1; RUN: opt -S -codegenprepare < %s | FileCheck %s
2
3target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64"
4target triple = "nvptx64-nvidia-cuda"
5
6; When we bypass slow div with a constant numerator which fits into the bypass
7; width, we still emit the bypass code, but we don't 'or' the numerator with
8; the denominator.
9; CHECK-LABEL: @small_constant_numer
10define i64 @small_constant_numer(i64 %a) {
11  ; CHECK: [[AND:%[0-9]+]] = and i64 %a, -4294967296
12  ; CHECK: icmp eq i64 [[AND]], 0
13
14  ; CHECK: [[TRUNC:%[0-9]+]] = trunc i64 %a to i32
15  ; CHECK: udiv i32 -1, [[TRUNC]]
16  %d = sdiv i64 4294967295, %a  ; 0xffff'ffff
17  ret i64 %d
18}
19
20; When we try to bypass slow div with a constant numerator which *doesn't* fit
21; into the bypass width, leave it as a plain 64-bit div with no bypass.
22; CHECK-LABEL: @large_constant_numer
23define i64 @large_constant_numer(i64 %a) {
24  ; CHECK-NOT: udiv i32
25  %d = sdiv i64 4294967296, %a  ; 0x1'0000'0000
26  ret i64 %d
27}
28
29; For good measure, try a value larger than 2^32.
30; CHECK-LABEL: @larger_constant_numer
31define i64 @larger_constant_numer(i64 %a) {
32  ; CHECK-NOT: udiv i32
33  %d = sdiv i64 5000000000, %a
34  ret i64 %d
35}
36