https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85640
Bug ID: 85640
Summary: Code size regression vs 7.3.1
Product: gcc
Version: 8.1.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c++
Assignee: unassigned at gcc dot gnu.org
Reporter: petschy at gmail dot com
Target Milestone: ---
Created attachment 44062
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=44062&action=edit
source
Attached the source of a simple Adler32 checksum class. The Update() fn is 32
bytes longer compared to the code generated with 7.3.1.
Dump of assembler code for function Adler32::Update(void const*, unsigned int):
7.3.1 0x0000000000400500 <+0>: test %edx,%edx
7.3.1 0x0000000000400502 <+2>: je 0x400578 <Adler32::Update(void
const*, unsigned int)+120>
7.3.1 0x0000000000400504 <+4>: mov (%rdi),%ecx
7.3.1 0x0000000000400506 <+6>: mov 0x4(%rdi),%r8d
7.3.1 0x000000000040050a <+10>: mov $0x80078071,%r10d
7.3.1 0x0000000000400510 <+16>: xor %r9d,%r9d
7.3.1 0x0000000000400513 <+19>: cmp $0x15af,%edx
7.3.1 0x0000000000400519 <+25>: jbe 0x400527 <Adler32::Update(void
const*, unsigned int)+39>
7.3.1 0x000000000040051b <+27>: lea -0x15b0(%rdx),%r9d
7.3.1 0x0000000000400522 <+34>: mov $0x15b0,%edx
7.3.1 0x0000000000400527 <+39>: lea -0x1(%rdx),%eax
7.3.1 0x000000000040052a <+42>: lea 0x1(%rsi,%rax,1),%rdx
7.3.1 0x000000000040052f <+47>: nop
7.3.1 0x0000000000400530 <+48>: add $0x1,%rsi
7.3.1 0x0000000000400534 <+52>: movzbl -0x1(%rsi),%eax
7.3.1 0x0000000000400538 <+56>: add %eax,%ecx
7.3.1 0x000000000040053a <+58>: add %ecx,%r8d
7.3.1 0x000000000040053d <+61>: cmp %rdx,%rsi
7.3.1 0x0000000000400540 <+64>: mov %ecx,(%rdi)
7.3.1 0x0000000000400542 <+66>: mov %r8d,0x4(%rdi)
7.3.1 0x0000000000400546 <+70>: jne 0x400530 <Adler32::Update(void
const*, unsigned int)+48>
7.3.1 0x0000000000400548 <+72>: mov %ecx,%eax
7.3.1 0x000000000040054a <+74>: mul %r10d
7.3.1 0x000000000040054d <+77>: mov %r8d,%eax
7.3.1 0x0000000000400550 <+80>: shr $0xf,%edx
7.3.1 0x0000000000400553 <+83>: imul $0xfff1,%edx,%edx
7.3.1 0x0000000000400559 <+89>: sub %edx,%ecx
7.3.1 0x000000000040055b <+91>: mul %r10d
7.3.1 0x000000000040055e <+94>: mov %ecx,(%rdi)
7.3.1 0x0000000000400560 <+96>: shr $0xf,%edx
7.3.1 0x0000000000400563 <+99>: imul $0xfff1,%edx,%edx
7.3.1 0x0000000000400569 <+105>: sub %edx,%r8d
7.3.1 0x000000000040056c <+108>: test %r9d,%r9d
7.3.1 0x000000000040056f <+111>: mov %r9d,%edx
7.3.1 0x0000000000400572 <+114>: mov %r8d,0x4(%rdi)
7.3.1 0x0000000000400576 <+118>: jne 0x400510 <Adler32::Update(void
const*, unsigned int)+16>
7.3.1 0x0000000000400578 <+120>: repz retq
Dump of assembler code for function Adler32::Update(void const*, unsigned int):
8.1.1 0x0000000000400500 <+0>: test %edx,%edx
8.1.1 0x0000000000400502 <+2>: je 0x400598 <Adler32::Update(void
const*, unsigned int)+152>
8.1.1 0x0000000000400508 <+8>: mov (%rdi),%ecx
8.1.1 0x000000000040050a <+10>: mov 0x4(%rdi),%r8d
8.1.1 0x000000000040050e <+14>: push %rbx
8.1.1 0x000000000040050f <+15>: mov $0x80078071,%ebx
8.1.1 0x0000000000400514 <+20>: nopl 0x0(%rax)
8.1.1 0x0000000000400518 <+24>: xor %r11d,%r11d
8.1.1 0x000000000040051b <+27>: cmp $0x15af,%edx
8.1.1 0x0000000000400521 <+33>: jbe 0x40052f <Adler32::Update(void
const*, unsigned int)+47>
8.1.1 0x0000000000400523 <+35>: lea -0x15b0(%rdx),%r11d
8.1.1 0x000000000040052a <+42>: mov $0x15b0,%edx
8.1.1 0x000000000040052f <+47>: mov %edx,%r10d
8.1.1 0x0000000000400532 <+50>: mov %rsi,%rax
8.1.1 0x0000000000400535 <+53>: add %rsi,%r10
8.1.1 0x0000000000400538 <+56>: nopl 0x0(%rax,%rax,1)
8.1.1 0x0000000000400540 <+64>: add $0x1,%rax
8.1.1 0x0000000000400544 <+68>: movzbl -0x1(%rax),%r9d
8.1.1 0x0000000000400549 <+73>: add %r9d,%ecx
8.1.1 0x000000000040054c <+76>: add %ecx,%r8d
8.1.1 0x000000000040054f <+79>: mov %ecx,(%rdi)
8.1.1 0x0000000000400551 <+81>: mov %r8d,0x4(%rdi)
8.1.1 0x0000000000400555 <+85>: cmp %r10,%rax
8.1.1 0x0000000000400558 <+88>: jne 0x400540 <Adler32::Update(void
const*, unsigned int)+64>
8.1.1 0x000000000040055a <+90>: lea -0x1(%rdx),%eax
8.1.1 0x000000000040055d <+93>: lea 0x1(%rsi,%rax,1),%rsi
8.1.1 0x0000000000400562 <+98>: mov %ecx,%eax
8.1.1 0x0000000000400564 <+100>: mul %ebx
8.1.1 0x0000000000400566 <+102>: mov %r8d,%eax
8.1.1 0x0000000000400569 <+105>: shr $0xf,%edx
8.1.1 0x000000000040056c <+108>: imul $0xfff1,%edx,%edx
8.1.1 0x0000000000400572 <+114>: sub %edx,%ecx
8.1.1 0x0000000000400574 <+116>: mul %ebx
8.1.1 0x0000000000400576 <+118>: mov %ecx,(%rdi)
8.1.1 0x0000000000400578 <+120>: shr $0xf,%edx
8.1.1 0x000000000040057b <+123>: imul $0xfff1,%edx,%edx
8.1.1 0x0000000000400581 <+129>: sub %edx,%r8d
8.1.1 0x0000000000400584 <+132>: mov %r11d,%edx
8.1.1 0x0000000000400587 <+135>: mov %r8d,0x4(%rdi)
8.1.1 0x000000000040058b <+139>: test %r11d,%r11d
8.1.1 0x000000000040058e <+142>: jne 0x400518 <Adler32::Update(void
const*, unsigned int)+24>
8.1.1 0x0000000000400590 <+144>: pop %rbx
8.1.1 0x0000000000400591 <+145>: retq
8.1.1 0x0000000000400592 <+146>: nopw 0x0(%rax,%rax,1)
8.1.1 0x0000000000400598 <+152>: retq
Here is an interwoven version, hopefully easier to follow:
7.3.1 0x0000000000400500 <+0>: test %edx,%edx
7.3.1 0x0000000000400502 <+2>: je 0x400578 <Adler32::Update(void
const*, unsigned int)+120>
7.3.1 0x0000000000400504 <+4>: mov (%rdi),%ecx
7.3.1 0x0000000000400506 <+6>: mov 0x4(%rdi),%r8d
7.3.1 0x000000000040050a <+10>: mov $0x80078071,%r10d
8.1.1 0x0000000000400500 <+0>: test %edx,%edx
8.1.1 0x0000000000400502 <+2>: je 0x400598 <Adler32::Update(void
const*, unsigned int)+152>
8.1.1 0x0000000000400508 <+8>: mov (%rdi),%ecx
8.1.1 0x000000000040050a <+10>: mov 0x4(%rdi),%r8d
8.1.1 0x000000000040050e <+14>: push %rbx
8.1.1 0x000000000040050f <+15>: mov $0x80078071,%ebx
8.1.1 0x0000000000400514 <+20>: nopl 0x0(%rax)
Two things so far:
- the je is 6 bytes in 8.1.1 vs 2 bytes in 7.3.1 because the jump offset can't
fit in a byte
- in 8.1.1 ebx is used for the modulo magic, which is callee saved, so have to
push before use
7.3.1 0x0000000000400510 <+16>: xor %r9d,%r9d
7.3.1 0x0000000000400513 <+19>: cmp $0x15af,%edx
7.3.1 0x0000000000400519 <+25>: jbe 0x400527 <Adler32::Update(void
const*, unsigned int)+39>
7.3.1 0x000000000040051b <+27>: lea -0x15b0(%rdx),%r9d
7.3.1 0x0000000000400522 <+34>: mov $0x15b0,%edx
7.3.1 0x0000000000400527 <+39>: lea -0x1(%rdx),%eax
7.3.1 0x000000000040052a <+42>: lea 0x1(%rsi,%rax,1),%rdx
7.3.1 0x000000000040052f <+47>: nop
8.1.1 0x0000000000400518 <+24>: xor %r11d,%r11d
8.1.1 0x000000000040051b <+27>: cmp $0x15af,%edx
8.1.1 0x0000000000400521 <+33>: jbe 0x40052f <Adler32::Update(void
const*, unsigned int)+47>
8.1.1 0x0000000000400523 <+35>: lea -0x15b0(%rdx),%r11d
8.1.1 0x000000000040052a <+42>: mov $0x15b0,%edx
8.1.1 0x000000000040052f <+47>: mov %edx,%r10d
8.1.1 0x0000000000400532 <+50>: mov %rsi,%rax
8.1.1 0x0000000000400535 <+53>: add %rsi,%r10
8.1.1 0x0000000000400538 <+56>: nopl 0x0(%rax,%rax,1)
This is the inner loop init, pretty similar.
7.3.1 0x0000000000400530 <+48>: add $0x1,%rsi
7.3.1 0x0000000000400534 <+52>: movzbl -0x1(%rsi),%eax
7.3.1 0x0000000000400538 <+56>: add %eax,%ecx
7.3.1 0x000000000040053a <+58>: add %ecx,%r8d
7.3.1 0x000000000040053d <+61>: cmp %rdx,%rsi
7.3.1 0x0000000000400540 <+64>: mov %ecx,(%rdi)
7.3.1 0x0000000000400542 <+66>: mov %r8d,0x4(%rdi)
7.3.1 0x0000000000400546 <+70>: jne 0x400530 <Adler32::Update(void
const*, unsigned int)+48>
8.1.1 0x0000000000400540 <+64>: add $0x1,%rax
8.1.1 0x0000000000400544 <+68>: movzbl -0x1(%rax),%r9d
8.1.1 0x0000000000400549 <+73>: add %r9d,%ecx
8.1.1 0x000000000040054c <+76>: add %ecx,%r8d
8.1.1 0x000000000040054f <+79>: mov %ecx,(%rdi)
8.1.1 0x0000000000400551 <+81>: mov %r8d,0x4(%rdi)
8.1.1 0x0000000000400555 <+85>: cmp %r10,%rax
8.1.1 0x0000000000400558 <+88>: jne 0x400540 <Adler32::Update(void
const*, unsigned int)+64>
These are the same, except the cmp is before/after the two stores, and the
movzbl and the first add is one byte cheaper in 7.3.1.
7.3.1 0x0000000000400548 <+72>: mov %ecx,%eax
7.3.1 0x000000000040054a <+74>: mul %r10d
7.3.1 0x000000000040054d <+77>: mov %r8d,%eax
7.3.1 0x0000000000400550 <+80>: shr $0xf,%edx
7.3.1 0x0000000000400553 <+83>: imul $0xfff1,%edx,%edx
7.3.1 0x0000000000400559 <+89>: sub %edx,%ecx
7.3.1 0x000000000040055b <+91>: mul %r10d
7.3.1 0x000000000040055e <+94>: mov %ecx,(%rdi)
7.3.1 0x0000000000400560 <+96>: shr $0xf,%edx
7.3.1 0x0000000000400563 <+99>: imul $0xfff1,%edx,%edx
7.3.1 0x0000000000400569 <+105>: sub %edx,%r8d
7.3.1 0x000000000040056c <+108>: test %r9d,%r9d
7.3.1 0x000000000040056f <+111>: mov %r9d,%edx
7.3.1 0x0000000000400572 <+114>: mov %r8d,0x4(%rdi)
7.3.1 0x0000000000400576 <+118>: jne 0x400510 <Adler32::Update(void
const*, unsigned int)+16>
7.3.1 0x0000000000400578 <+120>: repz retq
8.1.1 0x000000000040055a <+90>: lea -0x1(%rdx),%eax
8.1.1 0x000000000040055d <+93>: lea 0x1(%rsi,%rax,1),%rsi
8.1.1 0x0000000000400562 <+98>: mov %ecx,%eax
8.1.1 0x0000000000400564 <+100>: mul %ebx
8.1.1 0x0000000000400566 <+102>: mov %r8d,%eax
8.1.1 0x0000000000400569 <+105>: shr $0xf,%edx
8.1.1 0x000000000040056c <+108>: imul $0xfff1,%edx,%edx
8.1.1 0x0000000000400572 <+114>: sub %edx,%ecx
8.1.1 0x0000000000400574 <+116>: mul %ebx
8.1.1 0x0000000000400576 <+118>: mov %ecx,(%rdi)
8.1.1 0x0000000000400578 <+120>: shr $0xf,%edx
8.1.1 0x000000000040057b <+123>: imul $0xfff1,%edx,%edx
8.1.1 0x0000000000400581 <+129>: sub %edx,%r8d
8.1.1 0x0000000000400584 <+132>: mov %r11d,%edx
8.1.1 0x0000000000400587 <+135>: mov %r8d,0x4(%rdi)
8.1.1 0x000000000040058b <+139>: test %r11d,%r11d
8.1.1 0x000000000040058e <+142>: jne 0x400518 <Adler32::Update(void
const*, unsigned int)+24>
8.1.1 0x0000000000400590 <+144>: pop %rbx
8.1.1 0x0000000000400591 <+145>: retq
8.1.1 0x0000000000400592 <+146>: nopw 0x0(%rax,%rax,1)
8.1.1 0x0000000000400598 <+152>: retq
The loop variables in 8.1.1 handled a bit differently, that's why the minor
size increase. The two lea's here in the 8.1.1 version are very similar to the
ones before the inner loop in 7.3.1, those calculate the ptr range of the inner
loop. These lea's here calculate the start ptr for the next inner loop run,
however, this ptr is already present in eax. I don't know if this is some
special optimization, or eax should have been used but the compiler missed it.
The double retq at the end is interesting, but maybe the debug info is to
blame.
cmdline used to compile:
$ g++-7.3.1 -g -O3 -Wall a32.cpp
$ g++-8.1.1 -g -O3 -Wall a32.cpp
Platform is AMD64 (FX-8150), Debian 9.4
$ g++-7.3.1 -v
Using built-in specs.
COLLECT_GCC=g++-7.3.1
COLLECT_LTO_WRAPPER=/usr/local/libexec/gcc/x86_64-pc-linux-gnu/7.3.1/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: ../configure --enable-languages=c,c++ --disable-multilib
--program-suffix=-7.3.1 --disable-bootstrap CFLAGS='-O2 -march=native
-mtune=native' CXXFLAGS='-O2 -march=native -mtune=native'
Thread model: posix
gcc version 7.3.1 20180429 (GCC)
$ g++-8.1.1 -v
Using built-in specs.
COLLECT_GCC=g++-8.1.1
COLLECT_LTO_WRAPPER=/usr/local/libexec/gcc/x86_64-pc-linux-gnu/8.1.1/lto-wrapper
Target: x86_64-pc-linux-gnu
Configured with: ../configure --enable-languages=c,c++ --disable-multilib
--program-suffix=-8.1.1 --disable-bootstrap CFLAGS='-O2 -march=native
-mtune=native' CXXFLAGS='-O2 -march=native -mtune=native'
Thread model: posix
gcc version 8.1.1 20180502 (GCC)
