We have this optimization in if-conversion that turns SELECT(cond, 1 << a, 0) into cond << a. It only runs on RISCV:
|
// TryTransformSelectToOrdinaryOps: Try transforming the identified if-else expressions to a single expression |
|
// |
|
// This is meant mostly for RISC-V where the condition (1 or 0) is stored in a regular general-purpose register |
|
// which can be fed as an argument to standard operations, e.g. |
|
// * (cond ? 6 : 5) becomes (5 + cond) |
|
// * (cond ? -25 : -13) becomes (-25 >> cond) |
|
// * if (cond) a++; becomes (a + cond) |
|
// * (cond ? 1 << a : 0) becomes (cond << a) |
However our current cost-analysis always rejects such patterns. Example:
int Cond1Shift(bool cond, int a)
{
return cond ? (1 << a) : 0;
}If we look at JitDump:
***** BB03 [0002]
STMT00001 ( 0x005[E--] ... 0x00B )
N006 ( 11, 9) [000009] -----+----- * RETURN int $VN.Void
N005 ( 10, 8) [000008] -----+--R-- \--* LSH int $143
N004 ( 1, 2) [000004] -----+----- +--* CNS_INT int 1 $41
N003 ( 5, 5) [000007] -----+----- \--* AND int $142
N001 ( 3, 2) [000005] -----+----- +--* LCL_VAR int V01 arg1 u:1 (last use) $c0
N002 ( 1, 2) [000006] -----+----- \--* CNS_INT int 31 $44
Skipping if-conversion that will evaluate RHS unconditionally at costs 1,10
So the optimization above for RISCV is essentially dead code. We always bail out earlier which is probably not wanted.
We have this optimization in if-conversion that turns
SELECT(cond, 1 << a, 0)intocond << a. It only runs on RISCV:runtime/src/coreclr/jit/ifconversion.cpp
Lines 721 to 728 in b70c35e
However our current cost-analysis always rejects such patterns. Example:
If we look at JitDump:
So the optimization above for RISCV is essentially dead code. We always bail out earlier which is probably not wanted.