[Bug rtl-optimization/45434] x86 missed optimization: use high register (ah, bh, ch, dh) when available to make comparisons
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45434
--- Comment #7 from Adam Warner adam at consulting dot net.nz 2012-01-18
03:35:24 UTC ---
It does not appear to be possible to generate inline asm that leaves GCC to
choose the {ah, bh, ch, dh} register:
#include stdint.h
uint64_t u8l(uint64_t in) {
uint64_t out;
asm (movzbl %b1, %k0 : =q (out) : r (in));
return out;
}
uint64_t u8h(uint64_t in) {
uint64_t out;
asm (movzbl %h1, %k0 : =Q (out) : r (in)); //line number 11
return out;
}
int main(void) {
return 0;
}
$ gcc -O3 ah_bh_ch_dh.c
ah_bh_ch_dh.c: Assembler messages:
ah_bh_ch_dh.c:11: Error: operand type mismatch for `movzbl'
I found the h qualifier by reading:
www.cs.virginia.edu/~clc5q/gcc-inline-asm.pdf
%b0 prints the 8-bit form of an operand. %al, etc.
%h0 prints the high 8-bit form of a register. %ah, etc.
Q is the x86 machine constraint which means Any register accessible as rh: a,
b, c, and d.
http://gcc.gnu.org/onlinedocs/gcc/Machine-Constraints.html
Additionally, this inline assembly instruction causes LLVM-based clang to
segfault and request a bug submission report.
I haven't located the primary documentation for these %b, %h, %k, etc.
qualifiers. Is my usage of %h correct/is this actually a bug in GCC?
[Bug rtl-optimization/45434] x86 missed optimization: use high register (ah, bh, ch, dh) when available to make comparisons
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45434
--- Comment #8 from Adam Warner adam at consulting dot net.nz 2012-01-18
07:00:44 UTC ---
Apologies for the noise. I had my machine constraints around the wrong way.
Here is the fixed version of the code:
#include stdint.h
uint64_t u8l(uint64_t in) {
uint64_t out;
asm (movzbl %b1, %k0 : =r (out) : q (in));
return out;
}
uint64_t u8h(uint64_t in) {
uint64_t out;
asm (movzbl %h1, %k0 : =r (out) : Q (in));
return out;
}
int main(void) {
return 0;
}
With the X86-64 Linux ABI u8l() generates movzbl %dil,%eax and u8h() mov
%rdi,%rdx; movzbl %dh,%eax as expected.
[Bug rtl-optimization/45434] x86 missed optimization: use high register (ah, bh, ch, dh) when available to make comparisons
--- Comment #2 from ubizjak at gmail dot com 2010-08-29 16:40 ---
A bit simpler testcase:
int test(int i)
{
return (i 0xFF) == ((i 0xFF00) 8);
}
fails two combine attempts:
Trying 8 - 9:
Failed to match this instruction:
(set (reg:CCZ 17 flags)
(compare:CCZ (zero_extract:SI (subreg:DI (reg/v:SI 63 [ i ]) 0)
(const_int 8 [0x8])
(const_int 8 [0x8]))
(reg:SI 65)))
Trying 8, 6 - 9:
Failed to match this instruction:
(set (reg:CCZ 17 flags)
(compare:CCZ (subreg:QI (lshiftrt:SI (reg/v:SI 63 [ i ])
(const_int 8 [0x8])) 0)
(subreg:QI (reg/v:SI 63 [ i ]) 0)))
However, contrary to the first attempt, 8 - 9 substitution in the second
attempt is over-optimized from ZERO_EXTRACT to LSHIFTRT. i386.md expects
ZERO_EXTRACT form.
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45434
[Bug rtl-optimization/45434] x86 missed optimization: use high register (ah, bh, ch, dh) when available to make comparisons
--- Comment #3 from hjl dot tools at gmail dot com 2010-08-29 18:34 --- Combine seems to prefer LSHIFTRT over ZERO_EXTRACT. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45434
[Bug rtl-optimization/45434] x86 missed optimization: use high register (ah, bh, ch, dh) when available to make comparisons
--- Comment #4 from ubizjak at gmail dot com 2010-08-29 18:59 ---
(In reply to comment #3)
Combine seems to prefer LSHIFTRT over ZERO_EXTRACT.
It seems so. Slightly changed testcase:
char test(int a, int b, int i, int j)
{
return (i 0xFF) == ((j 0xFF00) 8);
}
and changed *cmpqi_ext_3_insn_rex64 pattern:
(define_insn *cmpqi_ext_3_insn_rex64
[(set (reg FLAGS_REG)
(compare
(subreg:QI
(lshiftrt:SI
;; (zero_extract:SI
(match_operand 0 ext_register_operand Q)
;;(const_int 8)
(const_int 8)) 0)
(match_operand:QI 1 nonmemory_operand Qn)))]
TARGET_64BIT ix86_match_ccmode (insn, CCmode)
cmp{b}\t{%1, %h0|%h0, %1}
[(set_attr type icmp)
(set_attr modrm 1)
(set_attr mode QI)])
result in:
cmpb%dl, %ch
sete%al
ret
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45434
[Bug rtl-optimization/45434] x86 missed optimization: use high register (ah, bh, ch, dh) when available to make comparisons
--- Comment #5 from ebotcazou at gcc dot gnu dot org 2010-08-29 19:12 --- Combine seems to prefer LSHIFTRT over ZERO_EXTRACT. The former is the canonical form of the latter for internal computations, but the latter should be reinstantiated when matching instructions. -- http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45434
[Bug rtl-optimization/45434] x86 missed optimization: use high register (ah, bh, ch, dh) when available to make comparisons
--- Comment #6 from ubizjak at gmail dot com 2010-08-29 19:20 ---
Some other examples:
char test(int a, int b, int i, int j)
{
return (i 0xFF) + ((j 0xFF00) 8);
}
movzbl %ch, %eax
addl%edx, %eax
ret
char test(int a, int b, int i, int j)
{
return (i 0xFF) - ((j 0xFF00) 8);
}
movzbl %ch, %ecx
movl%edx, %eax
subb%cl, %al
--
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45434
[Bug rtl-optimization/45434] x86 missed optimization: use high register (ah, bh, ch, dh) when available to make comparisons
--- Comment #1 from hjl dot tools at gmail dot com 2010-08-28 04:30 --- It may be related to PR 44659. -- hjl dot tools at gmail dot com changed: What|Removed |Added CC||ebotcazou at gcc dot gnu dot ||org http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45434
[Bug rtl-optimization/45434] x86 missed optimization: use high register (ah, bh, ch, dh) when available to make comparisons
-- hjl dot tools at gmail dot com changed: What|Removed |Added CC||hjl dot tools at gmail dot ||com Status|UNCONFIRMED |NEW Ever Confirmed|0 |1 Last reconfirmed|-00-00 00:00:00 |2010-08-28 04:35:30 date|| http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45434
