120 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_test_checks.py2; RUN: opt < %s -passes=instcombine -S | FileCheck %s3;4; Verify that the result of memchr calls used in equality expressions5; with either the first argument or null are optimally folded.6 7declare ptr @memchr(ptr, i32, i64)8 9 10@a5 = constant [5 x i8] c"12345"11 12; Fold memchr(a5, c, 5) == a5 to *a5 == c.13 14define i1 @fold_memchr_a_c_5_eq_a(i32 %c) {15; CHECK-LABEL: @fold_memchr_a_c_5_eq_a(16; CHECK-NEXT: [[TMP1:%.*]] = trunc i32 [[C:%.*]] to i817; CHECK-NEXT: [[CHAR0CMP:%.*]] = icmp eq i8 [[TMP1]], 4918; CHECK-NEXT: ret i1 [[CHAR0CMP]]19;20 %q = call ptr @memchr(ptr @a5, i32 %c, i64 5)21 %cmp = icmp eq ptr %q, @a522 ret i1 %cmp23}24 25 26; Fold memchr(a5, c, n) == a5 to n && *a5 == c. Unlike the case when27; the first argument is an arbitrary, including potentially past-the-end,28; pointer, this is safe because a5 is dereferenceable.29 30define i1 @fold_memchr_a_c_n_eq_a(i32 %c, i64 %n) {31; CHECK-LABEL: @fold_memchr_a_c_n_eq_a(32; CHECK-NEXT: [[TMP1:%.*]] = trunc i32 [[C:%.*]] to i833; CHECK-NEXT: [[CHAR0CMP:%.*]] = icmp eq i8 [[TMP1]], 4934; CHECK-NEXT: [[TMP2:%.*]] = icmp ne i64 [[N:%.*]], 035; CHECK-NEXT: [[TMP3:%.*]] = select i1 [[TMP2]], i1 [[CHAR0CMP]], i1 false36; CHECK-NEXT: ret i1 [[TMP3]]37;38 %q = call ptr @memchr(ptr @a5, i32 %c, i64 %n)39 %cmp = icmp eq ptr %q, @a540 ret i1 %cmp41}42 43 44; Do not fold memchr(a5 + i, c, n).45 46define i1 @call_memchr_api_c_n_eq_a(i64 %i, i32 %c, i64 %n) {47; CHECK-LABEL: @call_memchr_api_c_n_eq_a(48; CHECK-NEXT: [[P:%.*]] = getelementptr i8, ptr @a5, i64 [[I:%.*]]49; CHECK-NEXT: [[Q:%.*]] = call ptr @memchr(ptr [[P]], i32 [[C:%.*]], i64 [[N:%.*]])50; CHECK-NEXT: [[CMP:%.*]] = icmp eq ptr [[Q]], [[P]]51; CHECK-NEXT: ret i1 [[CMP]]52;53 %p = getelementptr [5 x i8], ptr @a5, i64 0, i64 %i54 %q = call ptr @memchr(ptr %p, i32 %c, i64 %n)55 %cmp = icmp eq ptr %q, %p56 ret i1 %cmp57}58 59 60; Fold memchr(s, c, 15) == s to *s == c.61 62define i1 @fold_memchr_s_c_15_eq_s(ptr %s, i32 %c) {63; CHECK-LABEL: @fold_memchr_s_c_15_eq_s(64; CHECK-NEXT: [[TMP1:%.*]] = load i8, ptr [[S:%.*]], align 165; CHECK-NEXT: [[TMP2:%.*]] = trunc i32 [[C:%.*]] to i866; CHECK-NEXT: [[CHAR0CMP:%.*]] = icmp eq i8 [[TMP1]], [[TMP2]]67; CHECK-NEXT: ret i1 [[CHAR0CMP]]68;69 %p = call ptr @memchr(ptr %s, i32 %c, i64 15)70 %cmp = icmp eq ptr %p, %s71 ret i1 %cmp72}73 74 75; Fold memchr(s, c, 17) != s to *s != c.76 77define i1 @fold_memchr_s_c_17_neq_s(ptr %s, i32 %c) {78; CHECK-LABEL: @fold_memchr_s_c_17_neq_s(79; CHECK-NEXT: [[TMP1:%.*]] = load i8, ptr [[S:%.*]], align 180; CHECK-NEXT: [[TMP2:%.*]] = trunc i32 [[C:%.*]] to i881; CHECK-NEXT: [[CHAR0CMP:%.*]] = icmp ne i8 [[TMP1]], [[TMP2]]82; CHECK-NEXT: ret i1 [[CHAR0CMP]]83;84 %p = call ptr @memchr(ptr %s, i32 %c, i64 17)85 %cmp = icmp ne ptr %p, %s86 ret i1 %cmp87}88 89 90; Fold memchr(s, c, n) == s to *s == c for nonzero n.91 92define i1 @fold_memchr_s_c_nz_eq_s(ptr %s, i32 %c, i64 %n) {93; CHECK-LABEL: @fold_memchr_s_c_nz_eq_s(94; CHECK-NEXT: [[TMP1:%.*]] = load i8, ptr [[S:%.*]], align 195; CHECK-NEXT: [[TMP2:%.*]] = trunc i32 [[C:%.*]] to i896; CHECK-NEXT: [[CHAR0CMP:%.*]] = icmp eq i8 [[TMP1]], [[TMP2]]97; CHECK-NEXT: ret i1 [[CHAR0CMP]]98;99 %nz = or i64 %n, 1100 %p = call ptr @memchr(ptr %s, i32 %c, i64 %nz)101 %cmp = icmp eq ptr %p, %s102 ret i1 %cmp103}104 105 106; But do not fold memchr(s, c, n) as above if n might be zero. This could107; be optimized to the equivalent of N && *S == C provided a short-circuiting108; AND, otherwise the load could read a byte just past the end of an array.109 110define i1 @call_memchr_s_c_n_eq_s(ptr %s, i32 %c, i64 %n) {111; CHECK-LABEL: @call_memchr_s_c_n_eq_s(112; CHECK-NEXT: [[P:%.*]] = call ptr @memchr(ptr [[S:%.*]], i32 [[C:%.*]], i64 [[N:%.*]])113; CHECK-NEXT: [[CMP:%.*]] = icmp eq ptr [[P]], [[S]]114; CHECK-NEXT: ret i1 [[CMP]]115;116 %p = call ptr @memchr(ptr %s, i32 %c, i64 %n)117 %cmp = icmp eq ptr %p, %s118 ret i1 %cmp119}120