https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108322
Bug ID: 108322
Summary: Using __register parameter with -ftree-vectorize
(default with -O2) results in massive code bloat
Product: gcc
Version: 12.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: gerbilsoft at gerbilsoft dot com
Target Milestone: ---
While examining some code using the bloaty tool, I found that a function for
deinterleaving Super Magic Drive ROM images was taking up ~5 KB when it should
have been less than 1 KB. On examining the disassembly, there appeared to be a
lot of unnecessary instructions; compiling with clang and MSVC resulted in
significantly fewer instructions. Either removing __restrict from the function
parameters (two pointers), or specifying -fno-tree-vectorize to disable
auto-vectorization, fixes this issue with gcc-12.
The generated code isn't buggy as far as I can tell, and it benchmarks around
the same as the non-bloated version.
I've narrowed it down to the following minimal test case:
#include <stdint.h>
#define SMD_BLOCK_SIZE 16384
void decodeBlock_cpp(uint8_t *__restrict pDest, const uint8_t *__restrict pSrc)
{
// First 8 KB of the source block is ODD bytes.
const uint8_t *pSrc_end = pSrc + (SMD_BLOCK_SIZE / 2);
for (uint8_t *pDest_odd = pDest + 1; pSrc < pSrc_end; pDest_odd += 2,
pSrc += 1) {
pDest_odd[0] = pSrc[0];
}
}
Assembly output with `g++ -O2 -fno-tree-vectorize` (or removing the __restrict
qualifiers):
decodeBlock_cpp(unsigned char*, unsigned char const*):
xor eax, eax
.L2:
movzx edx, BYTE PTR [rsi+rax]
mov BYTE PTR [rdi+1+rax*2], dl
add rax, 1
cmp rax, 8192
jne .L2
ret
Assembly output with `g++ -O2` (implying -ftree-vectorize with gcc-12) and
__restrict qualifiers:
decodeBlock_cpp(unsigned char*, unsigned char const*):
push r15
lea rax, [rsi+8192]
add rdi, 1
push r14
push r13
push r12
push rbp
push rbx
mov QWORD PTR [rsp-8], rax
.L2:
movzx ecx, BYTE PTR [rsi+10]
movzx eax, BYTE PTR [rsi+14]
add rsi, 16
add rdi, 32
movzx edx, BYTE PTR [rsi-3]
movzx r15d, BYTE PTR [rsi-1]
movzx r11d, BYTE PTR [rsi-10]
movzx ebx, BYTE PTR [rsi-11]
mov BYTE PTR [rsp-11], cl
movzx ecx, BYTE PTR [rsi-16]
movzx ebp, BYTE PTR [rsi-12]
mov BYTE PTR [rsp-9], al
movzx r12d, BYTE PTR [rsi-13]
movzx eax, BYTE PTR [rsi-4]
mov BYTE PTR [rsp-10], dl
movzx r13d, BYTE PTR [rsi-14]
movzx edx, BYTE PTR [rsi-5]
movzx r14d, BYTE PTR [rsi-15]
movzx r8d, BYTE PTR [rsi-7]
movzx r9d, BYTE PTR [rsi-8]
movzx r10d, BYTE PTR [rsi-9]
mov BYTE PTR [rdi-32], cl
movzx ecx, BYTE PTR [rsp-11]
mov BYTE PTR [rdi-10], dl
mov BYTE PTR [rdi-30], r14b
mov BYTE PTR [rdi-28], r13b
mov BYTE PTR [rdi-26], r12b
mov BYTE PTR [rdi-24], bpl
mov BYTE PTR [rdi-22], bl
mov BYTE PTR [rdi-20], r11b
mov BYTE PTR [rdi-18], r10b
mov BYTE PTR [rdi-16], r9b
mov BYTE PTR [rdi-14], r8b
mov BYTE PTR [rdi-12], cl
mov BYTE PTR [rdi-8], al
movzx eax, BYTE PTR [rsp-9]
movzx edx, BYTE PTR [rsp-10]
mov BYTE PTR [rdi-2], r15b
mov BYTE PTR [rdi-4], al
mov rax, QWORD PTR [rsp-8]
mov BYTE PTR [rdi-6], dl
cmp rsi, rax
jne .L2
pop rbx
pop rbp
pop r12
pop r13
pop r14
pop r15
ret
$ gcc --version
gcc (Gentoo Hardened 12.2.1_p20221008 p1) 12.2.1 20221008
Copyright (C) 2022 Free Software Foundation, Inc.
This is free software; see the source for copying conditions. There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
