[RISCV] Add new tests for RISCV zicond extension #143580
Conversation
|
Please test |
We plan to remove 'xventanacondops' and just go with zicondops |
Did xventanacondops ever ship in real hardware? |
No |
bababuck
force-pushed
the
bababuck/ZicondOptTests
branch
from
2a23721 to
98eed72
Compare
|
Pushed to address the above suggestions. |
bababuck
force-pushed
the
bababuck/ZicondOptTests
branch
from
98eed72 to
7d64e5c
Compare
| } | ||
|
|
||
| ; (select cond, -2048, t) -> (xor (czero_nez (xor t, -2048), cond), -2048) | ||
| define i64 @select_imm_reg_2048(i64 %0, i64 %1) { |
There was a problem hiding this comment.
I believe the constant operand of select will alway be last due to canonicalization of IR in the midend
There was a problem hiding this comment.
Isn't rhs same cost as lhs?
There was a problem hiding this comment.
Before:
%imm = li -2048
%res0 = czero.eqz %imm, %cond
%res1 = czero.nez %input, %cond
%return = or %res0, %res1
ret %return
After:
%xor = xori %input, -2048
%czero = czero.nez, %input, %cond
%return = xori %czero, -2048
ret %return
DAG of height 3 for both, transformed has one fewer instruction
There was a problem hiding this comment.
I believe the constant operand of select will alway be last due to canonicalization of IR in the midend
Will this always be true, I though the order of %select was dependent on the condition, or will the mid-end flip the SETCC condition code so that this is true?
There was a problem hiding this comment.
I believe the constant operand of select will alway be last due to canonicalization of IR in the midend
I don't want to rely on this. Selects can be created in the backend and I don't think we have the same canonicalization in DAGCombine, but I might be wrong.
There was a problem hiding this comment.
Oh, sorry, my bad. I don't know what I was thinking when I wrote this.
| ret i64 %4 | ||
| } | ||
|
|
||
| ; (select c, (and f, ~x), f) -> (andn f, (czero_nez x, c)) |
There was a problem hiding this comment.
note that "0" is identity for "andn" so in terms of implementation this case could be combined with conditional rotation.
bababuck
force-pushed
the
bababuck/ZicondOptTests
branch
from
7d64e5c to
cff8037
Compare
|
Updated:
|
bababuck
force-pushed
the
bababuck/ZicondOptTests
branch
from
cff8037 to
d3f5efa
Compare
|
Pushed one last time, I had a small typo in the rotl test. |
I have a few patches to improve compilation for these tests which I will be posting as separate MRs.
llvm#143581) See llvm#143580 for MR with the test commit. Performs the following transformations: (select c, c1, t) -> (add (czero_nez t - c1, c), c1) (select c, t, c1) -> (add (czero_eqz t - c1, c), c1) @mgudim
I have a few patches to improve compilation for these tests which I will be posting as separate MRs.
llvm#143581) See llvm#143580 for MR with the test commit. Performs the following transformations: (select c, c1, t) -> (add (czero_nez t - c1, c), c1) (select c, t, c1) -> (add (czero_eqz t - c1, c), c1) @mgudim
I have a few patches to improve compilation for these tests which I will be posting as separate MRs.