Add instructions needed to assemble sqlite3.

4 files changed, 110 insertions, 15 deletions

diff --git a/asm.peg b/asm.peg
index c00abec..0a80d3a 100644
--- a/asm.peg
+++ b/asm.peg
@@ -1,19 +1,19 @@
 
 line  =
-      ws? s:stmt eol { yy->v = s; }
+      ws? s:stmt { yy->v = s; }
     | eol { yy->v.kind = ASM_BLANK; }
     | . { yy->v.kind = ASM_SYNTAX_ERROR; }
 
-ws = 
-    [ \t]+
-  | "/*" (! "*/" . )* "*/" # XXX multiline comments break our line numbers.
+ws = ([ \t] | comment)+
+
+comment = "/*" (! "*/" . )* "*/"   # XXX multiline comments break our line numbers.
 
 eol = ws? "\n"
 
 stmt =
-    '.' d:directive {$$ = d;}
-  | i:instr { $$ = i; }
-  | l:label {  $$ = l; } 
+    '.' d:directive eol {$$ = d;}
+  | l:label eol {  $$ = l; } 
+  | i:instr eol { $$ = i; }
 
 directive =
    "glob" 'o'? 'l' ws i:ident
@@ -28,6 +28,8 @@ directive =
     { $$.balign = (Balign){.kind = ASM_DIR_BALIGN, .align = n.i64 }; }
   | "byte" ws v:value
     { $$.dirbyte = (Byte){.kind = ASM_DIR_BYTE, .value = v.value }; }
+  | "short" ws v:value
+    { $$.dirshort = (Short){.kind = ASM_DIR_BYTE, .value = v.value }; }
   | "int" ws v:value
     { $$.dirint = (Int){.kind = ASM_DIR_INT,  .value = v.value }; }
   | "quad" ws v:value
@@ -74,6 +76,8 @@ instr =
   | i:jmp      { $$ = i; }
   | i:add      { $$ = i; }
   | i:and      { $$ = i; }
+  | i:cvtsi2sd { $$ = i; }
+  | i:cvtsi2ss { $$ = i; }
   | i:cvtss2sd { $$ = i; }
   | i:cvtsd2ss { $$ = i; }
   | i:cmp      { $$ = i; }
@@ -85,12 +89,14 @@ instr =
   | i:mul      { $$ = i; }
   | i:imul     { $$ = i; }
   | i:neg      { $$ = i; }
+  | i:movaps   { $$ = i; }
   | i:movsd    { $$ = i; }
   | i:movss    { $$ = i; }
   | i:movsx    { $$ = i; }
   | i:movzx    { $$ = i; }
   | i:mov      { $$ = i; }
   | i:or       { $$ = i; }
+  | i:pxor     { $$ = i; }
   | i:set      { $$ = i; }
   | i:sub      { $$ = i; }
   | i:sal      { $$ = i; }
@@ -185,6 +191,20 @@ idiv = "idiv" (
 
 # TODO XXX Combine the xmm variants into single instructions??
 
+cvtsi2sd = "cvtsi2sd" (
+     ws s:r32 ws? ',' ws? d:xmm { $$ = INSTR2(0, s, d); }
+  |  ws s:m   ws? ',' ws? d:xmm { $$ = INSTR2(1, s, d); }
+  |  ws s:r64 ws? ',' ws? d:xmm { $$ = INSTR2(2, s, d); }
+  |  ws s:m   ws? ',' ws? d:xmm { $$ = INSTR2(3, s, d); }
+)  { $$.instr.kind = ASM_CVTSI2SD; }
+
+cvtsi2ss = "cvtsi2ss" (
+     ws s:r32 ws? ',' ws? d:xmm { $$ = INSTR2(0, s, d); }
+  |  ws s:m   ws? ',' ws? d:xmm { $$ = INSTR2(1, s, d); }
+  |  ws s:r64 ws? ',' ws? d:xmm { $$ = INSTR2(2, s, d); }
+  |  ws s:m   ws? ',' ws? d:xmm { $$ = INSTR2(3, s, d); }
+)  { $$.instr.kind = ASM_CVTSI2SS; }
+
 cvtss2sd = "cvtss2sd" (
     ws s:xmm ws? ',' ws? d:xmm { $$ = INSTR2(0, s, d); }
   | ws s:m     ws? ',' ws? d:xmm { $$ = INSTR2(1, s, d); }
@@ -195,6 +215,12 @@ cvtsd2ss = "cvtsd2ss" (
   | ws s:m     ws? ',' ws? d:xmm { $$ = INSTR2(1, s, d); }
 )  { $$.instr.kind = ASM_CVTSD2SS; }
 
+movaps = "movaps" (
+    ws s:xmm ws? ',' ws? d:xmm { $$ = INSTR2(0, s, d); }
+  | ws s:m   ws? ',' ws? d:xmm { $$ = INSTR2(1, s, d); }
+  | ws s:xmm ws? ',' ws? d:m   { $$ = INSTR2(2, s, d); }
+)  { $$.instr.kind = ASM_MOVAPS; }
+
 mulsd = "mulsd" (
     ws s:xmm ws? ',' ws? d:xmm { $$ = INSTR2(0, s, d); }
   | ws s:m     ws? ',' ws? d:xmm { $$ = INSTR2(1, s, d); }
@@ -227,6 +253,11 @@ ucomisd = "ucomisd" (
   | ws s:m     ws? ',' ws? d:xmm { $$ = INSTR2(1, s, d); }
 )  { $$.instr.kind = ASM_UCOMISD; }
 
+pxor = "pxor" (
+    ws s:xmm ws? ',' ws? d:xmm { $$ = INSTR2(0, s, d); }
+  | ws s:m   ws? ',' ws? d:xmm { $$ = INSTR2(1, s, d); }
+)  { $$.instr.kind = ASM_PXOR; }
+
 xorpd = "xorpd" (
     ws s:xmm ws? ',' ws? d:xmm { $$ = INSTR2(0, s, d); }
   | ws s:m   ws? ',' ws? d:xmm { $$ = INSTR2(1, s, d); }
diff --git a/main.c b/main.c
index 8ba72e2..a36250a 100644
--- a/main.c
+++ b/main.c
@@ -689,15 +689,19 @@ static void assembleshift(Instr *instr, uint8_t immreg) {
 }
 
 static void assemblexmmbasicop(Instr *instr, VarBytes prefix, VarBytes opcode) {
-  uint8_t rex, reg, rm;
+  uint8_t rexw, rex, reg, rm;
 
   if (instr->arg1->kind == ASM_MEMARG) {
     assemblemem(&instr->arg1->memarg, 0, prefix, opcode,
                 regbits(instr->arg2->kind));
+  } else if (instr->arg2->kind == ASM_MEMARG) {
+    assemblemem(&instr->arg2->memarg, 0, prefix, opcode,
+                regbits(instr->arg1->kind));
   } else {
+    rexw = isreg64(instr->arg1->kind);
     reg = regbits(instr->arg2->kind);
     rm = regbits(instr->arg1->kind);
-    rex = rexbyte(0, reg & (1 << 3), 0, rm & (1 << 3));
+    rex = rexbyte(rexw, reg & (1 << 3), 0, rm & (1 << 3));
     assemblemodregrm(rex, prefix, opcode, 0x03, reg, rm);
   }
 }
@@ -821,6 +825,9 @@ static void assemble(void) {
     case ASM_DIR_BYTE:
       assemblereloc(v->dirbyte.value.l, v->dirbyte.value.c, 1, R_X86_64_32);
       break;
+    case ASM_DIR_SHORT:
+      assemblereloc(v->dirshort.value.l, v->dirshort.value.c, 2, R_X86_64_32);
+      break;
     case ASM_DIR_INT:
       assemblereloc(v->dirint.value.l, v->dirint.value.c, 4, R_X86_64_32);
       break;
@@ -896,6 +903,18 @@ static void assemble(void) {
     case ASM_CQTO:
       sb2(0x48, 0x99);
       break;
+    case ASM_CVTSI2SD:
+      assemblexmmbasicop(&v->instr, 0xf2, 0x01000f2A);
+      break;
+    case ASM_CVTSI2SS:
+      assemblexmmbasicop(&v->instr, 0xf3, 0x01000f2A);
+      break;
+    case ASM_CVTSS2SD:
+      assemblexmmbasicop(&v->instr, 0xf3, 0x01000f5A);
+      break;
+    case ASM_CVTSD2SS:
+      assemblexmmbasicop(&v->instr, 0xf2, 0x01000f5A);
+      break;
     case ASM_RET:
       sb(0xc3);
       break;
@@ -905,15 +924,16 @@ static void assemble(void) {
       assemblerrm(&v->instr, prefix, 0x8d);
       break;
     }
+    case ASM_MOVAPS: {
+      VarBytes prefix, opcode;
+      prefix = EMPTY_VBYTES;
+      opcode = v->instr.variant == 2 ? 0x01000f29 : 0x01000f28;
+      assemblexmmbasicop(&v->instr, prefix, opcode);
+      break;
+    }
     case ASM_MOV:
       assemblemov(&v->instr);
       break;
-    case ASM_CVTSS2SD:
-      assemblexmmbasicop(&v->instr, 0xf3, 0x01000f5A);
-      break;
-    case ASM_CVTSD2SS:
-      assemblexmmbasicop(&v->instr, 0xf2, 0x01000f5A);
-      break;
     case ASM_MOVSD:
       assemblemovsmmx(&v->instr, 0xf2);
       break;
@@ -1013,6 +1033,10 @@ static void assemble(void) {
       assemblebasicop(&v->instr, variant2op[v->instr.variant], 0x01);
       break;
     }
+    case ASM_PXOR: {
+      assemblexmmbasicop(&v->instr, 0x66, 0x01000fef);
+      break;
+    }
     case ASM_SET: {
       VarBytes prefix, opcode;
       uint8_t rex, reg, rm;
diff --git a/minias.h b/minias.h
index b5eb779..cbaae9b 100644
--- a/minias.h
+++ b/minias.h
@@ -55,6 +55,7 @@ typedef enum {
   ASM_DIR_TEXT,
   ASM_DIR_FILL,
   ASM_DIR_BYTE,
+  ASM_DIR_SHORT,
   ASM_DIR_INT,
   ASM_DIR_QUAD,
   ASM_DIR_BALIGN,
@@ -73,6 +74,8 @@ typedef enum {
   ASM_CMP,
   ASM_CVTSS2SD,
   ASM_CVTSD2SS,
+  ASM_CVTSI2SD,
+  ASM_CVTSI2SS,
   ASM_DIV,
   ASM_IDIV,
   ASM_LEA,
@@ -81,12 +84,14 @@ typedef enum {
   ASM_MULSS,
   ASM_IMUL,
   ASM_MOV,
+  ASM_MOVAPS,
   ASM_MOVSD,
   ASM_MOVSS,
   ASM_MOVSX,
   ASM_MOVZX,
   ASM_NEG,
   ASM_OR,
+  ASM_PXOR,
   ASM_SAL,
   ASM_SAR,
   ASM_SET,
@@ -227,6 +232,11 @@ typedef struct {
 typedef struct {
   AsmKind kind;
   Value value;
+} Short;
+
+typedef struct {
+  AsmKind kind;
+  Value value;
 } Int;
 
 typedef struct {
@@ -305,6 +315,7 @@ union Parsev {
   Jmp jmp;
   Fill fill;
   Byte dirbyte;
+  Short dirshort;
   Int dirint;
   Quad dirquad;
   Imm imm;
diff --git a/test/test.sh b/test/test.sh
index 115be24..a916740 100644
--- a/test/test.sh
+++ b/test/test.sh
@@ -33,6 +33,32 @@ t () {
 t "callq *%rax"
 t "callq *%r10"
 
+t "cvtsi2sd %rax, %xmm1"
+t "cvtsi2sd %rax, %xmm10"
+t "cvtsi2sd (%rax), %xmm1"
+t "cvtsi2sd (%rax), %xmm10"
+t "cvtsi2sd %eax, %xmm1"
+t "cvtsi2sd %eax, %xmm10"
+
+t "cvtsi2ss %rax, %xmm1"
+t "cvtsi2ss %rax, %xmm10"
+t "cvtsi2ss (%rax), %xmm1"
+t "cvtsi2ss (%rax), %xmm10"
+t "cvtsi2ss %eax, %xmm1"
+t "cvtsi2ss %eax, %xmm10"
+
+t "pxor %xmm0, %xmm1"
+t "pxor %xmm10, %xmm1"
+t "pxor (%rax), %xmm0"
+t "pxor (%rax), %xmm10"
+
+t "movaps %xmm0, %xmm1"
+t "movaps %xmm0, (%rax)"
+t "movaps (%rax), %xmm0"
+t "movaps %xmm0, %xmm1"
+t "movaps %xmm10, (%rax)"
+t "movaps (%rax), %xmm10"
+
 t "cvtss2sd %xmm0, %xmm1"
 t "cvtss2sd %xmm10, %xmm1"
 t "cvtss2sd (%rax), %xmm0"
@@ -43,6 +69,9 @@ t "cvtsd2ss %xmm10, %xmm1"
 t "cvtsd2ss (%rax), %xmm0"
 t "cvtsd2ss (%rax), %xmm10"
 
+t "pxor %xmm0, %xmm1"
+t "pxor (%rax), %xmm1"
+
 t "xorps %xmm0, %xmm1"
 t "xorps %xmm10, %xmm1"
 t "xorps (%rax), %xmm0"