From aaca4da94fcdcdeedc2b9bff7dec1e3d7c7ecf9a Mon Sep 17 00:00:00 2001
From: Nick Gasson <nick@nickg.me.uk>
Date: Sun, 1 Jan 2023 10:15:46 +0000
Subject: [PATCH] Optimise $EXP to $SHL for powers-of-2

---
 src/jit/jit-llvm.c  | 12 ++++++------
 src/jit/jit-optim.c | 15 +++++++++++++++
 test/test_jit.c     |  8 ++++++++
 3 files changed, 29 insertions(+), 6 deletions(-)

diff --git a/src/jit/jit-llvm.c b/src/jit/jit-llvm.c
index c030475c..0c4f5549 100644
--- a/src/jit/jit-llvm.c
+++ b/src/jit/jit-llvm.c
@@ -1722,15 +1722,15 @@ static void cgen_op_call(llvm_obj_t *obj, cgen_block_t *cgb, jit_ir_t *ir)
 
 #if CLOSED_WORLD
       entry = llvm_add_fn(obj, istr(callee->name), obj->types[LLVM_ENTRY_FN]);
-#else
-      // Must have acquire semantics to synchronise with installing new code
-      entry = LLVMBuildLoad2(obj->builder, obj->types[LLVM_PTR], fptr, "entry");
-      LLVMSetOrdering(entry, LLVMAtomicOrderingAcquire);
 #endif
    }
-   else {
-      entry = llvm_ptr(obj, callee->entry);
+   else
       fptr = llvm_ptr(obj, callee);
+
+   if (entry == NULL) {
+      // Must have acquire semantics to synchronise with installing new code
+      entry = LLVMBuildLoad2(obj->builder, obj->types[LLVM_PTR], fptr, "entry");
+      LLVMSetOrdering(entry, LLVMAtomicOrderingAcquire);
    }
 
 #ifndef LLVM_HAS_OPAQUE_POINTERS
diff --git a/src/jit/jit-optim.c b/src/jit/jit-optim.c
index 3e47fcf0..0d722e62 100644
--- a/src/jit/jit-optim.c
+++ b/src/jit/jit-optim.c
@@ -724,6 +724,20 @@ static void jit_lvn_bzero(jit_ir_t *ir, lvn_state_t *state)
    state->regvn[ir->result] = lvn_new_value(state);
 }
 
+static void jit_lvn_exp(jit_ir_t *ir, lvn_state_t *state)
+{
+   int64_t base, exp;
+   if (lvn_can_fold(ir, state, &base, &exp))
+      lvn_convert_mov(ir, state, LVN_CONST(ipow(base, exp)));
+   else if (lvn_is_const(ir->arg1, state, &base) && base == 2) {
+      ir->op = J_SHL;
+      ir->arg1.int64 = 1;
+      jit_lvn_generic(ir, state, VN_INVALID);
+   }
+   else
+      jit_lvn_generic(ir, state, VN_INVALID);
+}
+
 void jit_do_lvn(jit_func_t *f)
 {
    lvn_state_t state = {
@@ -762,6 +776,7 @@ void jit_do_lvn(jit_func_t *f)
       case J_CLAMP: jit_lvn_clamp(ir, &state); break;
       case MACRO_COPY: jit_lvn_copy(ir, &state); break;
       case MACRO_BZERO: jit_lvn_bzero(ir, &state); break;
+      case MACRO_EXP: jit_lvn_exp(ir, &state); break;
       default: break;
       }
    }
diff --git a/test/test_jit.c b/test/test_jit.c
index cda54688..e093f972 100644
--- a/test/test_jit.c
+++ b/test/test_jit.c
@@ -1367,6 +1367,8 @@ START_TEST(test_lvn1)
       "    SUB     R1, R1, #0    \n"
       "    SUB     R2, #0, R2    \n"
       "    SUB.O   R2, #0, R2    \n"
+      "    $EXP    R3, #3, #4    \n"
+      "    $EXP    R3, #2, R2    \n"
       "    RET                   \n";
 
    jit_handle_t h1 = jit_assemble(j, ident_new("myfunc"), text1);
@@ -1408,6 +1410,12 @@ START_TEST(test_lvn1)
 
    ck_assert_int_eq(f->irbuf[15].op, J_SUB);
 
+   ck_assert_int_eq(f->irbuf[16].op, J_MOV);
+   ck_assert_int_eq(f->irbuf[16].arg1.int64, 81);
+
+   ck_assert_int_eq(f->irbuf[17].op, J_SHL);
+   ck_assert_int_eq(f->irbuf[17].arg1.int64, 1);
+
    jit_free(j);
 }
 END_TEST
-- 
2.39.2