; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
; RUN: opt -passes=slp-vectorizer,verify -S < %s -mtriple=x86_64-unknown-linux -mcpu=corei7-avx | FileCheck %s -check-prefix=ENABLED
;
; Without supernode operand reordering, this does not get fully vectorized.
; S[0] = (A[0] + B[0]) + C[0]
; S[1] = (B[1] + C[1]) + A[1]
define void @test_supernode_add(ptr %Aarray, ptr %Barray, ptr %Carray, ptr %Sarray) {
; ENABLED-LABEL: @test_supernode_add(
; ENABLED-NEXT: entry:
; ENABLED-NEXT: [[TMP0:%.*]] = load <2 x double>, ptr [[AARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP1:%.*]] = load <2 x double>, ptr [[BARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP2:%.*]] = load <2 x double>, ptr [[CARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP3:%.*]] = shufflevector <2 x double> [[TMP0]], <2 x double> [[TMP2]], <2 x i32> <i32 0, i32 3>
; ENABLED-NEXT: [[TMP4:%.*]] = fadd fast <2 x double> [[TMP1]], [[TMP3]]
; ENABLED-NEXT: [[TMP5:%.*]] = shufflevector <2 x double> [[TMP2]], <2 x double> [[TMP0]], <2 x i32> <i32 0, i32 3>
; ENABLED-NEXT: [[TMP6:%.*]] = fadd fast <2 x double> [[TMP4]], [[TMP5]]
; ENABLED-NEXT: store <2 x double> [[TMP6]], ptr [[SARRAY:%.*]], align 8
; ENABLED-NEXT: ret void
;
entry:
%idxA1 = getelementptr inbounds double, ptr %Aarray, i64 1
%idxB1 = getelementptr inbounds double, ptr %Barray, i64 1
%idxC1 = getelementptr inbounds double, ptr %Carray, i64 1
%idxS1 = getelementptr inbounds double, ptr %Sarray, i64 1
%A0 = load double, ptr %Aarray, align 8
%A1 = load double, ptr %idxA1, align 8
%B0 = load double, ptr %Barray, align 8
%B1 = load double, ptr %idxB1, align 8
%C0 = load double, ptr %Carray, align 8
%C1 = load double, ptr %idxC1, align 8
%addA0B0 = fadd fast double %A0, %B0
%addB1C1 = fadd fast double %B1, %C1
%add0 = fadd fast double %addA0B0, %C0
%add1 = fadd fast double %addB1C1, %A1
store double %add0, ptr %Sarray, align 8
store double %add1, ptr %idxS1, align 8
ret void
}
; Without supernode operand reordering, this does not get fully vectorized.
; S[0] = (A[0] - B[0]) + C[0]
; S[1] = (C[1] - B[1]) + A[1]
define void @test_supernode_addsub(ptr %Aarray, ptr %Barray, ptr %Carray, ptr %Sarray) {
; ENABLED-LABEL: @test_supernode_addsub(
; ENABLED-NEXT: entry:
; ENABLED-NEXT: [[TMP0:%.*]] = load <2 x double>, ptr [[AARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP1:%.*]] = load <2 x double>, ptr [[BARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP2:%.*]] = load <2 x double>, ptr [[CARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP3:%.*]] = shufflevector <2 x double> [[TMP0]], <2 x double> [[TMP2]], <2 x i32> <i32 0, i32 3>
; ENABLED-NEXT: [[TMP4:%.*]] = fsub fast <2 x double> [[TMP3]], [[TMP1]]
; ENABLED-NEXT: [[TMP5:%.*]] = shufflevector <2 x double> [[TMP2]], <2 x double> [[TMP0]], <2 x i32> <i32 0, i32 3>
; ENABLED-NEXT: [[TMP6:%.*]] = fadd fast <2 x double> [[TMP4]], [[TMP5]]
; ENABLED-NEXT: store <2 x double> [[TMP6]], ptr [[SARRAY:%.*]], align 8
; ENABLED-NEXT: ret void
;
entry:
%idxA1 = getelementptr inbounds double, ptr %Aarray, i64 1
%idxB1 = getelementptr inbounds double, ptr %Barray, i64 1
%idxC1 = getelementptr inbounds double, ptr %Carray, i64 1
%idxS1 = getelementptr inbounds double, ptr %Sarray, i64 1
%A0 = load double, ptr %Aarray, align 8
%A1 = load double, ptr %idxA1, align 8
%B0 = load double, ptr %Barray, align 8
%B1 = load double, ptr %idxB1, align 8
%C0 = load double, ptr %Carray, align 8
%C1 = load double, ptr %idxC1, align 8
%subA0B0 = fsub fast double %A0, %B0
%subC1B1 = fsub fast double %C1, %B1
%add0 = fadd fast double %subA0B0, %C0
%add1 = fadd fast double %subC1B1, %A1
store double %add0, ptr %Sarray, align 8
store double %add1, ptr %idxS1, align 8
ret void
}
; Without supernode operand reordering, this does not get fully vectorized.
; This checks that the super-node works with alternate sequences.
;
; S[0] = (A[0] - B[0]) - C[0]
; S[1] = (B[1] + C[1]) + A[1]
define void @test_supernode_addsub_alt(ptr %Aarray, ptr %Barray, ptr %Carray, ptr %Sarray) {
; ENABLED-LABEL: @test_supernode_addsub_alt(
; ENABLED-NEXT: entry:
; ENABLED-NEXT: [[TMP0:%.*]] = load <2 x double>, ptr [[AARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP1:%.*]] = load <2 x double>, ptr [[BARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP2:%.*]] = load <2 x double>, ptr [[CARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP3:%.*]] = shufflevector <2 x double> [[TMP0]], <2 x double> [[TMP2]], <2 x i32> <i32 0, i32 3>
; ENABLED-NEXT: [[TMP4:%.*]] = fsub fast <2 x double> [[TMP3]], [[TMP1]]
; ENABLED-NEXT: [[TMP5:%.*]] = fadd fast <2 x double> [[TMP3]], [[TMP1]]
; ENABLED-NEXT: [[TMP6:%.*]] = shufflevector <2 x double> [[TMP4]], <2 x double> [[TMP5]], <2 x i32> <i32 0, i32 3>
; ENABLED-NEXT: [[TMP7:%.*]] = shufflevector <2 x double> [[TMP2]], <2 x double> [[TMP0]], <2 x i32> <i32 0, i32 3>
; ENABLED-NEXT: [[TMP8:%.*]] = fsub fast <2 x double> [[TMP6]], [[TMP7]]
; ENABLED-NEXT: [[TMP9:%.*]] = fadd fast <2 x double> [[TMP6]], [[TMP7]]
; ENABLED-NEXT: [[TMP10:%.*]] = shufflevector <2 x double> [[TMP8]], <2 x double> [[TMP9]], <2 x i32> <i32 0, i32 3>
; ENABLED-NEXT: store <2 x double> [[TMP10]], ptr [[SARRAY:%.*]], align 8
; ENABLED-NEXT: ret void
;
entry:
%idxA1 = getelementptr inbounds double, ptr %Aarray, i64 1
%idxB1 = getelementptr inbounds double, ptr %Barray, i64 1
%idxC1 = getelementptr inbounds double, ptr %Carray, i64 1
%idxS1 = getelementptr inbounds double, ptr %Sarray, i64 1
%A0 = load double, ptr %Aarray, align 8
%A1 = load double, ptr %idxA1, align 8
%B0 = load double, ptr %Barray, align 8
%B1 = load double, ptr %idxB1, align 8
%C0 = load double, ptr %Carray, align 8
%C1 = load double, ptr %idxC1, align 8
%subA0B0 = fsub fast double %A0, %B0
%addB1C1 = fadd fast double %B1, %C1
%sub0 = fsub fast double %subA0B0, %C0
%add1 = fadd fast double %addB1C1, %A1
store double %sub0, ptr %Sarray, align 8
store double %add1, ptr %idxS1, align 8
ret void
}
; This checks that vectorizeTree() works correctly with the supernode
; and does not generate uses before defs.
; If all of the operands of the supernode are vectorizable, then the scheduler
; will fix their position in the program. If not, then the scheduler may not
; touch them, leading to uses before defs.
;
; A0 = ...
; C = ...
; t1 = A0 + C
; B0 = ...
; t2 = t1 + B0
; A1 = ...
; B1 = ...
; t3 = A1 + B1
; D = ...
; t4 = t3 + D
;
;
; A0 C A1 B1 A0 C A1 D A0:1 C,D
; \ / \ / Reorder \ / \ / Bundles \ /
; t1 + B0 t3 + D -------> t1 + B0 t3 + B1 ------> t1:3 + B0:1
; |/ |/ |/ |/ |/
; t2 + t4 + t2 + t4 + t2:4 +
;
; After reordering, 'D' conceptually becomes an operand of t3:
; t3 = A1 + D
; But D is defined *after* its use.
;
define void @supernode_scheduling(ptr %Aarray, ptr %Barray, ptr %Carray, ptr %Darray, ptr %Sarray) {
; ENABLED-LABEL: @supernode_scheduling(
; ENABLED-NEXT: entry:
; ENABLED-NEXT: [[C:%.*]] = load double, ptr [[CARRAY:%.*]], align 8
; ENABLED-NEXT: [[D:%.*]] = load double, ptr [[DARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP0:%.*]] = load <2 x double>, ptr [[AARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP1:%.*]] = load <2 x double>, ptr [[BARRAY:%.*]], align 8
; ENABLED-NEXT: [[TMP2:%.*]] = insertelement <2 x double> [[TMP1]], double [[C]], i32 0
; ENABLED-NEXT: [[TMP3:%.*]] = fadd fast <2 x double> [[TMP0]], [[TMP2]]
; ENABLED-NEXT: [[TMP4:%.*]] = insertelement <2 x double> [[TMP1]], double [[D]], i32 1
; ENABLED-NEXT: [[TMP5:%.*]] = fadd fast <2 x double> [[TMP3]], [[TMP4]]
; ENABLED-NEXT: store <2 x double> [[TMP5]], ptr [[SARRAY:%.*]], align 8
; ENABLED-NEXT: ret void
;
entry:
%idxA1 = getelementptr inbounds double, ptr %Aarray, i64 1
%idxB1 = getelementptr inbounds double, ptr %Barray, i64 1
%idxS1 = getelementptr inbounds double, ptr %Sarray, i64 1
%A0 = load double, ptr %Aarray, align 8
%C = load double, ptr %Carray, align 8
%t1 = fadd fast double %A0, %C
%B0 = load double, ptr %Barray, align 8
%t2 = fadd fast double %t1, %B0
%A1 = load double, ptr %idxA1, align 8
%B1 = load double, ptr %idxB1, align 8
%t3 = fadd fast double %A1, %B1
%D = load double, ptr %Darray, align 8
%t4 = fadd fast double %t3, %D
store double %t2, ptr %Sarray, align 8
store double %t4, ptr %idxS1, align 8
ret void
}
; The SLP scheduler has trouble moving instructions across blocks.
; Even though we can build a SuperNode for this example, we should not because the scheduler
; cannot handle the cross-block instruction motion that is required once the operands of the
; SuperNode are reordered.
;
; bb1:
; A0 = ...
; B1 = ...
; Tmp0 = A0 + 2.0
; Tmp1 = B1 + 2.0
;
; bb2:
; A1 = ...
; B0 = ...
; S[0] = Tmp0 + B0
; S[1] = Tmp1 + A1
define void @supernode_scheduling_cross_block(ptr %Aarray, ptr %Barray, ptr %Sarray) {
; ENABLED-LABEL: @supernode_scheduling_cross_block(
; ENABLED-NEXT: entry:
; ENABLED-NEXT: [[IDXA1:%.*]] = getelementptr inbounds double, ptr [[AARRAY:%.*]], i64 1
; ENABLED-NEXT: [[IDXB1:%.*]] = getelementptr inbounds double, ptr [[BARRAY:%.*]], i64 1
; ENABLED-NEXT: [[A0:%.*]] = load double, ptr [[AARRAY]], align 8
; ENABLED-NEXT: [[B1:%.*]] = load double, ptr [[IDXB1]], align 8
; ENABLED-NEXT: [[TMP0:%.*]] = insertelement <2 x double> poison, double [[A0]], i32 0
; ENABLED-NEXT: [[TMP1:%.*]] = insertelement <2 x double> [[TMP0]], double [[B1]], i32 1
; ENABLED-NEXT: [[TMP2:%.*]] = fadd fast <2 x double> [[TMP1]], <double 2.000000e+00, double 2.000000e+00>
; ENABLED-NEXT: br label [[BB:%.*]]
; ENABLED: bb:
; ENABLED-NEXT: [[A1:%.*]] = load double, ptr [[IDXA1]], align 8
; ENABLED-NEXT: [[B0:%.*]] = load double, ptr [[BARRAY]], align 8
; ENABLED-NEXT: [[TMP3:%.*]] = insertelement <2 x double> poison, double [[B0]], i32 0
; ENABLED-NEXT: [[TMP4:%.*]] = insertelement <2 x double> [[TMP3]], double [[A1]], i32 1
; ENABLED-NEXT: [[TMP5:%.*]] = fadd fast <2 x double> [[TMP2]], [[TMP4]]
; ENABLED-NEXT: store <2 x double> [[TMP5]], ptr [[SARRAY:%.*]], align 8
; ENABLED-NEXT: ret void
;
entry:
%idxA1 = getelementptr inbounds double, ptr %Aarray, i64 1
%idxB1 = getelementptr inbounds double, ptr %Barray, i64 1
%idxS1 = getelementptr inbounds double, ptr %Sarray, i64 1
%A0 = load double, ptr %Aarray, align 8
%B1 = load double, ptr %idxB1, align 8
%Tmp0 = fadd fast double %A0, 2.0
%Tmp1 = fadd fast double %B1, 2.0
br label %bb
bb:
%A1 = load double, ptr %idxA1, align 8
%B0 = load double, ptr %Barray, align 8
%Sum0 = fadd fast double %Tmp0, %B0
%Sum1 = fadd fast double %Tmp1, %A1
store double %Sum0, ptr %Sarray, align 8
store double %Sum1, ptr %idxS1, align 8
ret void
}
Codebase Browser
llvm