From f7167c9ec81a0e8cd99f6692fa60ae1d476c9758 Mon Sep 17 00:00:00 2001
From: lemon <lsof@mailbox.org>
Date: Sat, 20 Dec 2025 07:54:46 +0100
Subject: x86-64: peephole optimization for 3-address shift -> lea

---
 x86_64/emit.c | 15 ++++++++++++++-
 x86_64/isel.c |  1 +
 2 files changed, 15 insertions(+), 1 deletion(-)

(limited to 'x86_64')

diff --git a/x86_64/emit.c b/x86_64/emit.c
index f5b63fa..77f421e 100644
--- a/x86_64/emit.c
+++ b/x86_64/emit.c
@@ -1044,7 +1044,20 @@ emitinstr(uchar **pcode, struct function *fn, struct block *blk, int curi, struc
               mkoper(OMEM, .base = mkregoper(ins->l).reg, .index = NOINDEX, .disp = -intconval(ins->r)));
       }
       break;
-   case Oshl: X = Xshl; goto ALU2;
+   case Oshl:
+      dst = reg2oper(ins->reg-1);
+      src = mkregoper(ins->l);
+      if (dst.reg == src.reg)
+         Xshl(pcode, cls, dst, mkimmdatregoper(ins->r));
+      else {
+         uint sh = ins->r.i;
+         assert(ins->r.t == RICON && sh <= 3);
+         if (sh == 1) /* shl x, 1 -> lea [x + x] */
+            Xlea(pcode, cls, dst, mkoper(OMEM, .base = src.reg, .index = src.reg));
+         else /* shl x, n -> lea [x*(1<<n)+0x0] */
+            Xlea(pcode, cls, dst, mkoper(OMEM, .base = NOBASE, .index = src.reg, .shift = sh));
+      }
+      break;
    case Osar: X = Xsar; goto ALU2;
    case Oslr: X = Xshr; goto ALU2;
    case Oand:
diff --git a/x86_64/isel.c b/x86_64/isel.c
index 19692ed..2b087cd 100644
--- a/x86_64/isel.c
+++ b/x86_64/isel.c
@@ -507,6 +507,7 @@ sel(struct function *fn, struct instr *ins, struct block *blk, int *curi)
    ALU:
       if (!(op == Oadd && kisint(ins->cls))) /* 3-address add is lea */
       if (!(op == Omul && kisint(ins->cls) && isimm32(ins->r))) /* for (I)MUL r,r/m,imm */
+      if (!(op == Oshl && ins->r.t == RICON && ins->r.i <= 3)) /* can be lea */
          ins->inplace = 1;
       if (iscon(ins->l)) {
          fixarg(&ins->l, ins, blk, curi);
-- 
cgit v1.2.3