https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100331
Bug ID: 100331
Summary: 128 bit arithmetic --- suboptimal after shifting when
referencing other variables
Product: gcc
Version: 9.3.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: zero at smallinteger dot com
Target Milestone: ---
Created attachment 50706
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=50706&action=edit
Reproduction test case
Compile the given C program with -O2. Enabling the #if 1 branch results in:
compute(unsigned long, unsigned long):
mov ecx, edi
xor edx, edx
mov rax, rsi
xor esi, esi
and ecx, 63
shrd rax, rdx, cl
shr rdx, cl
test cl, 64
mov r8d, ecx
cmovne rax, rdx
cmovne rdx, rsi
and r8d, 63
mov rsi, rax
mov rax, r8
mov rdi, rdx
xor edx, edx
add rax, rsi
adc rdx, rdi
ret
Note test cl, 64 and the subsequent cmovs are unnecessary because the result of
the test is already known after and ecx, 63. Note also mov r8d, ecx, followed
by and r8d, 63, redoing work.
Enabling the #if 0 branch results in this code instead.
compute(unsigned long, unsigned long):
mov rcx, rdi
xor edx, edx
mov rax, rsi
shrd rax, rdx, cl
shr rdx, cl
ret
That is, now gcc realizes the range of possible values for cl and does not emit
the test, the cmovs, and the redoing of the and with r8d. One way or another,
the double precision shift is also unnecessary because only the lower 64 bits
of result may be non-zero.
Verified on Ubuntu 20.04 LTS, as well as Godbolt with gcc 9.3.0, gcc 11.1, and
gcc trunk. This issue is similar to other 128 bit arithmetic reported bugs,
but unlike those others this one seems to be controlled exclusively by the
addition in the #if 1 branch.
For the sake of comparison, clang trunk emits the following code for the #if 1
and #if 0 branches, as per Godbolt.
compute(unsigned long, unsigned long): #
@compute(unsigned long, unsigned long)
mov rcx, rdi
mov eax, ecx
shr rsi, cl
and eax, 63
xor edx, edx
add rax, rsi
setb dl
ret
compute(unsigned long, unsigned long): #
@compute(unsigned long, unsigned long)
mov rax, rsi
mov rcx, rdi
shr rax, cl
xor edx, edx
ret
