Inline memove for small constant sizes and refactor memcpy and memset.

tests_lit/llvm2ice_tests/nacl-mem-intrinsics.ll

Index: tests_lit/llvm2ice_tests/nacl-mem-intrinsics.ll
diff --git a/tests_lit/llvm2ice_tests/nacl-mem-intrinsics.ll b/tests_lit/llvm2ice_tests/nacl-mem-intrinsics.ll
index 8175eab856947136bd127f0831b37340a98caa6a..f2fe82e9a50d8dc560b77220645df3d0dc6e96a5 100644
--- a/tests_lit/llvm2ice_tests/nacl-mem-intrinsics.ll
+++ b/tests_lit/llvm2ice_tests/nacl-mem-intrinsics.ll
@@ -145,14 +145,12 @@ entry:
 ; ARM32-LABEL: test_memcpy_large_const_len
 ; ARM32: bl {{.*}} memcpy
 
-; TODO(jvoung) -- if we want to be clever, we can do memset without a function
-; call similar to memcpy.
 define void @test_memmove(i32 %iptr_dst, i32 %iptr_src, i32 %len) {
 entry:
   %dst = inttoptr i32 %iptr_dst to i8*
   %src = inttoptr i32 %iptr_src to i8*
   call void @llvm.memmove.p0i8.p0i8.i32(i8* %dst, i8* %src,
-                                        i32 %len, i32 1, i1 false)
+                                       i32 %len, i32 1, i1 false)
   ret void
 }
 ; CHECK-LABEL: test_memmove
@@ -160,17 +158,119 @@ entry:
 ; ARM32-LABEL: test_memmove
 ; ARM32: bl {{.*}} memmove
 
-define void @test_memmove_const_len_align(i32 %iptr_dst, i32 %iptr_src) {
+define void @test_memmove_long_const_len(i32 %iptr_dst, i32 %iptr_src) {
 entry:
   %dst = inttoptr i32 %iptr_dst to i8*
   %src = inttoptr i32 %iptr_src to i8*
   call void @llvm.memmove.p0i8.p0i8.i32(i8* %dst, i8* %src,
-                                        i32 32, i32 1, i1 false)
+                                       i32 4876, i32 1, i1 false)
   ret void
 }
-; CHECK-LABEL: test_memmove_const_len_align
+; CHECK-LABEL: test_memmove_long_const_len
 ; CHECK: call {{.*}} R_{{.*}} memmove
-; ARM32-LABEL: test_memmove_const_len_align
+; ARM32-LABEL: test_memmove_long_const_len
+; ARM32: bl {{.*}} memmove
+
+define void @test_memmove_very_small_const_len(i32 %iptr_dst, i32 %iptr_src) {
+entry:
+  %dst = inttoptr i32 %iptr_dst to i8*
+  %src = inttoptr i32 %iptr_src to i8*
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dst, i8* %src,
+                                       i32 2, i32 1, i1 false)
+  ret void
+}
+; CHECK-LABEL: test_memmove_very_small_const_len
+; CHECK: mov [[REG:[^,]*]],WORD PTR [{{.*}}]
+; CHECK-NEXT: mov WORD PTR [{{.*}}],[[REG]]
+; CHECK-NOT: mov
+; ARM32-LABEL: test_memmove_very_small_const_len
+; ARM32: bl {{.*}} memmove
+
+define void @test_memmove_const_len_3(i32 %iptr_dst, i32 %iptr_src) {
+entry:
+  %dst = inttoptr i32 %iptr_dst to i8*
+  %src = inttoptr i32 %iptr_src to i8*
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dst, i8* %src,
+                                       i32 3, i32 1, i1 false)
+  ret void
+}
+; CHECK-LABEL: test_memmove_const_len_3
+; CHECK: mov [[REG0:[^,]*]],WORD PTR [{{.*}}]
+; CHECK-NEXT: mov [[REG1:[^,]*]],BYTE PTR [{{.*}}+0x2]
+; CHECK-NEXT: mov WORD PTR [{{.*}}],[[REG0]]
+; CHECK-NEXT: mov BYTE PTR [{{.*}}+0x2],[[REG1]]
+; CHECK-NOT: mov
+; ARM32-LABEL: test_memmove_const_len_3
+; ARM32: bl {{.*}} memmove
+
+define void @test_memmove_mid_const_len(i32 %iptr_dst, i32 %iptr_src) {
+entry:
+  %dst = inttoptr i32 %iptr_dst to i8*
+  %src = inttoptr i32 %iptr_src to i8*
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dst, i8* %src,
+                                       i32 9, i32 1, i1 false)
+  ret void
+}
+; CHECK-LABEL: test_memmove_mid_const_len
+; CHECK: movq [[REG0:xmm[0-9]+]],QWORD PTR [{{.*}}]
+; CHECK-NEXT: mov [[REG1:[^,]*]],BYTE PTR [{{.*}}+0x8]
+; CHECK-NEXT: movq QWORD PTR [{{.*}}],[[REG0]]
+; CHECK-NEXT: mov BYTE PTR [{{.*}}+0x8],[[REG1]]
+; CHECK-NOT: mov
+; ARM32-LABEL: test_memmove_mid_const_len
+; ARM32: bl {{.*}} memmove
+
+define void @test_memmove_mid_const_len_overlap(i32 %iptr_dst, i32 %iptr_src) {
+entry:
+  %dst = inttoptr i32 %iptr_dst to i8*
+  %src = inttoptr i32 %iptr_src to i8*
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dst, i8* %src,
+                                       i32 15, i32 1, i1 false)
+  ret void
+}
+; CHECK-LABEL: test_memmove_mid_const_len_overlap
+; CHECK: movq [[REG0:xmm[0-9]+]],QWORD PTR [{{.*}}]
+; CHECK-NEXT: movq [[REG1:xmm[0-9]+]],QWORD PTR [{{.*}}+0x7]
+; CHECK-NEXT: movq QWORD PTR [{{.*}}],[[REG0]]
+; CHECK-NEXT: movq QWORD PTR [{{.*}}+0x7],[[REG1]]
+; CHECK-NOT: mov
+; ARM32-LABEL: test_memmove_mid_const_len_overlap
+; ARM32: bl {{.*}} memmove
+
+define void @test_memmove_large_const_len_overlap(i32 %iptr_dst, i32 %iptr_src) {
+entry:
+  %dst = inttoptr i32 %iptr_dst to i8*
+  %src = inttoptr i32 %iptr_src to i8*
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dst, i8* %src,
+                                       i32 30, i32 1, i1 false)
+  ret void
+}
+; CHECK-LABEL: test_memmove_large_const_len_overlap
+; CHECK: movups [[REG0:xmm[0-9]+]],XMMWORD PTR [{{.*}}]
+; CHECK-NEXT: movups [[REG1:xmm[0-9]+]],XMMWORD PTR [{{.*}}+0xe]
+; CHECK-NEXT: movups XMMWORD PTR [{{.*}}],[[REG0]]
+; CHECK-NEXT: movups XMMWORD PTR [{{.*}}+0xe],[[REG1]]
+; CHECK-NOT: mov
+; ARM32-LABEL: test_memmove_large_const_len_overlap
+; ARM32: bl {{.*}} memmove
+
+define void @test_memmove_large_const_len(i32 %iptr_dst, i32 %iptr_src) {
+entry:
+  %dst = inttoptr i32 %iptr_dst to i8*
+  %src = inttoptr i32 %iptr_src to i8*
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dst, i8* %src,
+                                       i32 33, i32 1, i1 false)
+  ret void
+}
+; CHECK-LABEL: test_memmove_large_const_len
+; CHECK: movups [[REG0:xmm[0-9]+]],XMMWORD PTR [{{.*}}+0x10]
+; CHECK-NEXT: movups [[REG1:xmm[0-9]+]],XMMWORD PTR [{{.*}}]
+; CHECK-NEXT: mov [[REG2:[^,]*]],BYTE PTR [{{.*}}+0x20]
+; CHECK-NEXT: movups XMMWORD PTR [{{.*}}+0x10],[[REG0]]
+; CHECK-NEXT: movups XMMWORD PTR [{{.*}}],[[REG1]]
+; CHECK-NEXT: mov BYTE PTR [{{.*}}+0x20],[[REG2]]
+; CHECK-NOT: mov
+; ARM32-LABEL: test_memmove_large_const_len
 ; ARM32: bl {{.*}} memmove
 
 define void @test_memset(i32 %iptr_dst, i32 %wide_val, i32 %len) {