llvmbot
commented
2 months ago @llvm/issue-subscribers-backend-aarch64
Author: Mamy Ratsimbazafy (mratsim)
Same IR as https://github.com/llvm/llvm-project/issues/103841 but applied to Aarch64 as an alternative to https://github.com/llvm/llvm-project/issues/103717
Unlike x86 there is always an extra instruction even for i256, and there are 2 unnecessary instruction for i320 or i384.
https://alive2.llvm.org/ce/z/-bGiUs
## Full code
### Original IR
```llvm
; ModuleID = 'x86_poc'
target triple = "arm64"
; target triple = "x86_64"
@bn254_snarks_fp_mod = constant i256 21888242871839275222246405745257275088696311157297823662689037894645226208583, section "ctt.bn254_snarks_fp.constants", align 64
@bls12_381_fp_mod = constant i384 4002409555221667393417789825735904156556882819939007885332058136124031650490837864442687629129015664037894272559787, section "ctt.bls12_381_fp.constants", align 64
@bls24_317_fp_mod = constant i320 136393071104295911515099765908274057061945112121419593977210139303905973197232025618026156731051, section "ctt.bls24_317_fp.constants", align 64
; Function Attrs: hot
define internal fastcc void @_modadd_noo.u64x4(ptr %0, ptr %1, ptr %2, ptr %3) #2 section "ctt.fields" {
%a = load i256, ptr %1, align 4
%b = load i256, ptr %2, align 4
%M = load i256, ptr %3, align 4
%a_plus_b = add i256 %a, %b
%5 = sub i256 %a_plus_b, %M
%6 = lshr i256 %5, 255
%7 = trunc i256 %6 to i1
%8 = select i1 %7, i256 %a_plus_b, i256 %5
store i256 %8, ptr %0, align 4
ret void
}
; Function Attrs: hot
define internal fastcc void @_modadd_noo.u64x5(ptr %0, ptr %1, ptr %2, ptr %3) #2 section "ctt.fields" {
%a = load i320, ptr %1, align 4
%b = load i320, ptr %2, align 4
%M = load i320, ptr %3, align 4
%a_plus_b = add i320 %a, %b
%5 = sub i320 %a_plus_b, %M
%6 = lshr i320 %5, 319
%7 = trunc i320 %6 to i1
%8 = select i1 %7, i320 %a_plus_b, i320 %5
store i320 %8, ptr %0, align 4
ret void
}
; Function Attrs: hot
define internal fastcc void @_modadd_noo.u64x6(ptr %0, ptr %1, ptr %2, ptr %3) #2 section "ctt.fields" {
%a = load i384, ptr %1, align 4
%b = load i384, ptr %2, align 4
%M = load i384, ptr %3, align 4
%a_plus_b = add i384 %a, %b
%5 = sub i384 %a_plus_b, %M
%6 = lshr i384 %5, 383
%7 = trunc i384 %6 to i1
%8 = select i1 %7, i384 %a_plus_b, i384 %5
store i384 %8, ptr %0, align 4
ret void
}
; Function Attrs: hot
define void @bn254_snarks_fp_add(ptr %0, ptr %1, ptr %2) #2 section "ctt.bn254_snarks_fp" {
call fastcc void @_modadd_noo.u64x4(ptr %0, ptr %1, ptr %2, ptr @bn254_snarks_fp_mod)
ret void
}
; Function Attrs: hot
define void @bls24_317_fp_add(ptr %0, ptr %1, ptr %2) #2 section "ctt.bls24_317_fp" {
call fastcc void @_modadd_noo.u64x5(ptr %0, ptr %1, ptr %2, ptr @bls24_317_fp_mod)
ret void
}
; Function Attrs: hot
define void @bls12_381_fp_add(ptr %0, ptr %1, ptr %2) #2 section "ctt.bls12_381_fp" {
call fastcc void @_modadd_noo.u64x6(ptr %0, ptr %1, ptr %2, ptr @bls12_381_fp_mod)
ret void
}
attributes #2 = { hot }
```
### After opt -O3
```llvm
target triple = "arm64"
; target triple = "x86_64"
define void @bn254_snarks_fp_add(ptr nocapture writeonly %0, ptr nocapture readonly %1, ptr nocapture readonly %2) local_unnamed_addr #0 section "ctt.bn254_snarks_fp" {
%.val = load i256, ptr %1, align 4
%.val1 = load i256, ptr %2, align 4
%a_plus_b.i = add i256 %.val1, %.val
%4 = add i256 %a_plus_b.i, -21888242871839275222246405745257275088696311157297823662689037894645226208583
%.not1.i = icmp slt i256 %4, 0
%5 = select i1 %.not1.i, i256 %a_plus_b.i, i256 %4
store i256 %5, ptr %0, align 4
ret void
}
define void @bls24_317_fp_add(ptr nocapture writeonly %0, ptr nocapture readonly %1, ptr nocapture readonly %2) local_unnamed_addr #0 section "ctt.bls24_317_fp" {
%.val = load i320, ptr %1, align 4
%.val1 = load i320, ptr %2, align 4
%a_plus_b.i = add i320 %.val1, %.val
%4 = add i320 %a_plus_b.i, -136393071104295911515099765908274057061945112121419593977210139303905973197232025618026156731051
%.not1.i = icmp slt i320 %4, 0
%5 = select i1 %.not1.i, i320 %a_plus_b.i, i320 %4
store i320 %5, ptr %0, align 4
ret void
}
define void @bls12_381_fp_add(ptr nocapture writeonly %0, ptr nocapture readonly %1, ptr nocapture readonly %2) local_unnamed_addr #0 section "ctt.bls12_381_fp" {
%.val = load i384, ptr %1, align 4
%.val1 = load i384, ptr %2, align 4
%a_plus_b.i = add i384 %.val1, %.val
%4 = add i384 %a_plus_b.i, -4002409555221667393417789825735904156556882819939007885332058136124031650490837864442687629129015664037894272559787
%.not1.i = icmp slt i384 %4, 0
%5 = select i1 %.not1.i, i384 %a_plus_b.i, i384 %4
store i384 %5, ptr %0, align 4
ret void
}
attributes #0 = { hot mustprogress nofree norecurse nosync nounwind willreturn memory(argmem: readwrite) }
```
### Assembly
```asm
bn254_snarks_fp_add: // @bn254_snarks_fp_add
ldp x8, x10, [x1]
mov x15, #24534 // =0x5fd6
ldp x9, x11, [x2]
movk x15, #7886, lsl #16
ldp x13, x14, [x2, #16]
movk x15, #45453, lsl #32
movk x15, #53147, lsl #48
adds x8, x9, x8
ldp x12, x9, [x1, #16]
adcs x10, x11, x10
mov x11, #697 // =0x2b9
movk x11, #10115, lsl #16
adcs x12, x13, x12
mov x13, #13682 // =0x3572
movk x11, #29673, lsl #32
movk x13, #38798, lsl #16
adc x9, x14, x9
mov x14, #42914 // =0xa7a2
movk x11, #50143, lsl #48
movk x13, #38254, lsl #32
movk x14, #32382, lsl #16
movk x13, #26750, lsl #48
adds x11, x8, x11
movk x14, #47689, lsl #32
movk x14, #18351, lsl #48
adcs x13, x10, x13
adcs x14, x12, x14
adc x15, x9, x15
cmp x15, #0
csel x12, x12, x14, lt
csel x9, x9, x15, lt
csel x8, x8, x11, lt
stp x12, x9, [x0, #16]
csel x9, x10, x13, lt
stp x8, x9, [x0]
ret
bls24_317_fp_add: // @bls24_317_fp_add
ldp x8, x10, [x1]
mov x16, #12230 // =0x2fc6
ldp x9, x11, [x2]
movk x16, #18438, lsl #16
ldp x13, x14, [x2, #16]
mov x17, #30419 // =0x76d3
ldr x15, [x2, #32]
movk x16, #14943, lsl #32
movk x17, #37023, lsl #16
adds x8, x9, x8
ldp x12, x9, [x1, #16]
adcs x10, x11, x10
ldr x11, [x1, #32]
movk x16, #3440, lsl #48
movk x17, #13789, lsl #32
adcs x12, x13, x12
mov x13, #54613 // =0xd555
movk x17, #61351, lsl #48
movk x13, #41556, lsl #16
adcs x9, x14, x9
mov x14, #16513 // =0x4081
movk x13, #53673, lsl #32
movk x14, #52187, lsl #16
adc x11, x15, x11
mov x15, #52153 // =0xcbb9
movk x13, #29358, lsl #48
movk x14, #50715, lsl #32
movk x15, #31544, lsl #16
movk x14, #10508, lsl #48
adds x13, x8, x13
movk x15, #40473, lsl #32
adcs x14, x10, x14
movk x15, #59749, lsl #48
adcs x15, x12, x15
adcs x16, x9, x16
adc x17, x11, x17
asr x18, x17, #63
cmp x18, #0
csel x11, x11, x17, lt
csel x9, x9, x16, lt
csel x12, x12, x15, lt
csel x8, x8, x13, lt
stp x9, x11, [x0, #24]
csel x9, x10, x14, lt
str x8, [x0]
stp x9, x12, [x0, #8]
ret
bls12_381_fp_add: // @bls12_381_fp_add
ldp x8, x10, [x1]
mov x17, #60736 // =0xed40
ldp x9, x11, [x2]
movk x17, #3194, lsl #16
ldp x13, x14, [x2, #16]
mov x18, #21288 // =0x5328
ldp x15, x16, [x2, #32]
movk x17, #46203, lsl #32
adds x8, x9, x8
ldp x12, x9, [x1, #16]
adcs x10, x11, x10
movk x18, #48308, lsl #16
movk x17, #39816, lsl #48
movk x18, #22601, lsl #32
adcs x12, x13, x12
ldp x11, x13, [x1, #32]
adcs x9, x14, x9
mov x14, #21845 // =0x5555
mov x1, #6501 // =0x1965
movk x14, #17921, lsl #48
movk x1, #50816, lsl #16
movk x18, #46308, lsl #48
adcs x11, x15, x11
mov x15, #1319895040 // =0x4eac0000
movk x1, #60949, lsl #32
adc x13, x16, x13
mov x16, #2523 // =0x9db
movk x15, #1, lsl #32
movk x16, #2383, lsl #16
movk x15, #57684, lsl #48
adds x14, x8, x14
movk x16, #11615, lsl #32
adcs x15, x10, x15
movk x1, #58878, lsl #48
movk x16, #39119, lsl #48
adcs x16, x12, x16
adcs x17, x9, x17
adcs x18, x11, x18
adc x1, x13, x1
asr x2, x1, #63
cmp x2, #0
csel x11, x11, x18, lt
csel x13, x13, x1, lt
csel x9, x9, x17, lt
stp x11, x13, [x0, #32]
csel x11, x12, x16, lt
csel x8, x8, x14, lt
stp x11, x9, [x0, #16]
csel x9, x10, x15, lt
stp x8, x9, [x0]
ret
```
## Analysis
With i256, the `cmp` is useless in this sequence
```asm
movk x11, #50143, lsl #48
movk x13, #38254, lsl #32
movk x14, #32382, lsl #16
movk x13, #26750, lsl #48
adds x11, x8, x11
movk x14, #47689, lsl #32
movk x14, #18351, lsl #48
adcs x13, x10, x13
adcs x14, x12, x14
adc x15, x9, x15
cmp x15, #0 // <----- unnecessary
csel x12, x12, x14, lt
csel x9, x9, x15, lt
csel x8, x8, x11, lt
stp x12, x9, [x0, #16]
csel x9, x10, x13, lt
stp x8, x9, [x0]
ret
```
as demonstrated by https://github.com/llvm/llvm-project/issues/103717
With i320, similar to x86 https://github.com/llvm/llvm-project/issues/103841, there is another additional `asr` instruction
```asm
movk x13, #29358, lsl #48
movk x14, #50715, lsl #32
movk x15, #31544, lsl #16
movk x14, #10508, lsl #48
adds x13, x8, x13
movk x15, #40473, lsl #32
adcs x14, x10, x14
movk x15, #59749, lsl #48
adcs x15, x12, x15
adcs x16, x9, x16
adc x17, x11, x17
asr x18, x17, #63 // <----- unnecessary
cmp x18, #0 // <----- unnecessary
csel x11, x11, x17, lt
csel x9, x9, x16, lt
csel x12, x12, x15, lt
csel x8, x8, x13, lt
stp x9, x11, [x0, #24]
csel x9, x10, x14, lt
str x8, [x0]
stp x9, x12, [x0, #8]
ret
```
Same IR as https://github.com/llvm/llvm-project/issues/103841 but applied to Aarch64 as an alternative to https://github.com/llvm/llvm-project/issues/103717
Unlike x86 there is always an extra instruction even for i256, and there are 2 unnecessary instruction for i320 or i384.
https://alive2.llvm.org/ce/z/-bGiUs
Full code
Original IR
After opt -O3
Assembly
Analysis
With i256, the
cmpis useless in this sequenceas demonstrated by https://github.com/llvm/llvm-project/issues/103717
With i320, similar to x86 https://github.com/llvm/llvm-project/issues/103841, there is another additional
asrinstruction