From 748091cfd93c64e9303dac5e3802e0b513f14c94 Mon Sep 17 00:00:00 2001
From: fsfod <fsfod11@gmail.com>
Date: Tue, 29 Mar 2016 12:15:40 +0100
Subject: [PATCH] CSE support for intrinsics only enabled for single value
 returning intrinsics that have no side effects and are not forced indirect
 ModRM which could be a load or store

---
 src/lj_opt_fold.c        | 22 +++++++++-
 tests/intrinsic_spec.lua | 92 ++++++++++++++++++++++++++++++++++++++++
 2 files changed, 112 insertions(+), 2 deletions(-)

diff --git a/src/lj_opt_fold.c b/src/lj_opt_fold.c
index ffe32788..999b223d 100644
--- a/src/lj_opt_fold.c
+++ b/src/lj_opt_fold.c
@@ -25,6 +25,7 @@
 #if LJ_HASFFI
 #include "lj_ctype.h"
 #include "lj_carith.h"
+#include "lj_intrinsic.h"
 #endif
 #include "lj_vm.h"
 #include "lj_strscan.h"
@@ -2331,6 +2332,25 @@ LJFOLDF(xload_kptr)
 LJFOLD(XLOAD any any)
 LJFOLDX(lj_opt_fwd_xload)
 
+/* -- Intrinsics ----------------------------------------------------------- */
+
+LJFOLD(INTRN any any)
+LJFOLDF(cse_intrin)
+{
+  CIntrinsic *intrin = lj_intrinsic_get(ctype_ctsG(J2G(J)), fins->op2);
+
+  /* NYI: CSE of multi return intrinsics */
+  if (!LJ_LIKELY(J->flags & JIT_F_OPT_CSE) || intrin->outsz != 1 || 
+      intrin->insz == 0 || (intrin->flags & (INTRINSFLAG_INDIRECT|
+                                INTRINSFLAG_HASSIDE|INTRINSFLAG_MEMORYSIDE))) {
+    return EMITFOLD;
+  }
+  /* This also works for template intrinsics since they append an extra CARG with
+  ** a pointer to there per instance code.
+  */
+  return lj_opt_cse(J);
+}
+
 /* -- Write barriers ------------------------------------------------------ */
 
 /* Write barriers are amenable to CSE, but not across any incremental
@@ -2371,7 +2391,6 @@ LJFOLDF(prof)
     return ref;
   return EMITFOLD;
 }
-
 /* -- Stores and allocations ---------------------------------------------- */
 
 /* Stores and allocations cannot be folded or passed on to CSE in general.
@@ -2405,7 +2424,6 @@ LJFOLD(TDUP any)
 LJFOLD(CNEW any any)
 LJFOLD(XSNEW any any)
 LJFOLD(BUFHDR any any)
-LJFOLD(INTRN any any)
 LJFOLD(ASMRET any any)
 LJFOLDX(lj_ir_emit)
 
diff --git a/tests/intrinsic_spec.lua b/tests/intrinsic_spec.lua
index bf621444..00416c73 100644
--- a/tests/intrinsic_spec.lua
+++ b/tests/intrinsic_spec.lua
@@ -1116,6 +1116,98 @@ context("mixed register type opcodes", function()
   end)
 end)
 
+context("Intrinsic CSE", function()
+
+  it("cse 1 input, same arg", function()
+    assert_cdef([[int32_t add3(int32_t n) __mcode("830mU", 3);]], "add_imm3")
+    
+    local function add3mul2(a)
+      return ffi.C.add3(a)+ffi.C.add3(a)
+    end
+    
+    assert_jit(10, add3mul2, 2)
+  end)
+  
+  it("cse 1 input, diff args", function()
+    assert_cdef([[int32_t add3(int32_t n) __mcode("830mU", 3);]], "add_imm3")
+    
+    local function add3mul2(a, b)
+      return ffi.C.add3(a)+ffi.C.add3(b)
+    end
+    
+    assert_jit(11, add3mul2, 2, 3)
+    assert_noexit(12, add3mul2, 3, 3)
+  end)
+  
+  
+  it("cse 1 input, same and diff args", function()
+    assert_cdef([[int32_t add3(int32_t n) __mcode("830mU", 3);]], "add_imm3")
+    
+    local function add3mul2(a, b)
+      return ffi.C.add3(a)+ffi.C.add3(b)+ffi.C.add3(a)
+    end
+    
+    assert_jit(16, add3mul2, 2, 3)
+    assert_noexit(10, add3mul2, -1, 3)
+  end)
+  
+  it("no cse, indirect ModRM", function()
+    assert_cdef([[int32_t xadd(int32_t* nptr, int32_t n) __mcode("0FC1mRI");]], "xadd")
+
+    local num = ffi.new("int32_t[1]", 0)
+    -- Pretend to use the output values so we don't get DCE'ed since we didn't flag xadd to have side effects
+    local function inc2(a, b)
+      local a = ffi.C.xadd(num, 1)
+      a = a + ffi.C.xadd(num, 1)
+      return num[0], a
+    end
+    
+    local function checker(i, n)
+      assert(i * 2 == n, n)
+    end
+    
+    assert_jitchecker(checker, inc2)
+  end)
+  
+  it("no cse, has side effects", function()
+    --Pretend we have an input register 
+    assert_cdef([[void rdtsc2(int32_t edx) __mcode("0f31_Es") __reglist(out, uint32_t eax) __reglist(mod, edx);]], "rdtsc2")
+
+    local function getticks()
+      local t1 = ffi.C.rdtsc2(0)
+      local t2 = ffi.C.rdtsc2(0)
+      return t2-t1
+    end
+    
+    local function checker(i, n)
+      assert(n > 0, tostring(n))
+    end
+    
+    assert_jitchecker(checker, getticks)
+  end)
+  
+  it("cse, fuse conflict", function()
+   -- assert_cdef([[int32_t add(int32_t i, int32_t n) __mcode("03rMc");]], "add")
+   assert_cdef([[double addsd(double n1, double n2) __mcode("F20F58rMvc");]], "addsd")
+
+    local num = ffi.new("double[2]", 0)
+    
+    local function testfuse(i, a, b)
+      local n = ffi.C.addsd(num[0], num[1])
+      num[1] = i
+      n = n + ffi.C.addsd(num[0], num[1])
+      return n
+    end
+    
+    local function checker(i, n)
+      assert(n > 0, tostring(n))
+    end
+    
+    assert_jitchecker(checker, testfuse, 2.5)
+  end)
+
+end)
+
 end)