[llvm-bugs] [Bug 51104] New: Suboptimal codegen for CAS loop

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=51104

            Bug ID: 51104
           Summary: Suboptimal codegen for CAS loop
           Product: clang
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: LLVM Codegen
          Assignee: unassignedclangbugs at nondot.org
          Reporter: gonzalo.gadeschi at gmail.com
                CC: llvm-bugs at lists.llvm.org, neeilans at live.com,
                    richard-llvm at metafoo.co.uk

Given this CAS loop:

#include <atomic>
int fetch_max(std::atomic<int>& mem, int val) {
    int read = mem.load(std::memory_order_relaxed);
    int write;
    do {
        write = std::max(read, val);
    } while( !mem.compare_exchange_weak(read, write) );
    return read;
}

Using -O3 -fno-exceptions -g0 -march=skylake, gcc generates:

fetch_max(std::atomic<int>&, int):
        mov     eax, DWORD PTR [rdi]
.L2:
        cmp     eax, esi
        mov     edx, esi
        cmovge  edx, eax
        lock cmpxchg    DWORD PTR [rdi], edx
        jne     .L2
        ret

but clang generates:

fetch_max(std::atomic<int>&, int):             # @fetch_max(std::atomic<int>&,
int)
        movl    (%rdi), %ecx
        cmpl    %esi, %ecx
        movl    %ecx, %edx
        cmovll  %esi, %edx
        movl    %ecx, %eax
        lock            cmpxchgl        %edx, (%rdi)
        je      .LBB0_3
.LBB0_1:                                # =>This Inner Loop Header: Depth=1
        movl    %eax, %ecx
        cmpl    %esi, %eax
        movl    %eax, %edx
        cmovll  %esi, %edx
        lock            cmpxchgl        %edx, (%rdi)
        jne     .LBB0_1
.LBB0_3:
        movl    %ecx, %eax
        retq

If instead of return read, I change the example to return write, then the
codegen improves significantly.

-- 
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/20210715/86610eea/attachment-0001.html>