x86-64: micro-optimize LEA DWORD-REG,EA

If we are loading a DWORD EA into a DWORD register, we may not need the
REX prefix.  Unlikely to occur in current SBCL code, but does come up
when you're writing your own VOPs for 32-bit operations.

diff --git a/src/compiler/x86-64/insts.lisp b/src/compiler/x86-64/insts.lisp
index a23f5f6..15efc7d 100644 (file)
--- a/src/compiler/x86-64/insts.lisp
+++ b/src/compiler/x86-64/insts.lisp
@@ -1799,7 +1799,7 @@
   (:emitter
    (aver (or (dword-reg-p dst) (qword-reg-p dst)))
    (maybe-emit-rex-for-ea segment src dst
-                          :operand-size :qword)
+                          :operand-size (if (dword-reg-p dst) :dword :qword))
    (emit-byte segment #b10001101)
    (emit-ea segment src (reg-tn-encoding dst))))