Allow movsd/movss patching in asm template

Author: shirok

lib/lang/asm/x86_64.scm

@@ -123,9 +123,18 @@
          [patches (asm-template-patches tmpl)])
     ;; Apply each patch whose keyword appears in PARAMS.
     (for-each (match-lambda
+                ;; Integer patch: write value as little-endian bytes.
                 [(kw offset width)
                  (and-let1 val (assq kw params)
-                   (patch-bytes! bytes offset width (cdr val)))])
+                   (patch-bytes! bytes offset width (cdr val)))]
+                ;; movs_ patch: translate symbol movsd->#xf2, movss->#xf3.
+                [(kw offset 1 'movs_)
+                 (let* ([v   (assq-ref params kw 'movsd)]
+                        [pfx (case v
+                               [(movsd) #xf2]
+                               [(movss) #xf3]
+                               [else (error "movs_ value must be movsd or movss:" v)])])
+                   (u8vector-set! bytes offset pfx))])
               patches)
     (values bytes labels)))
 
@@ -300,6 +309,15 @@
     [`(movsd (mem . ,x) (sse ,dst)) (! (opc'(#xf2 #x0f #x10)) (reg dst) (mem x))]
     [`(movsd (sse ,src) (mem . ,x)) (! (opc'(#xf2 #x0f #x11)) (reg src) (mem x))]
 
+    ;; movs_ -- switchable movsd/movss: prefix byte (#xf2/#xf3) is patchable.
+    ;; Default encoding uses movsd (#xf2); asm-instantiate flips to #xf3 for movss.
+    [`((movs_ ,kw) (sse ,src) (sse ,dst))
+     (op-movs_ kw (! (opc '(#xf2 #x0f #x10)) (reg dst) (r/m-reg dst)))]
+    [`((movs_ ,kw) (mem . ,x) (sse ,dst))
+     (op-movs_ kw (! (opc '(#xf2 #x0f #x10)) (reg dst) (mem x)))]
+    [`((movs_ ,kw) (sse ,src) (mem . ,x))
+     (op-movs_ kw (! (opc '(#xf2 #x0f #x11)) (reg src) (mem x)))]
+
     ;; calculations
     [('addq _ _)                   (op-add pinsn 0)]
     [('orq _ _)                    (op-add pinsn 1)]
@@ -364,6 +382,20 @@
         ((! w (rex.b dst) (opc+rq #xb8 dst) (imm64 laddr)) 0 0))
       ((! w (rex.b dst) (opc+rq #xb8 dst) (imm64 0)) 0 0)))) ;; dummy
 
+;; op-movs_ :: keyword, closure -> closure
+;;   Wraps an inner expand-spec closure for a movsd instruction.  In pass 2,
+;;   records a patch of kind 'movs_ at the prefix byte (offset 0 within the
+;;   instruction, i.e. addr - length(bytes)).
+;;   asm-instantiate translates movsd->#xf2, movss->#xf3 when applying it.
+(define (op-movs_ kw inner)
+  (^[a t]
+    (let1 bytes (inner a t)
+      (when (and t (patch-collector))
+        (let1 acc (patch-collector)
+          (set-car! acc (cons (list kw (- a (length bytes)) 1 'movs_)
+                              (car acc)))))
+      bytes)))
+
 ;; addq family
 ;;  opcode variations are derived from a single number, regc.
 (define (op-add pinsn regc)

tests/lang.scm

@@ -331,7 +331,7 @@
              #f
              (eq? b1 b2)))))
 
-;; --- Stage 2: immediate-value placeholders ---
+;; --- immediate-value placeholders ---
 
 ;; imm64 placeholder: movq (imm64 :val), %rax
 ;; Encoding: REX.W(48) B8 <8 bytes immediate>  -- 10 bytes total
@@ -421,7 +421,7 @@
            '((start . 0))
            labels)))
 
-;; --- Stage 3: data-directive placeholders ---
+;; --- data-directive placeholders ---
 
 ;; .dataq placeholder: 8-byte hole
 (let ()
@@ -479,6 +479,74 @@
            '(#x08 #x07 #x06 #x05 #x04 #x03 #x02 #x01)
            (u8vector->list b))))
 
+;; --- movs_ instruction-variant placeholder ---
+
+;; ((movs_ :op) %xmm0 %xmm1)  -- sse->sse form
+;; Default encoding: F2 0F 10 ModRM
+;; ModRM: mod=3(reg-reg), reg=1(xmm1/dst), r/m=1(xmm1/dst) -> 0b11_001_001 = #xc9
+(let ()
+  (define tmpl (asm-template '(((movs_ :op) %xmm0 %xmm1))))
+  ;; Default is movsd
+  (receive (b _) (asm-instantiate tmpl '())
+    (test* "movs_ sse->sse default (movsd)"
+           '(#xf2 #x0f #x10 #xc9)
+           (u8vector->list b)))
+  ;; Explicit movsd
+  (receive (b _) (asm-instantiate tmpl '((:op . movsd)))
+    (test* "movs_ sse->sse explicit movsd"
+           '(#xf2 #x0f #x10 #xc9)
+           (u8vector->list b)))
+  ;; Switch to movss
+  (receive (b _) (asm-instantiate tmpl '((:op . movss)))
+    (test* "movs_ sse->sse switched to movss"
+           '(#xf3 #x0f #x10 #xc9)
+           (u8vector->list b)))
+  ;; Re-instantiate as movsd again (template not mutated)
+  (receive (b _) (asm-instantiate tmpl '((:op . movsd)))
+    (test* "movs_ sse->sse back to movsd"
+           '(#xf2 #x0f #x10 #xc9)
+           (u8vector->list b))))
+
+;; ((movs_ :op) (%rax) %xmm1)  -- mem->sse form
+;; Default encoding: F2 0F 10 ModRM
+;; ModRM: mod=0, reg=1(xmm1/dst), r/m=0(%rax) -> 0b00_001_000 = #x08
+(let ()
+  (define tmpl (asm-template '(((movs_ :op) (%rax) %xmm1))))
+  (receive (b _) (asm-instantiate tmpl '())
+    (test* "movs_ mem->sse default (movsd)"
+           '(#xf2 #x0f #x10 #x08)
+           (u8vector->list b)))
+  (receive (b _) (asm-instantiate tmpl '((:op . movss)))
+    (test* "movs_ mem->sse switched to movss"
+           '(#xf3 #x0f #x10 #x08)
+           (u8vector->list b))))
+
+;; ((movs_ :op) %xmm0 (%rax))  -- sse->mem form
+;; Default encoding: F2 0F 11 ModRM
+;; ModRM: mod=0, reg=0(xmm0/src), r/m=0(%rax) -> #x00
+(let ()
+  (define tmpl (asm-template '(((movs_ :op) %xmm0 (%rax)))))
+  (receive (b _) (asm-instantiate tmpl '())
+    (test* "movs_ sse->mem default (movsd)"
+           '(#xf2 #x0f #x11 #x00)
+           (u8vector->list b)))
+  (receive (b _) (asm-instantiate tmpl '((:op . movss)))
+    (test* "movs_ sse->mem switched to movss"
+           '(#xf3 #x0f #x11 #x00)
+           (u8vector->list b))))
+
+;; movs_ combined with other placeholders in a sequence
+(let ()
+  (define tmpl (asm-template '((.dataq :fn-ptr)
+                                ((movs_ :variant) (%rax) %xmm0)
+                                (ret))))
+  (receive (b _) (asm-instantiate tmpl `((:fn-ptr . #xdeadbeef) (:variant . movss)))
+    (test* "movs_ combined with .dataq placeholder"
+           (append '(#xef #xbe #xad #xde 0 0 0 0) ; .dataq :fn-ptr
+                   '(#xf3 #x0f #x10 #x00)          ; movss (%rax),%xmm0
+                   '(#xc3))                         ; ret
+           (u8vector->list b))))
+
 ;;----------------------------------------------------------------------
 (test-section "lang.c")