Subzero: Partial implementation of global initializers.

tests_lit/llvm2ice_tests/nacl-atomic-fence-all.ll

 ; Test that loads/stores don't move across a nacl.atomic.fence.all.
 ; This should apply to both atomic and non-atomic loads/stores
 ; (unlike the non-"all" variety of nacl.atomic.fence, which only
 ; applies to atomic load/stores).
 ;
 ; RUN: %llvm2ice -O2 --verbose none %s | FileCheck %s
 
 declare void @llvm.nacl.atomic.fence.all()
 declare i32 @llvm.nacl.atomic.load.i32(i32*, i32)
 declare void @llvm.nacl.atomic.store.i32(i32, i32*, i32)
@@ skipping 25 matching lines @@
   call void @llvm.nacl.atomic.fence.all()
   store i32 %l_c2, i32* %p_c, align 1
 
   ret i32 %l_c2
 }
 ; CHECK-LABEL: test_fused_load_add_a
 ;    alloca store
 ; CHECK: mov {{.*}}, esp
 ; CHECK: mov dword ptr {{.*}}, 999
 ;    atomic store (w/ its own mfence)
-; CHECK: mov {{.*}}, g32_a
+; CHECK: lea {{.*}}, g32_a
 ; The load + add are optimized into one everywhere.
 ; CHECK: add {{.*}}, dword ptr
 ; CHECK: mov dword ptr
 ; CHECK: mfence
-; CHECK: mov {{.*}}, g32_b
+; CHECK: lea {{.*}}, g32_b
 ; CHECK: add {{.*}}, dword ptr
 ; CHECK: mov dword ptr
-; CHECK: mov {{.*}}, g32_c
+; CHECK: lea {{.*}}, g32_c
 ; CHECK: add {{.*}}, dword ptr
 ; CHECK: mfence
 ; CHECK: mov dword ptr
 
 ; Test with the fence moved up a bit.
 define i32 @test_fused_load_add_b() {
 entry:
   %p_alloca = alloca i8, i32 4, align 4
   %p_alloca_bc = bitcast i8* %p_alloca to i32*
   store i32 999, i32* %p_alloca_bc, align 1
@@ skipping 14 matching lines @@
   %l_c2 = add i32 %l_c, 1
   store i32 %l_c2, i32* %p_c, align 1
 
   ret i32 %l_c2
 }
 ; CHECK-LABEL: test_fused_load_add_b
 ;    alloca store
 ; CHECK: mov {{.*}}, esp
 ; CHECK: mov dword ptr {{.*}}, 999
 ;    atomic store (w/ its own mfence)
-; CHECK: mov {{.*}}, g32_a
+; CHECK: lea {{.*}}, g32_a
 ; CHECK: add {{.*}}, dword ptr
 ; CHECK: mov dword ptr
 ; CHECK: mfence
-; CHECK: mov {{.*}}, g32_b
+; CHECK: lea {{.*}}, g32_b
 ; CHECK: add {{.*}}, dword ptr
 ; CHECK: mov dword ptr
-; CHECK: mov {{.*}}, g32_c
+; CHECK: lea {{.*}}, g32_c
 ; CHECK: mfence
 ; Load + add can still be optimized into one instruction
 ; because it is not separated by a fence.
 ; CHECK: add {{.*}}, dword ptr
 ; CHECK: mov dword ptr
 
 ; Test with the fence splitting a load/add.
 define i32 @test_fused_load_add_c() {
 entry:
   %p_alloca = alloca i8, i32 4, align 4
@@ skipping 16 matching lines @@
   %l_c2 = add i32 %l_c, 1
   store i32 %l_c2, i32* %p_c, align 1
 
   ret i32 %l_c2
 }
 ; CHECK-LABEL: test_fused_load_add_c
 ;    alloca store
 ; CHECK: mov {{.*}}, esp
 ; CHECK: mov dword ptr {{.*}}, 999
 ;    atomic store (w/ its own mfence)
-; CHECK: mov {{.*}}, g32_a
+; CHECK: lea {{.*}}, g32_a
 ; CHECK: add {{.*}}, dword ptr
 ; CHECK: mov dword ptr
 ; CHECK: mfence
-; CHECK: mov {{.*}}, g32_b
+; CHECK: lea {{.*}}, g32_b
 ; This load + add are no longer optimized into one,
 ; though perhaps it should be legal as long as
 ; the load stays on the same side of the fence.
 ; CHECK: mov {{.*}}, dword ptr
 ; CHECK: mfence
 ; CHECK: add {{.*}}, 1
 ; CHECK: mov dword ptr
-; CHECK: mov {{.*}}, g32_c
+; CHECK: lea {{.*}}, g32_c
 ; CHECK: add {{.*}}, dword ptr
 ; CHECK: mov dword ptr
 
 
 ; Test where a bunch of i8 loads could have been fused into one
 ; i32 load, but a fence blocks that.
 define i32 @could_have_fused_loads() {
 entry:
   %ptr1 = bitcast [4 x i8]* @g32_d to i8*
   %b1 = load i8* %ptr1
@@ skipping 19 matching lines @@
   %b12 = or i32 %b1.ext, %b2.shift
   %b3.ext = zext i8 %b3 to i32
   %b3.shift = shl i32 %b3.ext, 16
   %b123 = or i32 %b12, %b3.shift
   %b4.ext = zext i8 %b4 to i32
   %b4.shift = shl i32 %b4.ext, 24
   %b1234 = or i32 %b123, %b4.shift
   ret i32 %b1234
 }
 ; CHECK-LABEL: could_have_fused_loads
-; CHECK: mov {{.*}}, g32_d
+; CHECK: lea {{.*}}, g32_d
 ; CHECK: mov {{.*}}, byte ptr
 ; CHECK: mov {{.*}}, byte ptr
 ; CHECK: mov {{.*}}, byte ptr
 ; CHECK: mfence
 ; CHECK: mov {{.*}}, byte ptr
 
 
 ; Test where an identical load from two branches could have been hoisted
 ; up, and then the code merged, but a fence prevents it.
 define i32 @could_have_hoisted_loads(i32 %x) {
 entry:
   %ptr = bitcast [4 x i8]* @g32_d to i32*
   %cmp = icmp eq i32 %x, 1
   br i1 %cmp, label %branch1, label %branch2
 branch1:
   %y = load i32* %ptr
   ret i32 %y
 branch2:
   call void @llvm.nacl.atomic.fence.all()
   %z = load i32* %ptr
   ret i32 %z
 }
 ; CHECK-LABEL: could_have_hoisted_loads
-; CHECK: mov {{.*}}, g32_d
+; CHECK: lea {{.*}}, g32_d
 ; CHECK: je {{.*}}
 ; CHECK: jmp {{.*}}
 ; CHECK: mov {{.*}}, dword ptr
 ; CHECK: ret
 ; CHECK: mfence
 ; CHECK: mov {{.*}}, dword ptr
 ; CHECK: ret