[llvm-bugs] [Bug 43238] New: Integer promotion quirk prevents efficient power of 2 division

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=43238

            Bug ID: 43238
           Summary: Integer promotion quirk prevents efficient power of 2
                    division
           Product: clang
           Version: 8.0
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: LLVM Codegen
          Assignee: unassignedclangbugs at nondot.org
          Reporter: kim.nilsson at effnet.com
                CC: llvm-bugs at lists.llvm.org, neeilans at live.com,
                    richard-llvm at metafoo.co.uk

Created attachment 22476
  --> https://bugs.llvm.org/attachment.cgi?id=22476&action=edit
Example LLVM IR

For the function,

uint_fast32_t foo_a(uint_fast8_t x) {
    const uint_fast32_t q = 1 << x;
    return 256 / q;
}

GCC-{8.*,9.*} generates inefficient division instructions for both x86_64 and
ARM. Since 'q' is a power of 2, the result should be a single right shift.
Indeed, if the function is changed to,

uint_fast32_t foo_b(uint_fast8_t x) {
    return 256 / (1 << x); // This is, of course, equivalent to (256 >> x)
},

the expected instructions are generated. Counter-intuitively, if the original
function is instead changed slightly to,

uint_fast32_t foo_c(uint_fast8_t x) {
    const uint_fast32_t q = (uint_fast32_t)1 << x;
    return 256 / q;
},

the expected instructions are once again generated.

See https://godbolt.org/z/X-It3b for examples.


(FWIW, GCC also has this behavior and I've reported the same thing there
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91680)

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20190906/db520762/attachment.html>