| Issue |
204486
|
| Summary |
[RISCV] Performance regression from change in getRegisterCostTableIndex
|
| Labels |
new issue
|
| Assignees |
|
| Reporter |
KavinTheG
|
Noticed this regression on the BPI-F3 (rva22u64_v):
http://lnt.llvm.org/db_default/v4/nts/profile/1438/121/117
<img width="2517" height="843" alt="Image" src="" />
It looks like that the code generated is the identical, but the `s0` and `s2` registers are swapped. I have traced this change to this commit https://github.com/llvm/llvm-project/pull/201501.
Here is the diff I have been able to produce locally.
```diff
diff --git a/pre/fib2.s b/fib2.s
index a727287..57c5b31 100644
--- a/pre/fib2.s
+++ b/fib2.s
@@ -19,8 +19,8 @@ fib: # @fib
.cfi_offset s1, -24
.cfi_offset s2, -32
.cfi_remember_state
- li s2, 2
- bgeu a0, s2, .LBB0_2
+ li s0, 2
+ bgeu a0, s0, .LBB0_2
# %bb.1:
li a0, 1
ld ra, 24(sp) # 8-byte Folded Reload
@@ -37,14 +37,14 @@ fib: # @fib
.LBB0_2: # %if.else.preheader
.cfi_restore_state
li s1, 0
- addi s0, a0, 1
+ addi s2, a0, 1
.LBB0_3: # %if.else
# =>This Inner Loop Header: Depth=1
- addi a0, s0, -3
+ addi a0, s2, -3
call fib
- addi s0, s0, -1
+ addi s2, s2, -1
add s1, s1, a0
- bltu s2, s0, .LBB0_3
+ bltu s0, s2, .LBB0_3
# %bb.4: # %return.loopexit
addi a0, s1, 1
ld ra, 24(sp) # 8-byte Folded Reload
@@ -103,6 +103,6 @@ main: # @main
.asciz "%ld\n"
.size .L.str, 5
- .ident "clang version 23.0.0git ([email protected]:KavinTheG/llvm-project.git 5e7ec28c5e3a70588bbe9368d3816b009cf3670f)"
+ .ident "clang version 23.0.0git ([email protected]:KavinTheG/llvm-project.git 86c073af08bac41fe0349eee2bc0279e5f88d6d9)"
.section ".note.GNU-stack","",@progbits
.addrsig
```
This is the command I use to compile the code, which I am fairly confident is the same used in these benchmarks (except without -flto, which I don't think is relevant here).
```zsh
$LBD/bin/clang --target=riscv64-unknown-linux-gnu \
--sysroot="$RISCV_SYSROOT" \
-march=rva22u64_v \
-O3 \
-S fib2.c -o fib2.s
```
I've only taken a cursory glance at the benchmarks below but I believe the regressions stem from the same commit.
http://lnt.llvm.org/db_default/v4/nts/profile/1422/141/2
http://lnt.llvm.org/db_default/v4/nts/profile/1354/141/2
![]()
_______________________________________________
llvm-bugs mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
