First draft of string instructions

kenardi · May 24, 2018 · f62a273 · f62a273
1 parent c1f6c4a
commit f62a273
Show file tree

Hide file tree

Showing 8 changed files with 101 additions and 43 deletions.
diff --git a/emu/cpu.h b/emu/cpu.h
@@ -73,6 +73,7 @@ struct cpu_state {
         };
         // for asm
 #define AF_FLAG (1 << 4)
+#define DF_FLAG (1 << 10)
     };
     // for maximum efficiency these are stored in bytes
     byte_t cf;

diff --git a/emu/decode.h b/emu/decode.h
@@ -459,17 +459,17 @@ __no_instrument DECODER_RET glue(DECODER_NAME, OP_SIZE)(DECODER_ARGS) {
                    READADDR; MOV(reg_a, mem_addr,8); break;
         case 0xa3: TRACEI("mov oax, mem\t");
                    READADDR; MOV(reg_a, mem_addr,oz); break;
-        case 0xa4: TRACEI("movsb"); MOVS(8); break;
-        case 0xa5: TRACEI("movs"); MOVS(OP_SIZE); break;
+        case 0xa4: TRACEI("movsb"); STR(movs, 8); break;
+        case 0xa5: TRACEI("movs"); STR(movs, oz); break;
 
         case 0xa8: TRACEI("test imm8, al");
                    READIMM8; TEST(imm, reg_a,8); break;
         case 0xa9: TRACEI("test imm, oax");
                    READIMM; TEST(imm, reg_a,oz); break;
 
-        case 0xaa: TRACEI("stosb"); STOS(8); break;
-        case 0xab: TRACEI("stos"); STOS(OP_SIZE); break;
-        case 0xac: TRACEI("lodsb"); LODS(8); break;
+        case 0xaa: TRACEI("stosb"); STR(stos, 8); break;
+        case 0xab: TRACEI("stos"); STR(stos, oz); break;
+        case 0xac: TRACEI("lodsb"); STR(lods, 8); break;
 
         case 0xb0: TRACEI("mov imm, al\t");
                    READIMM8; MOV(imm, reg_a,8); break;
@@ -633,7 +633,7 @@ __no_instrument DECODER_RET glue(DECODER_NAME, OP_SIZE)(DECODER_ARGS) {
                         default: TRACE("undefined"); UNDEFINED;
                     }
                     break;
-                case 0xae: TRACEI("repnz scasb"); REPNZ(SCAS(8)); break;
+                case 0xae: TRACEI("repnz scasb"); REPNZ(scas, 8); break;
                 default: TRACE("undefined"); UNDEFINED;
             }
             break;
@@ -654,13 +654,13 @@ __no_instrument DECODER_RET glue(DECODER_NAME, OP_SIZE)(DECODER_ARGS) {
 
                 case 0x90: TRACEI("pause"); break;
 
-                case 0xa4: TRACEI("rep movsb"); REP(MOVS(8)); break;
-                case 0xa5: TRACEI("rep movs"); REP(MOVS(OP_SIZE)); break;
+                case 0xa4: TRACEI("rep movsb"); REP(movs, 8); break;
+                case 0xa5: TRACEI("rep movs"); REP(movs, oz); break;
 
-                case 0xa6: TRACEI("repz cmpsb"); REPZ(CMPS(8)); break;
+                case 0xa6: TRACEI("repz cmpsb"); REPZ(cmps, 8); break;
 
-                case 0xaa: TRACEI("rep stosb"); REP(STOS(8)); break;
-                case 0xab: TRACEI("rep stos"); REP(STOS(OP_SIZE)); break;
+                case 0xaa: TRACEI("rep stosb"); REP(stos, 8); break;
+                case 0xab: TRACEI("rep stos"); REP(stos, oz); break;
 
                 // repz ret is equivalent to ret but on some amd chips there's
                 // a branch prediction penalty if the target of a branch is a

diff --git a/emu/gadgets-x86/gadgets.h b/emu/gadgets-x86/gadgets.h
@@ -42,9 +42,9 @@
         cmpl TLB_ENTRY_page_if_writable(%_tlb,%r14), %r15d
     .endif
     movl %r15d, -TLB_entries+TLB_dirty_page(%_tlb)
-    je 1f
+    je 2f
     call handle_\type\()_miss
-1:
+2:
     addq TLB_ENTRY_data_minus_addr(%_tlb,%r14), %_addrq
 .endm
 
@@ -80,15 +80,19 @@
     .popsection
 .endm
 
-.macro .gadget_array type
+.macro .gadget_list_size type, list:vararg
     _gadget_array_start \type
         # sync with enum size
-        gadgets \type\()8, GADGET_LIST
-        gadgets \type\()16, GADGET_LIST
-        gadgets \type\()32, GADGET_LIST
+        gadgets \type\()8, \list
+        gadgets \type\()16, \list
+        gadgets \type\()32, \list
     .popsection
 .endm
 
+.macro .gadget_array type
+    .gadget_list_size \type, GADGET_LIST
+.endm
+
 .macro _invoke reg, post, macro:vararg
     \macro reg_\reg, e\reg\post
 .endm

diff --git a/emu/gadgets-x86/string.S b/emu/gadgets-x86/string.S
@@ -0,0 +1,50 @@
+#include "gadgets.h"
+
+#define REP_LIST once,rep,repnz
+
+.gadget cld
+    andl $~DF_FLAG, CPU_eflags(%_cpu)
+    gret
+
+.macro do_strop op, size, rep
+    # repnz is only a thing for cmps and scas
+    .ifc \rep,repnz
+        .ifnc \rep,cmps; .ifnc \rep,scas
+            .exitm
+        .endif; .endif
+    .endif
+
+    .gadget \op\size\()_\rep
+    1:
+        .ifc \op,stos
+            movl %edi, %_addr
+            write_prep
+            .if \size == 8
+                movb %al, (%_addrq)
+            .elseif \size == 16
+                movw %ax, (%_addrq)
+            .elseif \size == 32
+                movl %eax, (%_addrq)
+            .endif
+            addl $(\size/8), %edi
+        .endif
+        .ifc \rep,rep
+            decl %ecx
+            jnz 1b
+        .endif
+        gret
+.endm
+
+.irp op, stos
+    .irp size, 8,16,32
+        .irp rep, REP_LIST
+            do_strop \op, \size, \rep
+        .endr
+    .endr
+    .gadget_list_size \op, REP_LIST
+.endr
+# temporary
+.gadget_list_size movs, REP_LIST
+.gadget_list_size lods, REP_LIST
+.gadget_list_size scas, REP_LIST
+.gadget_list_size cmps, REP_LIST
diff --git a/emu/gen.c b/emu/gen.c
@@ -23,6 +23,12 @@ enum cond {
     cond_count,
 };
 
+enum repeat {
+    rep_once, rep_repz, rep_repnz,
+    rep_count,
+    rep_rep = rep_repz,
+};
+
 typedef void (*gadget_t)();
 
 #define GEN(thing) gen(state, (unsigned long) (thing))
@@ -151,7 +157,7 @@ static inline void gen_op(struct gen_state *state, gadget_t *gadgets, enum arg a
 #define PUSHF() UNDEFINED
 #define POPF() UNDEFINED
 #define SAHF UNDEFINED
-#define CLD UNDEFINED
+#define CLD g(cld)
 #define STD UNDEFINED
 
 #define MUL18(val,z) UNDEFINED
@@ -185,14 +191,11 @@ static inline void gen_op(struct gen_state *state, gadget_t *gadgets, enum arg a
 
 #define BSWAP(dst) UNDEFINED
 
-#define SCAS(z) UNDEFINED
-#define MOVS(z) UNDEFINED
-#define LODS(z) UNDEFINED
-#define STOS(z) UNDEFINED
-#define CMPS(z) UNDEFINED
-#define REP(op) UNDEFINED
-#define REPZ(op) UNDEFINED
-#define REPNZ(op) UNDEFINED
+#define strop(op, rep, z) ga(op, sz(z) * size_count + rep_##rep)
+#define STR(op, z) strop(op, once, z)
+#define REP(op, z) strop(op, rep, z)
+#define REPZ(op, z) strop(op, repz, z)
+#define REPNZ(op, z) strop(op, repnz, z)
 
 #define CMPXCHG(src, dst,z) UNDEFINED
 #define XADD(src, dst,z) UNDEFINED

diff --git a/emu/interp.c b/emu/interp.c
@@ -428,42 +428,40 @@ static bool modrm_compute(struct cpu_state *cpu, struct tlb *tlb, addr_t *addr_o
 #define BUMP_SI_DI(size) \
     BUMP_SI(size); BUMP_DI(size)
 
-#define MOVS(z) \
+#define str_movs(z) \
     mem_write(cpu->edi, mem_read(cpu->esi, z), z); \
     BUMP_SI_DI(z)
-
-#define STOS(z) \
+#define str_stos(z) \
     mem_write(cpu->edi, get(reg_a,z),z); \
     BUMP_DI(z)
-
-#define LODS(z) \
+#define str_lods(z) \
     set(reg_a, mem_read(cpu->esi, z),z); \
     BUMP_SI(z)
-
-#define SCAS(z) \
+#define str_scas(z) \
     CMP(reg_a, mem_di,z); \
     BUMP_DI(z)
-
-#define CMPS(z) \
+#define str_cmps(z) \
     CMP(mem_di, mem_si,z); \
     BUMP_SI_DI(z)
 
-#define REP(OP) \
+#define STR(op, z) str_##op(z)
+
+#define REP(op, z) \
     while (cpu->ecx != 0) { \
-        OP; \
+        STR(op, z); \
         cpu->ecx--; \
     }
 
-#define REPNZ(OP) \
+#define REPNZ(op, z) \
     while (cpu->ecx != 0) { \
-        OP; \
+        STR(op, z); \
         cpu->ecx--; \
         if (ZF) break; \
     }
 
-#define REPZ(OP) \
+#define REPZ(op, z) \
     while (cpu->ecx != 0) { \
-        OP; \
+        STR(op, z); \
         cpu->ecx--; \
         if (!ZF) break; \
     }

diff --git a/emu/offsets.c b/emu/offsets.c
@@ -33,6 +33,7 @@ void cpu() {
     MACRO(SF_RES);
     MACRO(AF_OPS);
     MACRO(AF_FLAG);
+    MACRO(DF_FLAG);
 
     OFFSET(JIT_BLOCK, jit_block, code);
 

diff --git a/meson.build b/meson.build
@@ -93,10 +93,11 @@ if get_option('jit')
         'emu/jit.c',
         'emu/gen.c',
         'emu/gadgets-x86/entry.S',
+        'emu/gadgets-x86/memory.S',
+        'emu/gadgets-x86/control.S',
         'emu/gadgets-x86/math.S',
         'emu/gadgets-x86/bits.S',
-        'emu/gadgets-x86/control.S',
-        'emu/gadgets-x86/memory.S',
+        'emu/gadgets-x86/string.S',
         'emu/gadgets-x86/misc.S',
         offsets,
     ]