Loop alignment: appears to be a small win for reverse-complement

basis/compiler/cfg/builder/builder.factor

@@ -100,6 +100,7 @@ GENERIC: emit-node ( node -- next )
     basic-block get swap loops get set-at ;
 
 : compile-loop ( node -- next )
+    ##loop-entry
     begin-basic-block
     [ label>> id>> remember-loop ] [ child>> emit-nodes ] bi
     iterate-next ;

basis/compiler/cfg/instructions/instructions.factor

@@ -160,6 +160,8 @@ INSN: ##epilogue ;
 
 INSN: ##branch ;
 
+INSN: ##loop-entry ;
+
 ! Condition codes
 SYMBOL: cc<
 SYMBOL: cc<=

basis/compiler/codegen/codegen.factor

@@ -205,6 +205,8 @@ M: ##write-barrier generate-insn
 
 M: _gc generate-insn drop %gc ;
 
+M: ##loop-entry generate-insn drop %loop-entry ;
+
 ! ##alien-invoke
 GENERIC: reg-size ( register-class -- n )
 

basis/cpu/architecture/architecture.factor

@@ -130,6 +130,8 @@ HOOK: %spill-float cpu ( src n -- )
 HOOK: %reload-integer cpu ( dst n -- )
 HOOK: %reload-float cpu ( dst n -- )
 
+HOOK: %loop-entry cpu ( -- )
+
 ! FFI stuff
 
 ! Is this integer small enough to appear in value template

basis/cpu/x86/architecture/architecture.factor

@@ -51,8 +51,8 @@ M: x86 %call ( label -- ) CALL ;
 M: x86 %jump-label ( label -- ) JMP ;
 M: x86 %return ( -- ) 0 RET ;
 
-: code-alignment ( -- n )
-    building get length dup cell align swap - ;
+: code-alignment ( align -- n )
+    [ building get [ integer? ] count dup ] dip align swap - ;
 
 : align-code ( n -- )
     0 <repetition> % ;
@@ -66,7 +66,7 @@ M:: x86 %dispatch ( src temp -- )
     src temp ADD
     src HEX: 7f [+] JMP
     ! Fix up the displacement above
-    code-alignment dup bootstrap-cell 8 = 15 9 ? +
+    cell code-alignment dup bootstrap-cell 8 = 15 9 ? +
     building get dup pop* push
     align-code ;
 
@@ -466,6 +466,9 @@ M: x86 %reload-integer ( dst n -- ) spill-integer@ MOV ;
 M: x86 %spill-float spill-float@ swap MOVSD ;
 M: x86 %reload-float spill-float@ MOVSD ;
 
+M: x86 %loop-entry
+    16 code-alignment [ NOP ] times ;
+
 M: int-regs %save-param-reg drop >r stack@ r> MOV ;
 M: int-regs %load-param-reg drop swap stack@ MOV ;
 

basis/cpu/x86/assembler/assembler.factor

@@ -446,6 +446,10 @@ M: operand IMUL2 OCT: 257 extended-opcode (2-operand) ;
 
 : CPUID ( -- ) HEX: a2 extended-opcode, ;
 
+! Misc
+
+: NOP ( -- ) HEX: 90 , ;
+
 ! x87 Floating Point Unit
 
 : FSTPS ( operand -- ) { BIN: 011 f HEX: d9 } 1-operand ;