.S b/jit/gadgets-aarch64/bits.S
@@ -226,6 +226,7 @@
         ldr w8, [_xaddr]
         do_bt_op \op, w8, \size, \s
         .ifin(\op, btc,bts,btr)
+            str w8, [_xaddr]
             write_done \size, \op\size\()_mem
         .endifin
         gret 1
@@ -254,6 +255,43 @@
     .gadget_array \op
 .endr
 
+# atomic versions of the above
+
+.macro do_bt_atomic op, size, s
+    .gadget atomic_\op\size\()_mem
+        bic w8, _tmp, 0x1f
+        add _addr, _addr, w8, lsr 3
+        write_prep \size, atomic_\op\size\()_mem
+        # this is simple enough that I'm comfortable doing it with ldaxr/stlxr
+    1:
+        ldaxr w8, [_xaddr]
+        mov w9, 1
+        and _tmp, _tmp, \size-1
+        lsl w9, w9, _tmp
+        tst w8, w9
+        .ifc \op,btc
+            eor w8, w8, w9
+        .else N .ifc \op,bts
+            orr w8, w8, w9
+        .else N .ifc \op,btr
+            bic w8, w8, w9
+        .endif N .endif N .endif
+        cset w9, ne
+        stlxr w10, w8, [_xaddr]
+        cbnz w10, 1b
+        strb w9, [_cpu, CPU_cf]
+        write_done \size, atomic_\op\size\()_mem
+        gret 1
+        write_bullshit \size, atomic_\op\size\()_mem
+.endm
+
+.irp op, btc,bts,btr
+    .irp size, 16,32
+        ss \size, do_bt_atomic, \op
+    .endr
+    .gadget_array atomic_\op
+.endr
+
 .macro x name reg
     .gadget bswap_\name
         rev \reg, \reg

diff --git a/jit/gadgets-x86_64/bits.S b/jit/gadgets-x86_64/bits.S
@@ -113,6 +113,28 @@
     .gadget_array \op
 .endr
 
+# atomic versions of the above
+
+.macro do_bt_atomic op, size, s, ss
+    .gadget atomic_\op\size\()_mem
+        mov\ss %tmp\s, %r14\s
+        andl $~(\size-1), %r14d
+        shrl $3, %r14d
+        addl %r14d, %_addr
+        read_prep \size, atomic_\op\size\()_mem
+        and\ss $(\size-1), %tmp\s
+        lock \op\ss %tmp\s, (%_addrq)
+        setf_c
+        gret 1
+.endm
+
+.irp op, btc,bts,btr
+    .irp size, 16,32
+        ss \size, do_bt_atomic, \op
+    .endr
+    .gadget_array atomic_\op
+.endr
+
 .macro x name reg
     .gadget bswap_\name
         bswap %\reg

diff --git a/jit/gen.c b/jit/gen.c
@@ -331,6 +331,9 @@ void helper_rdtsc(struct cpu_state *cpu);
 #define ATOMIC_DEC(val,z) op(atomic_dec, val, z)
 #define ATOMIC_CMPXCHG(src, dst,z) atomic_op(cmpxchg, src, dst, z)
 #define ATOMIC_XADD(src, dst,z) load(src, z); op(atomic_xadd, dst, z); store(src, z)
+#define ATOMIC_BTC(bit, val,z) lo(atomic_btc, val, bit, z)
+#define ATOMIC_BTS(bit, val,z) lo(atomic_bts, val, bit, z)
+#define ATOMIC_BTR(bit, val,z) lo(atomic_btr, val, bit, z)
 
 // sse
 #define XORP(src, dst) UNDEFINED