https://gcc.gnu.org/bugzilla/show_bug.cgi?id=95793
Bug ID: 95793
Summary: Nested function multi-versioning doesn't work
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: middle-end
Assignee: unassigned at gcc dot gnu.org
Reporter: hjl.tools at gmail dot com
Target Milestone: ---
[hjl@gnu-cfl-2 pr95790]$ cat x.cc
#include <x86intrin.h>
extern char *buf;
__attribute__ ((target ("default")))
static int foo(const char *buf) {
return 1;
}
__attribute__ ((target ("avx2")))
static int foo(const char *buf) {
__m256i x = *(const __m256i_u *) buf;
return __builtin_ia32_pmovmskb256 (x);
}
__attribute__ ((target ("avx512f")))
static int foo(const char *buf) {
__m128i x = *(const __m128i_u *) buf;
return __builtin_ia32_pmovmskb128 (x);
}
__attribute__ ((target ("default")))
int bar(int unsigned size) {
int acc = 0;
for (int i = 0; i < size; i++) {
acc += foo(&buf[i]);
}
return acc;
}
__attribute__ ((target ("avx2")))
int bar(int unsigned size) {
int acc = 0;
for (int i = 0; i < size; i++) {
acc += foo(&buf[i]);
}
return acc;
}
[hjl@gnu-cfl-2 pr95790]$ /usr/gcc-10.1.1-x32/bin/gcc -O2
-flax-vector-conversions -fno-asynchronous-unwind-tables -S x.cc
cat[hjl@gnu-cfl-2 pr95790]$ cat x.s
.file "x.cc"
.text
.p2align 4
.globl _Z3barj
.type _Z3barj, @function
_Z3barj:
.LFB5599:
.cfi_startproc
movl %edi, %eax
ret
.cfi_endproc
.LFE5599:
.size _Z3barj, .-_Z3barj
.p2align 4
.globl _Z3barj.avx2
.type _Z3barj.avx2, @function
_Z3barj.avx2:
.LFB5600:
.cfi_startproc
testl %edi, %edi
je .L7
movq buf(%rip), %rdx
leal -1(%rdi), %ecx
xorl %r8d, %r8d
leaq 1(%rdx), %rax
addq %rax, %rcx
jmp .L6
.p2align 4,,10
.p2align 3
.L11:
addq $1, %rax
.L6:
vmovdqu (%rdx), %xmm1
vinserti128 $0x1, 16(%rdx), %ymm1, %ymm0
vpmovmskb %ymm0, %edx
addl %edx, %r8d
movq %rax, %rdx
cmpq %rcx, %rax
jne .L11
vzeroupper
movl %r8d, %eax
ret
.p2align 4,,10
.p2align 3
.L7:
xorl %r8d, %r8d
movl %r8d, %eax
ret
.cfi_endproc
.LFE5600:
.size _Z3barj.avx2, .-_Z3barj.avx2
.ident "GCC: (GNU) 10.1.1 20200523"
.section .note.GNU-stack,"",@progbits
[hjl@gnu-cfl-2 pr95790]$
_Z3barj.avx2 is unused.
