https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101758
Bug ID: 101758
Summary: Inconsistent optimizations with UBSan
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: sanitizer
Assignee: unassigned at gcc dot gnu.org
Reporter: blubban at gmail dot com
CC: dodji at gcc dot gnu.org, dvyukov at gcc dot gnu.org,
jakub at gcc dot gnu.org, kcc at gcc dot gnu.org, marxin at
gcc dot gnu.org
Target Milestone: ---
Input:
gcc -fsanitize=undefined -O2 -fno-strict-aliasing -Wall
float b(unsigned a) { return *(float*)&a; }
float c(unsigned a) { return *(float*)&a; }
float d(unsigned a) { return *(float*)&a; }
Output:
b:
mov DWORD PTR [rsp-4], edi
movss xmm0, DWORD PTR [rsp-4]
ret
c:
movd xmm0, edi
ret
d:
movd xmm0, edi
ret
or, on ARM
b:
sub sp, sp, #8
str r0, [sp, #4]
vldr.32 s0, [sp, #4]
add sp, sp, #8
bx lr
c:
sub sp, sp, #8
vmov s0, r0
add sp, sp, #8
bx lr
d:
sub sp, sp, #8
vmov s0, r0
add sp, sp, #8
bx lr
Expected output:
Same assembly for all three. And maybe don't set up an unused stack frame on
ARM (it's correctly optimized out without UBSan).
Compiler Explorer: https://godbolt.org/z/Tfs5qazqM
(Found it while trying to determine if that function is a strict aliasing
violation (it is), and whether it can be fixed with memcpy or a union (both
work).)
