| Summary: | inefficient use of registers induces size and time overhead | ||
|---|---|---|---|
| Product: | gcc | Reporter: | willy tarreau <willy> |
| Component: | rtl-optimization | Assignee: | Not yet assigned to anyone <unassigned> |
| Status: | NEW --- | ||
| Severity: | enhancement | CC: | gcc-bugs, ian |
| Priority: | P2 | Keywords: | missed-optimization, ra |
| Version: | 3.3.1 | ||
| Target Milestone: | --- | ||
| Host: | Target: | i586-linux-gnu | |
| Build: | Known to work: | ||
| Known to fail: | Last reconfirmed: | 2011-05-22 16:27:10 | |
| Bug Depends on: | 15792 | ||
| Bug Blocks: | |||
I can confirm this on the mainline (20030809). GCC is not really good at optimizing long long's, I have some improvements but it seems not to help in this case. I filed 15792 to track part of this bug. I don't know why this was put in waiting but it should not have been. This has improved (-O2 -fomit-frame-pointer): test: movl 4(%esp), %eax # 32 *movsi_1/1 [length = 4] movl 8(%esp), %edx # 44 *movsi_1/1 [length = 4] orl %eax, %edx # 6 *iorsi_1/1 [length = 2] addl $1, %eax # 35 *addsi_1/1 [length = 3] cmpl $1, %edx # 38 *cmpsi_1_insn/1 [length = 3] sbbl %edx, %edx # 39 x86_movsicc_0_m1 [length = 2] notl %edx # 40 *one_cmplsi2_1 [length = 2] andl %edx, %eax # 41 *andsi_1/1 [length = 2] ret # 47 return_internal [length = 1] .ident "GCC: (GNU) 4.3.0 20071102 (experimental)" With -Os -fomit-frame-pointer we get: test: movl 4(%esp), %edx # 32 *movsi_1/1 [length = 4] xorl %eax, %eax # 48 *movsi_xor [length = 2] movl 8(%esp), %ecx # 43 *movsi_1/1 [length = 4] orl %edx, %ecx # 7 *iorsi_3 [length = 2] je .L3 # 8 *jcc_1 [length = 2] leal 1(%edx), %eax # 44 *lea_1 [length = 3] .L3: ret # 47 return_internal [length = 1] With -O2/-Os -fomit-frame-pointer -march=pentiumpro: test: movl 4(%esp), %edx # 32 *movsi_1/1 [length = 4] xorl %eax, %eax # 46 *movsi_xor [length = 2] leal 1(%edx), %ecx # 41 *lea_1 [length = 3] orl 8(%esp), %edx # 36 *iorsi_3 [length = 4] cmovne %ecx, %eax # 38 *movsicc_noc/1 [length = 3] ret # 44 return_internal [length = 1] I would probably code it like so: movl 4(%esp), %eax ; 4 movl 8(%esp), %edx ; 4 orl %eax, %edx ; 2 addl $-1, %edx ; 3 adcl $0, %eax ; 3 ret ; 1 |
I wrote a simple test function which uses an unsigned long long arg and returns its integer part + 1 if not 0, otherwise 0. GCC 3.3.1 passes some intermediate values through several registers while it's unneeded. Here comes all my analysis. I hope this can help improving the optimizer. The following code : int test(unsigned long long x) { if (x) { return (int)x + 1; } else { return (int)x; } } processed this way : # 1 "bool.c" # 1 "<built-in>" # 1 "<command line>" # 1 "bool.c" int test(unsigned long long x) { if (x) { return (int)x + 1; } else { return (int)x; } } produces the following code when compiled with gcc 2.95.3 : (gcc -c -O2 -fomit-frame-pointer bool.c) 00000000 <test>: 0: 8b 54 24 04 mov 0x4(%esp,1),%edx 4: 8b 4c 24 08 mov 0x8(%esp,1),%ecx 8: 89 d0 mov %edx,%eax a: 09 c8 or %ecx,%eax c: 75 03 jne 11 <test+0x11> e: 89 d0 mov %edx,%eax 10: c3 ret 11: 8d 42 01 lea 0x1(%edx),%eax 14: c3 ret 15: 8d 76 00 lea 0x0(%esi),%esi 18: and this one with gcc-3.3.1 : (gcc-3.3.1 -c -O2 -fomit-frame-pointer bool.c) 00000000 <test>: 0: 53 push %ebx 1: 8b 4c 24 0c mov 0xc(%esp,1),%ecx 5: 8b 54 24 08 mov 0x8(%esp,1),%edx 9: 89 cb mov %ecx,%ebx b: 89 d0 mov %edx,%eax d: 09 d3 or %edx,%ebx f: 74 03 je 14 <test+0x14> 11: 8d 42 01 lea 0x1(%edx),%eax 14: 5b pop %ebx 15: c3 ret 16: 8d 76 00 lea 0x0(%esi),%esi 19: 8d bc 27 00 00 00 00 lea 0x0(%edi,1),%edi 20: the EBX register is used and clobbered for nothing here. The same code could be written this way, which is fully equivalent and saves some cycles and bytes : 00000000 <test>: 0: 8b 4c 24 0c mov 0xc(%esp,1),%ecx 4: 8b 54 24 08 mov 0x8(%esp,1),%edx 8: 89 d0 mov %edx,%eax a: 09 d1 or %edx,%ecx c: 74 03 je 11 <test+0x11> e: 8d 42 01 lea 0x1(%edx),%eax 11: c3 ret 12: Now we can also save EDX and some more bytes : 00000000 <test>: 0: 8b 4c 24 0c mov 0xc(%esp,1),%ecx 4: 8b 54 24 08 mov 0x8(%esp,1),%eax a: 09 c1 or %eax,%ecx c: 74 01 je f <test+0xf> e: 41 inc %eax f: c3 ret 10: Here are the compilers versions : wtap:~/dev$ gcc -v Reading specs from /usr/lib/gcc-lib/i586-pc-linux-gnu/2.95.3/specs gcc version 2.95.3 20010315 (release) wtap:~/dev$ gcc-3.3.1 -v Reading specs from /usr/lib/gcc-lib/i586-pc-linux-gnu/3.3.1/specs Configured with: ./configure --prefix=/usr --with-cpu=i386 --host=i586-pc-linux-gnu --enable-languages=c,c++ --disable-nls --disable-locale --enable-shared --enable-target-optspace --enable-version-specific-runtime-libs --program-suffix=-3.3.1 --enable-threads Thread model: posix gcc version 3.3.1