[llvm-bugs] [Bug 46655] New: [ARM] [v7m] Missed optimization - condition check of `do-while` case

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=46655

            Bug ID: 46655
           Summary: [ARM] [v7m] Missed optimization - condition check of
                    `do-while` case
           Product: clang
           Version: trunk
          Hardware: All
                OS: All
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: C
          Assignee: unassignedclangbugs at nondot.org
          Reporter: alex_lop at walla.com
                CC: blitzrakete at gmail.com, dgregor at apple.com,
                    erik.pilkington at gmail.com, llvm-bugs at lists.llvm.org,
                    richard-llvm at metafoo.co.uk

I've noticed that when compiling the below code with clang for armv7
architecture, using -Os or -Oz optimizations, the produced assembly contains
unneeded `cmp` instruction:

    #include <stdint.h>
    #include <limits.h>

    #define NUM_BITS(type)                  (CHAR_BIT * sizeof(type))

    #define MOST_SIGNIFICANT_ONE(x)         (NUM_BITS(x) - 1 -
(uint32_t)_Generic((x), \
                                             unsigned int       :
__builtin_clz,   \
                                             unsigned long      :
__builtin_clzl,  \
                                             unsigned long long :
__builtin_clzll)(x))

    typedef void (*fn_ptr)(void);

    void boo(uint32_t);

    void foo(uint32_t mask)
    {
        uint32_t idx;

        // it is guarnteed that mask != 0 and mask < 0x40
        do {
            idx = MOST_SIGNIFICANT_ONE(mask);
            boo(idx);
            mask &= ~(1u << idx);
        } while(mask);
    }

The produced assembly is

        .section        __TEXT,__text,regular,pure_instructions
        .syntax unified
        .globl  _foo                    @ -- Begin function foo
        .p2align        1
        .code   16                      @ @foo
        .thumb_func     _foo
    _foo:
        .cfi_startproc
    @ %bb.0:
        push    {r4, r5, r6, r7, lr}
        .cfi_def_cfa_offset 20
        .cfi_offset lr, -4
        .cfi_offset r7, -8
        .cfi_offset r6, -12
        .cfi_offset r5, -16
        .cfi_offset r4, -20
        add     r7, sp, #12
        .cfi_def_cfa r7, 8
        str     r11, [sp, #-4]!
        .cfi_offset r11, -24
        mov     r4, r0
        mov.w   r5, #-2147483648
    LBB0_1:                                 @ =>This Inner Loop Header: Depth=1
        clz     r6, r4
        rsb.w   r0, r6, #31
        bl      _boo
        lsr.w   r0, r5, r6
        bics    r4, r0
        cmp     r4, #0           @ <========= THIS `cmp` instruction is not
needed since `bics` updates the zero flag
        bne     LBB0_1
    @ %bb.2:
        ldr     r11, [sp], #4
        pop     {r4, r5, r6, r7, pc}
        .cfi_endproc
                                            @ -- End function
    .subsections_via_symbols

In this function, the condition for `while(mask)` is checked using `cmp`
assembly instruction but the previous instruction `bics` has already updated
the zero flag for r4.

I tried to reproduce it for armv7a, there was used `bic` instead of `bics`:
https://godbolt.org/z/kmMis2

Clang verion:
  Apple clang version 12.0.0 (clang-1200.0.22.21)
  Target: x86_64-apple-darwin19.5.0
  Thread model: posix

Compilation command: 
clang -arch armv7m -munaligned-access -Wall -Wextra -Wshadow -Wpointer-arith
-Wconversion -Werror -fno-strict-aliasing -Oz -std=c11 -S test.c -o test.s

-- 
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/20200709/43f35c2f/attachment.html>