Skip to content

[LV] Add support for uniform parameters on vectorized function variants #72891

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Nov 28, 2023
Merged

[LV] Add support for uniform parameters on vectorized function variants #72891

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 8 additions & 0 deletions llvm/lib/Transforms/Vectorize/LoopVectorize.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -7027,6 +7027,14 @@ void LoopVectorizationCostModel::setVectorizedCallDecision(ElementCount VF) {
switch (Param.ParamKind) {
case VFParamKind::Vector:
break;
case VFParamKind::OMP_Uniform: {
Value *ScalarParam = CI->getArgOperand(Param.ParamPos);
// Make sure the scalar parameter in the loop is invariant.
if (!PSE.getSE()->isLoopInvariant(PSE.getSCEV(ScalarParam),
TheLoop))
ParamsOk = false;
break;
}
case VFParamKind::GlobalPredicate:
UsesMask = true;
break;
Expand Down
81 changes: 68 additions & 13 deletions llvm/test/Transforms/LoopVectorize/AArch64/uniform-args-call-variants.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,17 +9,25 @@ define void @test_uniform(ptr noalias %dst, ptr readonly %src, i64 %uniform , i6
; CHECK-LABEL: define void @test_uniform
; CHECK-SAME: (ptr noalias [[DST:%.*]], ptr readonly [[SRC:%.*]], i64 [[UNIFORM:%.*]], i64 [[N:%.*]]) #[[ATTR0:[0-9]+]] {
; CHECK-NEXT: entry:
; CHECK-NEXT: br label [[FOR_BODY:%.*]]
; CHECK: for.body:
; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; CHECK-NEXT: [[GEPSRC:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDVARS_IV]]
; CHECK-NEXT: [[DATA:%.*]] = load double, ptr [[GEPSRC]], align 8
; CHECK-NEXT: [[CALL:%.*]] = call double @foo(double [[DATA]], i64 [[UNIFORM]]) #[[ATTR1:[0-9]+]]
; CHECK-NEXT: [[GEPDST:%.*]] = getelementptr inbounds double, ptr [[DST]], i64 [[INDVARS_IV]]
; CHECK-NEXT: store double [[CALL]], ptr [[GEPDST]], align 8
; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
; CHECK-NEXT: [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT]], [[N]]
; CHECK-NEXT: br i1 [[EXITCOND]], label [[FOR_COND_CLEANUP:%.*]], label [[FOR_BODY]]
; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
; CHECK-NEXT: [[TMP1:%.*]] = shl i64 [[TMP0]], 1
; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.usub.sat.i64(i64 [[N]], i64 [[TMP1]])
; CHECK-NEXT: [[ACTIVE_LANE_MASK_ENTRY:%.*]] = call <vscale x 2 x i1> @llvm.get.active.lane.mask.nxv2i1.i64(i64 0, i64 [[N]])
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
; CHECK: vector.body:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
; CHECK-NEXT: [[ACTIVE_LANE_MASK:%.*]] = phi <vscale x 2 x i1> [ [[ACTIVE_LANE_MASK_ENTRY]], [[ENTRY]] ], [ [[ACTIVE_LANE_MASK_NEXT:%.*]], [[VECTOR_BODY]] ]
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDEX]]
; CHECK-NEXT: [[WIDE_MASKED_LOAD:%.*]] = call <vscale x 2 x double> @llvm.masked.load.nxv2f64.p0(ptr [[TMP3]], i32 8, <vscale x 2 x i1> [[ACTIVE_LANE_MASK]], <vscale x 2 x double> poison)
; CHECK-NEXT: [[TMP4:%.*]] = call <vscale x 2 x double> @foo_uniform(<vscale x 2 x double> [[WIDE_MASKED_LOAD]], i64 [[UNIFORM]], <vscale x 2 x i1> [[ACTIVE_LANE_MASK]])
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds double, ptr [[DST]], i64 [[INDEX]]
; CHECK-NEXT: call void @llvm.masked.store.nxv2f64.p0(<vscale x 2 x double> [[TMP4]], ptr [[TMP5]], i32 8, <vscale x 2 x i1> [[ACTIVE_LANE_MASK]])
; CHECK-NEXT: [[TMP6:%.*]] = call i64 @llvm.vscale.i64()
; CHECK-NEXT: [[TMP7:%.*]] = shl i64 [[TMP6]], 1
; CHECK-NEXT: [[INDEX_NEXT]] = add i64 [[INDEX]], [[TMP7]]
; CHECK-NEXT: [[ACTIVE_LANE_MASK_NEXT]] = call <vscale x 2 x i1> @llvm.get.active.lane.mask.nxv2i1.i64(i64 [[INDEX]], i64 [[TMP2]])
; CHECK-NEXT: [[TMP8:%.*]] = extractelement <vscale x 2 x i1> [[ACTIVE_LANE_MASK_NEXT]], i64 0
; CHECK-NEXT: br i1 [[TMP8]], label [[VECTOR_BODY]], label [[FOR_COND_CLEANUP:%.*]], !llvm.loop [[LOOP0:![0-9]+]]
; CHECK: for.cond.cleanup:
; CHECK-NEXT: ret void
;
Expand All @@ -41,6 +49,50 @@ for.cond.cleanup:
ret void
}

define void @test_uniform_smaller_scalar(ptr noalias %dst, ptr readonly %src, i32 %uniform , i64 %n) #0 {
; CHECK-LABEL: define void @test_uniform_smaller_scalar
; CHECK-SAME: (ptr noalias [[DST:%.*]], ptr readonly [[SRC:%.*]], i32 [[UNIFORM:%.*]], i64 [[N:%.*]]) #[[ATTR0]] {
; CHECK-NEXT: entry:
; CHECK-NEXT: [[TMP0:%.*]] = call i64 @llvm.vscale.i64()
; CHECK-NEXT: [[TMP1:%.*]] = shl i64 [[TMP0]], 1
; CHECK-NEXT: [[TMP2:%.*]] = call i64 @llvm.usub.sat.i64(i64 [[N]], i64 [[TMP1]])
; CHECK-NEXT: [[ACTIVE_LANE_MASK_ENTRY:%.*]] = call <vscale x 2 x i1> @llvm.get.active.lane.mask.nxv2i1.i64(i64 0, i64 [[N]])
; CHECK-NEXT: br label [[VECTOR_BODY:%.*]]
; CHECK: vector.body:
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
; CHECK-NEXT: [[ACTIVE_LANE_MASK:%.*]] = phi <vscale x 2 x i1> [ [[ACTIVE_LANE_MASK_ENTRY]], [[ENTRY]] ], [ [[ACTIVE_LANE_MASK_NEXT:%.*]], [[VECTOR_BODY]] ]
; CHECK-NEXT: [[TMP3:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDEX]]
; CHECK-NEXT: [[WIDE_MASKED_LOAD:%.*]] = call <vscale x 2 x double> @llvm.masked.load.nxv2f64.p0(ptr [[TMP3]], i32 8, <vscale x 2 x i1> [[ACTIVE_LANE_MASK]], <vscale x 2 x double> poison)
; CHECK-NEXT: [[TMP4:%.*]] = call <vscale x 2 x double> @bar_uniform(<vscale x 2 x double> [[WIDE_MASKED_LOAD]], i32 [[UNIFORM]], <vscale x 2 x i1> [[ACTIVE_LANE_MASK]])
; CHECK-NEXT: [[TMP5:%.*]] = getelementptr inbounds double, ptr [[DST]], i64 [[INDEX]]
; CHECK-NEXT: call void @llvm.masked.store.nxv2f64.p0(<vscale x 2 x double> [[TMP4]], ptr [[TMP5]], i32 8, <vscale x 2 x i1> [[ACTIVE_LANE_MASK]])
; CHECK-NEXT: [[TMP6:%.*]] = call i64 @llvm.vscale.i64()
; CHECK-NEXT: [[TMP7:%.*]] = shl i64 [[TMP6]], 1
; CHECK-NEXT: [[INDEX_NEXT]] = add i64 [[INDEX]], [[TMP7]]
; CHECK-NEXT: [[ACTIVE_LANE_MASK_NEXT]] = call <vscale x 2 x i1> @llvm.get.active.lane.mask.nxv2i1.i64(i64 [[INDEX]], i64 [[TMP2]])
; CHECK-NEXT: [[TMP8:%.*]] = extractelement <vscale x 2 x i1> [[ACTIVE_LANE_MASK_NEXT]], i64 0
; CHECK-NEXT: br i1 [[TMP8]], label [[VECTOR_BODY]], label [[FOR_COND_CLEANUP:%.*]], !llvm.loop [[LOOP3:![0-9]+]]
; CHECK: for.cond.cleanup:
; CHECK-NEXT: ret void
;
entry:
br label %for.body

for.body:
%indvars.iv = phi i64 [ 0, %entry ], [ %indvars.iv.next, %for.body ]
%gepsrc = getelementptr double, ptr %src, i64 %indvars.iv
%data = load double, ptr %gepsrc, align 8
%call = call double @bar(double %data, i32 %uniform) #2
%gepdst = getelementptr inbounds double, ptr %dst, i64 %indvars.iv
store double %call, ptr %gepdst
%indvars.iv.next = add nuw nsw i64 %indvars.iv, 1
%exitcond = icmp eq i64 %indvars.iv.next, %n
br i1 %exitcond, label %for.cond.cleanup, label %for.body

for.cond.cleanup:
ret void
}

; If the parameter is not uniform, then we can't use the vector variant.
define void @test_uniform_not_invariant(ptr noalias %dst, ptr readonly %src, i64 %n) #0 {
; CHECK-LABEL: define void @test_uniform_not_invariant
Expand All @@ -51,7 +103,7 @@ define void @test_uniform_not_invariant(ptr noalias %dst, ptr readonly %src, i64
; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ 0, [[ENTRY:%.*]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; CHECK-NEXT: [[GEPSRC:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDVARS_IV]]
; CHECK-NEXT: [[DATA:%.*]] = load double, ptr [[GEPSRC]], align 8
; CHECK-NEXT: [[CALL:%.*]] = call double @foo(double [[DATA]], i64 [[INDVARS_IV]]) #[[ATTR1]]
; CHECK-NEXT: [[CALL:%.*]] = call double @foo(double [[DATA]], i64 [[INDVARS_IV]]) #[[ATTR5:[0-9]+]]
; CHECK-NEXT: [[GEPDST:%.*]] = getelementptr inbounds double, ptr [[DST]], i64 [[INDVARS_IV]]
; CHECK-NEXT: store double [[CALL]], ptr [[GEPDST]], align 8
; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
Expand Down Expand Up @@ -80,11 +132,14 @@ for.cond.cleanup:

; Scalar functions
declare double @foo(double, i64)
declare double @bar(double, i32)

; Vector variants
declare <vscale x 2 x double> @foo_uniform(<vscale x 2 x double>, i64, <vscale x 2 x i1>)
declare <vscale x 2 x double> @bar_uniform(<vscale x 2 x double>, i32, <vscale x 2 x i1>)

attributes #0 = { "target-features"="+sve" }

; Mappings
attributes #1 = { nounwind "vector-function-abi-variant"="_ZGV_LLVM_Mxvu_foo(foo_uniform)" }
attributes #1 = { nounwind "vector-function-abi-variant"="_ZGVsMxvu_foo(foo_uniform)" }
attributes #2 = { nounwind "vector-function-abi-variant"="_ZGVsMxvu_bar(bar_uniform)" }
17 changes: 6 additions & 11 deletions llvm/test/Transforms/LoopVectorize/uniform-args-call-variants.ll
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,17 +16,12 @@ define void @test_uniform(ptr noalias %dst, ptr readonly %src, i64 %uniform , i6
; CHECK-NEXT: [[INDEX:%.*]] = phi i64 [ 0, [[VECTOR_PH]] ], [ [[INDEX_NEXT:%.*]], [[VECTOR_BODY]] ]
; CHECK-NEXT: [[TMP0:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDEX]]
; CHECK-NEXT: [[WIDE_LOAD:%.*]] = load <2 x double>, ptr [[TMP0]], align 8
; CHECK-NEXT: [[TMP1:%.*]] = extractelement <2 x double> [[WIDE_LOAD]], i64 0
; CHECK-NEXT: [[TMP2:%.*]] = call double @foo(double [[TMP1]], i64 [[UNIFORM]]) #[[ATTR0:[0-9]+]]
; CHECK-NEXT: [[TMP3:%.*]] = extractelement <2 x double> [[WIDE_LOAD]], i64 1
; CHECK-NEXT: [[TMP4:%.*]] = call double @foo(double [[TMP3]], i64 [[UNIFORM]]) #[[ATTR0]]
; CHECK-NEXT: [[TMP5:%.*]] = insertelement <2 x double> poison, double [[TMP2]], i64 0
; CHECK-NEXT: [[TMP6:%.*]] = insertelement <2 x double> [[TMP5]], double [[TMP4]], i64 1
; CHECK-NEXT: [[TMP7:%.*]] = getelementptr inbounds double, ptr [[DST]], i64 [[INDEX]]
; CHECK-NEXT: store <2 x double> [[TMP6]], ptr [[TMP7]], align 8
; CHECK-NEXT: [[TMP1:%.*]] = call <2 x double> @foo_uniform(<2 x double> [[WIDE_LOAD]], i64 [[UNIFORM]])
; CHECK-NEXT: [[TMP2:%.*]] = getelementptr inbounds double, ptr [[DST]], i64 [[INDEX]]
; CHECK-NEXT: store <2 x double> [[TMP1]], ptr [[TMP2]], align 8
; CHECK-NEXT: [[INDEX_NEXT]] = add nuw i64 [[INDEX]], 2
; CHECK-NEXT: [[TMP8:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP8]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; CHECK-NEXT: [[TMP3:%.*]] = icmp eq i64 [[INDEX_NEXT]], [[N_VEC]]
; CHECK-NEXT: br i1 [[TMP3]], label [[MIDDLE_BLOCK:%.*]], label [[VECTOR_BODY]], !llvm.loop [[LOOP0:![0-9]+]]
; CHECK: middle.block:
; CHECK-NEXT: [[CMP_N:%.*]] = icmp eq i64 [[N_VEC]], [[N]]
; CHECK-NEXT: br i1 [[CMP_N]], label [[FOR_COND_CLEANUP:%.*]], label [[SCALAR_PH]]
Expand All @@ -37,7 +32,7 @@ define void @test_uniform(ptr noalias %dst, ptr readonly %src, i64 %uniform , i6
; CHECK-NEXT: [[INDVARS_IV:%.*]] = phi i64 [ [[BC_RESUME_VAL]], [[SCALAR_PH]] ], [ [[INDVARS_IV_NEXT:%.*]], [[FOR_BODY]] ]
; CHECK-NEXT: [[GEPSRC:%.*]] = getelementptr double, ptr [[SRC]], i64 [[INDVARS_IV]]
; CHECK-NEXT: [[DATA:%.*]] = load double, ptr [[GEPSRC]], align 8
; CHECK-NEXT: [[CALL:%.*]] = call double @foo(double [[DATA]], i64 [[UNIFORM]]) #[[ATTR0]]
; CHECK-NEXT: [[CALL:%.*]] = call double @foo(double [[DATA]], i64 [[UNIFORM]]) #[[ATTR0:[0-9]+]]
; CHECK-NEXT: [[GEPDST:%.*]] = getelementptr inbounds double, ptr [[DST]], i64 [[INDVARS_IV]]
; CHECK-NEXT: store double [[CALL]], ptr [[GEPDST]], align 8
; CHECK-NEXT: [[INDVARS_IV_NEXT]] = add nuw nsw i64 [[INDVARS_IV]], 1
Expand Down