Skip to content

Commit

Permalink
[AArch64][SVE] Workaround incorrect types when lowering fixed length …
Browse files Browse the repository at this point in the history 
…gather/scatter

When lowering a fixed length gather/scatter the index type is assumed to
be the same as the memory type, this is incorrect in cases where the
extension of the index has been folded into the addressing mode.

For now add a temporary workaround to fix the codegen faults caused by
this by preventing the removal of this extension. At a later date the
lowering for SVE gather/scatters will be redesigned to improve the way
addressing modes are handled.

As a short term side effect of this change, the addressing modes
generated for fixed length gather/scatters will not be optimal.

Differential Revision: https://reviews.llvm.org/D109145
  • Loading branch information
@brads55
brads55 committed Sep 2, 2021
1 parent 3fd27ec commit 14e1a4a
Show file tree
Hide file tree
Showing 3 changed files with 115 additions and 77 deletions.
3 changes: 2 additions & 1 deletion llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
View file
Expand Up @@ -4222,7 +4222,8 @@ bool AArch64TargetLowering::shouldExtendGSIndex(EVT VT, EVT &EltTy) const {

bool AArch64TargetLowering::shouldRemoveExtendFromGSIndex(EVT VT) const {
if (VT.getVectorElementType() == MVT::i32 &&
VT.getVectorElementCount().getKnownMinValue() >= 4)
VT.getVectorElementCount().getKnownMinValue() >= 4 &&
!VT.isFixedLengthVector())
return true;

return false;
Expand Down
100 changes: 59 additions & 41 deletions llvm/test/CodeGen/AArch64/sve-fixed-length-masked-gather.ll
View file
Expand Up @@ -917,19 +917,22 @@ define void @masked_gather_v32f64(<32 x double>* %a, <32 x double*>* %b) #0 {
; The above tests test the types, the below tests check that the addressing
; modes still function

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_gather_32b_scaled_sext_f16(<32 x half>* %a, <32 x i32>* %b, half* %base) #0 {
; CHECK-LABEL: masked_gather_32b_scaled_sext_f16:
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].h, vl32
; VBITS_GE_2048-NEXT: ld1h { [[VALS:z[0-9]+]].h }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1sw { [[PTRS:z[0-9]+]].d }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[CMP:p[0-9]+]].h, [[PG0]]/z, [[VALS]].h, #0.0
; VBITS_GE_2048-NEXT: mov [[MONE:z[0-9]+]].h, [[CMP]]/z, #-1
; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].s, [[PG1]]/z, [[UPK]].s, #0
; VBITS_GE_2048-NEXT: ld1h { [[RES:z[0-9]+]].s }, [[MASK]]/z, [x2, [[PTRS]].s, sxtw #1]
; VBITS_GE_2048-NEXT: uzp1 [[UZP:z[0-9]+]].h, [[RES]].h, [[RES]].h
; VBITS_GE_2048-NEXT: st1h { [[UZP]].h }, [[PG0]], [x0]
; VBITS_GE_2048-NEXT: uunpklo [[UPK1:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].d, [[PG1]]/z, [[UPK2]].d, #0
; VBITS_GE_2048-NEXT: ld1h { [[RES:z[0-9]+]].d }, [[MASK]]/z, [x2, [[PTRS]].d, lsl #1]
; VBITS_GE_2048-NEXT: uzp1 [[UZP1:z[0-9]+]].s, [[RES]].s, [[RES]].s
; VBITS_GE_2048-NEXT: uzp1 [[UZP2:z[0-9]+]].h, [[UZP1]].h, [[UZP1]].h
; VBITS_GE_2048-NEXT: st1h { [[UZP2]].h }, [[PG0]], [x0]
; VBITS_GE_2048-NEXT: ret
%cvals = load <32 x half>, <32 x half>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand All @@ -941,14 +944,20 @@ define void @masked_gather_32b_scaled_sext_f16(<32 x half>* %a, <32 x i32>* %b,
ret void
}

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_gather_32b_scaled_sext_f32(<32 x float>* %a, <32 x i32>* %b, float* %base) #0 {
; CHECK-LABEL: masked_gather_32b_scaled_sext_f32:
; VBITS_GE_2048: ptrue [[PG:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[VALS:z[0-9]+]].s }, [[PG]]/z, [x0]
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[MASK:p[0-9]+]].s, [[PG]]/z, [[VALS]].s, #0.0
; VBITS_GE_2048-NEXT: ld1w { [[RES:z[0-9]+]].s }, [[MASK]]/z, [x2, [[PTRS]].s, sxtw #2]
; VBITS_GE_2048-NEXT: st1w { [[RES]].s }, [[PG]], [x0]
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[VALS:z[0-9]+]].s }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1sw { [[PTRS:z[0-9]+]].d }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[CMP:p[0-9]+]].s, [[PG0]]/z, [[VALS]].s, #0.0
; VBITS_GE_2048-NEXT: mov [[MONE:z[0-9]+]].s, [[CMP]]/z, #-1
; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].d, [[MONE]].s
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].d, [[PG1]]/z, [[UPK]].d, #0
; VBITS_GE_2048-NEXT: ld1w { [[RES:z[0-9]+]].d }, [[MASK]]/z, [x2, [[PTRS]].d, lsl #2]
; VBITS_GE_2048-NEXT: uzp1 [[UZP:z[0-9]+]].s, [[RES]].s, [[RES]].s
; VBITS_GE_2048-NEXT: st1w { [[UZP]].s }, [[PG0]], [x0]
; VBITS_GE_2048-NEXT: ret
%cvals = load <32 x float>, <32 x float>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand All @@ -960,15 +969,15 @@ define void @masked_gather_32b_scaled_sext_f32(<32 x float>* %a, <32 x i32>* %b,
ret void
}

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_gather_32b_scaled_sext_f64(<32 x double>* %a, <32 x i32>* %b, double* %base) #0 {
; CHECK-LABEL: masked_gather_32b_scaled_sext_f64:
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1d { [[VALS:z[0-9]+]].d }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[MASK:p[0-9]+]].d, [[PG0]]/z, [[VALS]].d, #0.0
; VBITS_GE_2048-NEXT: ld1d { [[RES:z[0-9]+]].d }, [[MASK]]/z, [x2, [[PTRS]].d, sxtw #3]
; VBITS_GE_2048-NEXT: st1d { [[RES]].d }, [[PG0]], [x0]
; VBITS_GE_2048: ptrue [[PG:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1d { [[VALS:z[0-9]+]].d }, [[PG]]/z, [x0]
; VBITS_GE_2048-NEXT: ld1sw { [[PTRS:z[0-9]+]].d }, [[PG]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[MASK:p[0-9]+]].d, [[PG]]/z, [[VALS]].d, #0.0
; VBITS_GE_2048-NEXT: ld1d { [[RES:z[0-9]+]].d }, [[MASK]]/z, [x2, [[PTRS]].d, lsl #3]
; VBITS_GE_2048-NEXT: st1d { [[RES]].d }, [[PG]], [x0]
; VBITS_GE_2048-NEXT: ret
%cvals = load <32 x double>, <32 x double>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand All @@ -980,19 +989,22 @@ define void @masked_gather_32b_scaled_sext_f64(<32 x double>* %a, <32 x i32>* %b
ret void
}

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_gather_32b_scaled_zext(<32 x half>* %a, <32 x i32>* %b, half* %base) #0 {
; CHECK-LABEL: masked_gather_32b_scaled_zext:
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].h, vl32
; VBITS_GE_2048-NEXT: ld1h { [[VALS:z[0-9]+]].h }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].d }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[CMP:p[0-9]+]].h, [[PG0]]/z, [[VALS]].h, #0.0
; VBITS_GE_2048-NEXT: mov [[MONE:z[0-9]+]].h, [[CMP]]/z, #-1
; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].s, [[PG1]]/z, [[UPK]].s, #0
; VBITS_GE_2048-NEXT: ld1h { [[RES:z[0-9]+]].s }, [[MASK]]/z, [x2, [[PTRS]].s, uxtw #1]
; VBITS_GE_2048-NEXT: uzp1 [[UZP:z[0-9]+]].h, [[RES]].h, [[RES]].h
; VBITS_GE_2048-NEXT: st1h { [[UZP]].h }, [[PG0]], [x0]
; VBITS_GE_2048-NEXT: uunpklo [[UPK1:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].d, [[PG1]]/z, [[UPK2]].d, #0
; VBITS_GE_2048-NEXT: ld1h { [[RES:z[0-9]+]].d }, [[MASK]]/z, [x2, [[PTRS]].d, lsl #1]
; VBITS_GE_2048-NEXT: uzp1 [[UZP1:z[0-9]+]].s, [[RES]].s, [[RES]].s
; VBITS_GE_2048-NEXT: uzp1 [[UZP2:z[0-9]+]].h, [[UZP1]].h, [[UZP1]].h
; VBITS_GE_2048-NEXT: st1h { [[UZP2]].h }, [[PG0]], [x0]
; VBITS_GE_2048-NEXT: ret
%cvals = load <32 x half>, <32 x half>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand All @@ -1004,19 +1016,22 @@ define void @masked_gather_32b_scaled_zext(<32 x half>* %a, <32 x i32>* %b, half
ret void
}

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_gather_32b_unscaled_sext(<32 x half>* %a, <32 x i32>* %b, i8* %base) #0 {
; CHECK-LABEL: masked_gather_32b_unscaled_sext:
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].h, vl32
; VBITS_GE_2048-NEXT: ld1h { [[VALS:z[0-9]+]].h }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1sw { [[PTRS:z[0-9]+]].d }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[CMP:p[0-9]+]].h, [[PG0]]/z, [[VALS]].h, #0.0
; VBITS_GE_2048-NEXT: mov [[MONE:z[0-9]+]].h, [[CMP]]/z, #-1
; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].s, [[PG1]]/z, [[UPK]].s, #0
; VBITS_GE_2048-NEXT: ld1h { [[RES:z[0-9]+]].s }, [[MASK]]/z, [x2, [[PTRS]].s, sxtw]
; VBITS_GE_2048-NEXT: uzp1 [[UZP:z[0-9]+]].h, [[RES]].h, [[RES]].h
; VBITS_GE_2048-NEXT: st1h { [[UZP]].h }, [[PG0]], [x0]
; VBITS_GE_2048-NEXT: uunpklo [[UPK1:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].d, [[PG1]]/z, [[UPK2]].d, #0
; VBITS_GE_2048-NEXT: ld1h { [[RES:z[0-9]+]].d }, [[MASK]]/z, [x2, [[PTRS]].d]
; VBITS_GE_2048-NEXT: uzp1 [[UZP1:z[0-9]+]].s, [[RES]].s, [[RES]].s
; VBITS_GE_2048-NEXT: uzp1 [[UZP2:z[0-9]+]].h, [[UZP1]].h, [[UZP1]].h
; VBITS_GE_2048-NEXT: st1h { [[UZP2]].h }, [[PG0]], [x0]
; VBITS_GE_2048-NEXT: ret
%cvals = load <32 x half>, <32 x half>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand All @@ -1029,19 +1044,22 @@ define void @masked_gather_32b_unscaled_sext(<32 x half>* %a, <32 x i32>* %b, i8
ret void
}

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_gather_32b_unscaled_zext(<32 x half>* %a, <32 x i32>* %b, i8* %base) #0 {
; CHECK-LABEL: masked_gather_32b_unscaled_zext:
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].h, vl32
; VBITS_GE_2048-NEXT: ld1h { [[VALS:z[0-9]+]].h }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].d }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[CMP:p[0-9]+]].h, [[PG0]]/z, [[VALS]].h, #0.0
; VBITS_GE_2048-NEXT: mov [[MONE:z[0-9]+]].h, [[CMP]]/z, #-1
; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].s, [[PG1]]/z, [[UPK]].s, #0
; VBITS_GE_2048-NEXT: ld1h { [[RES:z[0-9]+]].s }, [[MASK]]/z, [x2, [[PTRS]].s, uxtw]
; VBITS_GE_2048-NEXT: uzp1 [[UZP:z[0-9]+]].h, [[RES]].h, [[RES]].h
; VBITS_GE_2048-NEXT: st1h { [[UZP]].h }, [[PG0]], [x0]
; VBITS_GE_2048-NEXT: uunpklo [[UPK1:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].d, [[PG1]]/z, [[UPK2]].d, #0
; VBITS_GE_2048-NEXT: ld1h { [[RES:z[0-9]+]].d }, [[MASK]]/z, [x2, [[PTRS]].d]
; VBITS_GE_2048-NEXT: uzp1 [[UZP1:z[0-9]+]].s, [[RES]].s, [[RES]].s
; VBITS_GE_2048-NEXT: uzp1 [[UZP2:z[0-9]+]].h, [[UZP1]].h, [[UZP1]].h
; VBITS_GE_2048-NEXT: st1h { [[UZP2]].h }, [[PG0]], [x0]
; VBITS_GE_2048-NEXT: ret
%cvals = load <32 x half>, <32 x half>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand Down
89 changes: 54 additions & 35 deletions llvm/test/CodeGen/AArch64/sve-fixed-length-masked-scatter.ll
View file
Expand Up @@ -839,18 +839,22 @@ define void @masked_scatter_v32f64(<32 x double>* %a, <32 x double*>* %b) #0 {

; The above tests test the types, the below tests check that the addressing
; modes still function

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_scatter_32b_scaled_sext_f16(<32 x half>* %a, <32 x i32>* %b, half* %base) #0 {
; CHECK-LABEL: masked_scatter_32b_scaled_sext_f16:
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].h, vl32
; VBITS_GE_2048-NEXT: ld1h { [[VALS:z[0-9]+]].h }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1sw { [[PTRS:z[0-9]+]].d }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[CMP:p[0-9]+]].h, [[PG0]]/z, [[VALS]].h, #0.0
; VBITS_GE_2048-NEXT: mov [[MONE:z[0-9]+]].h, [[PG0]]/z, #-1
; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].s, [[PG1]]/z, [[UPK]].s, #0
; VBITS_GE_2048-NEXT: uunpklo [[UPKV:z[0-9]+]].s, [[VALS]].h
; VBITS_GE_2048-NEXT: st1h { [[UPKV]].s }, [[MASK]], [x2, [[PTRS]].s, sxtw #1]
; VBITS_GE_2048-NEXT: uunpklo [[UPK1:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPKV1:z[0-9]+]].s, [[VALS]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].d, [[PG1]]/z, [[UPK2]].d, #0
; VBITS_GE_2048-NEXT: uunpklo [[UPKV2:z[0-9]+]].d, [[UPKV1]].s
; VBITS_GE_2048-NEXT: st1h { [[UPKV2]].d }, [[MASK]], [x2, [[PTRS]].d, lsl #1]
; VBITS_GE_2048-NEXT: ret
%vals = load <32 x half>, <32 x half>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand All @@ -861,13 +865,19 @@ define void @masked_scatter_32b_scaled_sext_f16(<32 x half>* %a, <32 x i32>* %b,
ret void
}

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_scatter_32b_scaled_sext_f32(<32 x float>* %a, <32 x i32>* %b, float* %base) #0 {
; CHECK-LABEL: masked_scatter_32b_scaled_sext_f32:
; VBITS_GE_2048: ptrue [[PG:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[VALS:z[0-9]+]].s }, [[PG]]/z, [x0]
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[MASK:p[0-9]+]].s, [[PG]]/z, [[VALS]].s, #0.0
; VBITS_GE_2048-NEXT: st1w { [[VALS]].s }, [[MASK]], [x2, [[PTRS]].s, sxtw #2]
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[VALS:z[0-9]+]].s }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1sw { [[PTRS:z[0-9]+]].d }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[CMP:p[0-9]+]].s, [[PG0]]/z, [[VALS]].s, #0.0
; VBITS_GE_2048-NEXT: mov [[MONE:z[0-9]+]].s, [[PG0]]/z, #-1
; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].d, [[MONE]].s
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].d, [[PG1]]/z, [[UPK]].d, #0
; VBITS_GE_2048-NEXT: uunpklo [[UPKV:z[0-9]+]].d, [[VALS]].s
; VBITS_GE_2048-NEXT: st1w { [[UPKV]].d }, [[MASK]], [x2, [[PTRS]].d, lsl #2]
; VBITS_GE_2048-NEXT: ret
%vals = load <32 x float>, <32 x float>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand All @@ -878,14 +888,14 @@ define void @masked_scatter_32b_scaled_sext_f32(<32 x float>* %a, <32 x i32>* %b
ret void
}

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_scatter_32b_scaled_sext_f64(<32 x double>* %a, <32 x i32>* %b, double* %base) #0 {
; CHECK-LABEL: masked_scatter_32b_scaled_sext_f64:
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1d { [[VALS:z[0-9]+]].d }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[MASK:p[0-9]+]].d, [[PG0]]/z, [[VALS]].d, #0.0
; VBITS_GE_2048-NEXT: st1d { [[VALS]].d }, [[MASK]], [x2, [[PTRS]].d, sxtw #3]
; VBITS_GE_2048: ptrue [[PG:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1d { [[VALS:z[0-9]+]].d }, [[PG]]/z, [x0]
; VBITS_GE_2048-NEXT: ld1sw { [[PTRS:z[0-9]+]].d }, [[PG]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[MASK:p[0-9]+]].d, [[PG]]/z, [[VALS]].d, #0.0
; VBITS_GE_2048-NEXT: st1d { [[VALS]].d }, [[MASK]], [x2, [[PTRS]].d, lsl #3]
; VBITS_GE_2048-NEXT: ret
%vals = load <32 x double>, <32 x double>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand All @@ -896,18 +906,21 @@ define void @masked_scatter_32b_scaled_sext_f64(<32 x double>* %a, <32 x i32>* %
ret void
}

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_scatter_32b_scaled_zext(<32 x half>* %a, <32 x i32>* %b, half* %base) #0 {
; CHECK-LABEL: masked_scatter_32b_scaled_zext:
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].h, vl32
; VBITS_GE_2048-NEXT: ld1h { [[VALS:z[0-9]+]].h }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].d }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[CMP:p[0-9]+]].h, [[PG0]]/z, [[VALS]].h, #0.0
; VBITS_GE_2048-NEXT: mov [[MONE:z[0-9]+]].h, [[PG0]]/z, #-1
; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].s, [[PG1]]/z, [[UPK]].s, #0
; VBITS_GE_2048-NEXT: uunpklo [[UPKV:z[0-9]+]].s, [[VALS]].h
; VBITS_GE_2048-NEXT: st1h { [[UPKV]].s }, [[MASK]], [x2, [[PTRS]].s, uxtw #1]
; VBITS_GE_2048-NEXT: uunpklo [[UPK1:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPKV1:z[0-9]+]].s, [[VALS]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].d, [[PG1]]/z, [[UPK2]].d, #0
; VBITS_GE_2048-NEXT: uunpklo [[UPKV2:z[0-9]+]].d, [[UPKV1]].s
; VBITS_GE_2048-NEXT: st1h { [[UPKV2]].d }, [[MASK]], [x2, [[PTRS]].d, lsl #1]
; VBITS_GE_2048-NEXT: ret
%vals = load <32 x half>, <32 x half>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand All @@ -918,18 +931,21 @@ define void @masked_scatter_32b_scaled_zext(<32 x half>* %a, <32 x i32>* %b, hal
ret void
}

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_scatter_32b_unscaled_sext(<32 x half>* %a, <32 x i32>* %b, i8* %base) #0 {
; CHECK-LABEL: masked_scatter_32b_unscaled_sext:
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].h, vl32
; VBITS_GE_2048-NEXT: ld1h { [[VALS:z[0-9]+]].h }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1sw { [[PTRS:z[0-9]+]].d }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[CMP:p[0-9]+]].h, [[PG0]]/z, [[VALS]].h, #0.0
; VBITS_GE_2048-NEXT: mov [[MONE:z[0-9]+]].h, [[PG0]]/z, #-1
; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].s, [[PG1]]/z, [[UPK]].s, #0
; VBITS_GE_2048-NEXT: uunpklo [[UPKV:z[0-9]+]].s, [[VALS]].h
; VBITS_GE_2048-NEXT: st1h { [[UPKV]].s }, [[MASK]], [x2, [[PTRS]].s, sxtw]
; VBITS_GE_2048-NEXT: uunpklo [[UPK1:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPKV1:z[0-9]+]].s, [[VALS]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].d, [[PG1]]/z, [[UPK2]].d, #0
; VBITS_GE_2048-NEXT: uunpklo [[UPKV2:z[0-9]+]].d, [[UPKV1]].s
; VBITS_GE_2048-NEXT: st1h { [[UPKV2]].d }, [[MASK]], [x2, [[PTRS]].d]
; VBITS_GE_2048-NEXT: ret
%vals = load <32 x half>, <32 x half>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand All @@ -941,18 +957,21 @@ define void @masked_scatter_32b_unscaled_sext(<32 x half>* %a, <32 x i32>* %b, i
ret void
}

; NOTE: This produces an non-optimal addressing mode due to a temporary workaround
define void @masked_scatter_32b_unscaled_zext(<32 x half>* %a, <32 x i32>* %b, i8* %base) #0 {
; CHECK-LABEL: masked_scatter_32b_unscaled_zext:
; VBITS_GE_2048: ptrue [[PG0:p[0-9]+]].h, vl32
; VBITS_GE_2048-NEXT: ld1h { [[VALS:z[0-9]+]].h }, [[PG0]]/z, [x0]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].s, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].s }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: ptrue [[PG1:p[0-9]+]].d, vl32
; VBITS_GE_2048-NEXT: ld1w { [[PTRS:z[0-9]+]].d }, [[PG1]]/z, [x1]
; VBITS_GE_2048-NEXT: fcmeq [[CMP:p[0-9]+]].h, [[PG0]]/z, [[VALS]].h, #0.0
; VBITS_GE_2048-NEXT: mov [[MONE:z[0-9]+]].h, [[PG0]]/z, #-1
; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].s, [[PG1]]/z, [[UPK]].s, #0
; VBITS_GE_2048-NEXT: uunpklo [[UPKV:z[0-9]+]].s, [[VALS]].h
; VBITS_GE_2048-NEXT: st1h { [[UPKV]].s }, [[MASK]], [x2, [[PTRS]].s, uxtw]
; VBITS_GE_2048-NEXT: uunpklo [[UPK1:z[0-9]+]].s, [[MONE]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPKV1:z[0-9]+]].s, [[VALS]].h
; VBITS_GE_2048-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
; VBITS_GE_2048-NEXT: cmpne [[MASK:p[0-9]+]].d, [[PG1]]/z, [[UPK2]].d, #0
; VBITS_GE_2048-NEXT: uunpklo [[UPKV2:z[0-9]+]].d, [[UPKV1]].s
; VBITS_GE_2048-NEXT: st1h { [[UPKV2]].d }, [[MASK]], [x2, [[PTRS]].d]
; VBITS_GE_2048-NEXT: ret
%vals = load <32 x half>, <32 x half>* %a
%idxs = load <32 x i32>, <32 x i32>* %b
Expand Down

0 comments on commit 14e1a4a

Please sign in to comment.