[Bug target/82038] New: Very poor optimization of constant multiply on ARM Cortex-M7

[headch at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82038

            Bug ID: 82038
           Summary: Very poor optimization of constant multiply on ARM
                    Cortex-M7
           Product: gcc
           Version: 7.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: headch at gmail dot com
  Target Milestone: ---

Consider the following source code:

#include <stdint.h>

int64_t f(int32_t x) {
        return x * 16384LL;
}

int64_t g(int32_t x) {
        return static_cast<int64_t>(x) << 14;
}

Compile it with the following command:

armv7m-none-eabihf-g++ -ffreestanding -Wall -Wextra -O2 -mcpu=cortex-m7
-std=c++17 -c test.cpp

It produces the following code:

00000000 <_Z1fl>:
   0:   b430            push    {r4, r5}
   2:   17c5            asrs    r5, r0, #31
   4:   4603            mov     r3, r0
   6:   0380            lsls    r0, r0, #14
   8:   03a9            lsls    r1, r5, #14
   a:   bc30            pop     {r4, r5}
   c:   ea41 4193       orr.w   r1, r1, r3, lsr #18
  10:   4770            bx      lr
  12:   bf00            nop

00000014 <_Z1gl>:
  14:   4601            mov     r1, r0
  16:   0380            lsls    r0, r0, #14
  18:   1489            asrs    r1, r1, #18
  1a:   4770            bx      lr

The implementation of f could be the same as g, yet it’s really quite awful.

Changing -mcpu=cortex-m7 to -mcpu=cortex-m4 doesn’t affect g. It yields rather
better code for f than the M7 build, but still not as good as g.

I could just use g, but that isn’t really a good option because left-shifting a
negative number is UB.