1 ; RUN: opt -S -passes='require<profile-summary>,function(codegenprepare)' < %s | FileCheck %s
3 target datalayout = "e-i64:64-v16:16-v32:32-n16:32:64"
4 target triple = "nvptx64-nvidia-cuda"
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
9 ; CHECK-LABEL: @small_constant_numer
10 define i64 @small_constant_numer(i64 %a) {
11 ; CHECK: [[AND:%[0-9]+]] = and i64 %a, -4294967296
12 ; CHECK: icmp eq i64 [[AND]], 0
14 ; CHECK: [[TRUNC:%[0-9]+]] = trunc i64 %a to i32
15 ; CHECK: udiv i32 -1, [[TRUNC]]
16 %d = sdiv i64 4294967295, %a ; 0xffff'ffff
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
23 define i64 @large_constant_numer(i64 %a) {
25 %d = sdiv i64 4294967296, %a ; 0x1'0000'0000
29 ; For good measure, try a value larger than 2^32.
30 ; CHECK-LABEL: @larger_constant_numer
31 define i64 @larger_constant_numer(i64 %a) {
33 %d = sdiv i64 5000000000, %a