[AArch64][GlobalISel] Full reverse shuffles. (#119083)

davemgreen · web-flow · commit 01c8cd664a9b · 2024-12-29T15:56:12.000Z
A full shuffle reverse needs to use EXT+REV64. This adds handling for more types than SDAG so long as the mask is isReverseMask to make the patterns simpler.
diff --git a/llvm/lib/Target/AArch64/AArch64Combine.td b/llvm/lib/Target/AArch64/AArch64Combine.td
@@ -131,6 +131,15 @@ def ext: GICombineRule <
   (apply [{ applyEXT(*${root}, ${matchinfo}); }])
 >;
 
+def fullrev: GICombineRule <
+  (defs root:$root, shuffle_matchdata:$matchinfo),
+  (match (G_IMPLICIT_DEF $src2),
+         (G_SHUFFLE_VECTOR $src, $src1, $src2, $mask):$root,
+         [{ return ShuffleVectorInst::isReverseMask(${mask}.getShuffleMask(),
+                                                    ${mask}.getShuffleMask().size()); }]),
+  (apply [{ applyFullRev(*${root}, MRI); }])
+>;
+
 def insertelt_nonconst: GICombineRule <
   (defs root:$root, shuffle_matchdata:$matchinfo),
   (match (wip_match_opcode G_INSERT_VECTOR_ELT):$root,
@@ -163,7 +172,7 @@ def form_duplane : GICombineRule <
   (apply [{ applyDupLane(*${root}, MRI, B, ${matchinfo}); }])
 >;
 
-def shuffle_vector_lowering : GICombineGroup<[dup, rev, ext, zip, uzp, trn,
+def shuffle_vector_lowering : GICombineGroup<[dup, rev, ext, zip, uzp, trn, fullrev,
                                               form_duplane, shuf_to_ins]>;
 
 // Turn G_UNMERGE_VALUES -> G_EXTRACT_VECTOR_ELT's
diff --git a/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerLowering.cpp b/llvm/lib/Target/AArch64/GISel/AArch64PostLegalizerLowering.cpp
@@ -405,6 +405,19 @@ void applyEXT(MachineInstr &MI, ShuffleVectorPseudo &MatchInfo) {
   MI.eraseFromParent();
 }
 
+void applyFullRev(MachineInstr &MI, MachineRegisterInfo &MRI) {
+  Register Dst = MI.getOperand(0).getReg();
+  Register Src = MI.getOperand(1).getReg();
+  LLT DstTy = MRI.getType(Dst);
+  assert(DstTy.getSizeInBits() == 128 &&
+         "Expected 128bit vector in applyFullRev");
+  MachineIRBuilder MIRBuilder(MI);
+  auto Cst = MIRBuilder.buildConstant(LLT::scalar(32), 8);
+  auto Rev = MIRBuilder.buildInstr(AArch64::G_REV64, {DstTy}, {Src});
+  MIRBuilder.buildInstr(AArch64::G_EXT, {Dst}, {Rev, Rev, Cst});
+  MI.eraseFromParent();
+}
+
 bool matchNonConstInsert(MachineInstr &MI, MachineRegisterInfo &MRI) {
   assert(MI.getOpcode() == TargetOpcode::G_INSERT_VECTOR_ELT);
 
diff --git a/llvm/test/CodeGen/AArch64/aarch64-dup-ext.ll b/llvm/test/CodeGen/AArch64/aarch64-dup-ext.ll
@@ -440,11 +440,10 @@ define <8 x i16> @shufsext_v8i8_v8i16(<8 x i8> %src, <8 x i8> %b) {
 ;
 ; CHECK-GI-LABEL: shufsext_v8i8_v8i16:
 ; CHECK-GI:       // %bb.0: // %entry
-; CHECK-GI-NEXT:    adrp x8, .LCPI14_0
-; CHECK-GI-NEXT:    sshll v2.8h, v0.8b, #0
+; CHECK-GI-NEXT:    sshll v0.8h, v0.8b, #0
 ; CHECK-GI-NEXT:    sshll v1.8h, v1.8b, #0
-; CHECK-GI-NEXT:    ldr q0, [x8, :lo12:.LCPI14_0]
-; CHECK-GI-NEXT:    tbl v0.16b, { v2.16b, v3.16b }, v0.16b
+; CHECK-GI-NEXT:    rev64 v0.8h, v0.8h
+; CHECK-GI-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
 ; CHECK-GI-NEXT:    mul v0.8h, v0.8h, v1.8h
 ; CHECK-GI-NEXT:    ret
 entry:
@@ -493,11 +492,10 @@ define <8 x i16> @shufzext_v8i8_v8i16(<8 x i8> %src, <8 x i8> %b) {
 ;
 ; CHECK-GI-LABEL: shufzext_v8i8_v8i16:
 ; CHECK-GI:       // %bb.0: // %entry
-; CHECK-GI-NEXT:    adrp x8, .LCPI16_0
-; CHECK-GI-NEXT:    ushll v2.8h, v0.8b, #0
+; CHECK-GI-NEXT:    ushll v0.8h, v0.8b, #0
 ; CHECK-GI-NEXT:    ushll v1.8h, v1.8b, #0
-; CHECK-GI-NEXT:    ldr q0, [x8, :lo12:.LCPI16_0]
-; CHECK-GI-NEXT:    tbl v0.16b, { v2.16b, v3.16b }, v0.16b
+; CHECK-GI-NEXT:    rev64 v0.8h, v0.8h
+; CHECK-GI-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
 ; CHECK-GI-NEXT:    mul v0.8h, v0.8h, v1.8h
 ; CHECK-GI-NEXT:    ret
 entry:
diff --git a/llvm/test/CodeGen/AArch64/neon-reverseshuffle.ll b/llvm/test/CodeGen/AArch64/neon-reverseshuffle.ll
@@ -23,19 +23,11 @@ entry:
 }
 
 define <4 x i32> @v4i32(<4 x i32> %a) {
-; CHECK-SD-LABEL: v4i32:
-; CHECK-SD:       // %bb.0: // %entry
-; CHECK-SD-NEXT:    rev64 v0.4s, v0.4s
-; CHECK-SD-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
-; CHECK-SD-NEXT:    ret
-;
-; CHECK-GI-LABEL: v4i32:
-; CHECK-GI:       // %bb.0: // %entry
-; CHECK-GI-NEXT:    adrp x8, .LCPI2_0
-; CHECK-GI-NEXT:    // kill: def $q0 killed $q0 def $q0_q1
-; CHECK-GI-NEXT:    ldr q2, [x8, :lo12:.LCPI2_0]
-; CHECK-GI-NEXT:    tbl v0.16b, { v0.16b, v1.16b }, v2.16b
-; CHECK-GI-NEXT:    ret
+; CHECK-LABEL: v4i32:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    rev64 v0.4s, v0.4s
+; CHECK-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    ret
 entry:
   %V128 = shufflevector <4 x i32> %a, <4 x i32> undef, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
   ret <4 x i32> %V128
@@ -52,19 +44,11 @@ entry:
 }
 
 define <8 x i16> @v8i16(<8 x i16> %a) {
-; CHECK-SD-LABEL: v8i16:
-; CHECK-SD:       // %bb.0: // %entry
-; CHECK-SD-NEXT:    rev64 v0.8h, v0.8h
-; CHECK-SD-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
-; CHECK-SD-NEXT:    ret
-;
-; CHECK-GI-LABEL: v8i16:
-; CHECK-GI:       // %bb.0: // %entry
-; CHECK-GI-NEXT:    adrp x8, .LCPI4_0
-; CHECK-GI-NEXT:    // kill: def $q0 killed $q0 def $q0_q1
-; CHECK-GI-NEXT:    ldr q2, [x8, :lo12:.LCPI4_0]
-; CHECK-GI-NEXT:    tbl v0.16b, { v0.16b, v1.16b }, v2.16b
-; CHECK-GI-NEXT:    ret
+; CHECK-LABEL: v8i16:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    rev64 v0.8h, v0.8h
+; CHECK-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    ret
 entry:
   %V128 = shufflevector <8 x i16> %a, <8 x i16> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
   ret <8 x i16> %V128
@@ -93,6 +77,22 @@ entry:
   ret <8 x i16> %V128
 }
 
+define <4 x i16> @v8i16_3(<8 x i16> %a) {
+; CHECK-SD-LABEL: v8i16_3:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    rev64 v0.4h, v0.4h
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: v8i16_3:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    rev64 v0.8h, v0.8h
+; CHECK-GI-NEXT:    // kill: def $d0 killed $d0 killed $q0
+; CHECK-GI-NEXT:    ret
+entry:
+  %V128 = shufflevector <8 x i16> %a, <8 x i16> poison, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
+  ret <4 x i16> %V128
+}
+
 define <4 x i16> @v4i16(<4 x i16> %a) {
 ; CHECK-LABEL: v4i16:
 ; CHECK:       // %bb.0: // %entry
@@ -104,19 +104,11 @@ entry:
 }
 
 define <16 x i8> @v16i8(<16 x i8> %a) {
-; CHECK-SD-LABEL: v16i8:
-; CHECK-SD:       // %bb.0: // %entry
-; CHECK-SD-NEXT:    rev64 v0.16b, v0.16b
-; CHECK-SD-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
-; CHECK-SD-NEXT:    ret
-;
-; CHECK-GI-LABEL: v16i8:
-; CHECK-GI:       // %bb.0: // %entry
-; CHECK-GI-NEXT:    adrp x8, .LCPI7_0
-; CHECK-GI-NEXT:    // kill: def $q0 killed $q0 def $q0_q1
-; CHECK-GI-NEXT:    ldr q2, [x8, :lo12:.LCPI7_0]
-; CHECK-GI-NEXT:    tbl v0.16b, { v0.16b, v1.16b }, v2.16b
-; CHECK-GI-NEXT:    ret
+; CHECK-LABEL: v16i8:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    rev64 v0.16b, v0.16b
+; CHECK-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    ret
 entry:
   %V128 = shufflevector <16 x i8> %a, <16 x i8> undef, <16 x i32> <i32 15, i32 14, i32 13, i32 12, i32 11, i32 10, i32 9, i32 8, i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
   ret <16 x i8> %V128
@@ -125,18 +117,18 @@ entry:
 define <16 x i8> @v16i8_2(<8 x i8> %a, <8 x i8> %b) {
 ; CHECK-SD-LABEL: v16i8_2:
 ; CHECK-SD:       // %bb.0: // %entry
-; CHECK-SD-NEXT:    adrp x8, .LCPI8_0
+; CHECK-SD-NEXT:    adrp x8, .LCPI9_0
 ; CHECK-SD-NEXT:    // kill: def $d1 killed $d1 killed $q0_q1 def $q0_q1
-; CHECK-SD-NEXT:    ldr q2, [x8, :lo12:.LCPI8_0]
+; CHECK-SD-NEXT:    ldr q2, [x8, :lo12:.LCPI9_0]
 ; CHECK-SD-NEXT:    // kill: def $d0 killed $d0 killed $q0_q1 def $q0_q1
 ; CHECK-SD-NEXT:    tbl v0.16b, { v0.16b, v1.16b }, v2.16b
 ; CHECK-SD-NEXT:    ret
 ;
 ; CHECK-GI-LABEL: v16i8_2:
 ; CHECK-GI:       // %bb.0: // %entry
-; CHECK-GI-NEXT:    adrp x8, .LCPI8_0
+; CHECK-GI-NEXT:    adrp x8, .LCPI9_0
 ; CHECK-GI-NEXT:    // kill: def $d0 killed $d0 killed $q0_q1 def $q0_q1
-; CHECK-GI-NEXT:    ldr q2, [x8, :lo12:.LCPI8_0]
+; CHECK-GI-NEXT:    ldr q2, [x8, :lo12:.LCPI9_0]
 ; CHECK-GI-NEXT:    // kill: def $d1 killed $d1 killed $q0_q1 def $q0_q1
 ; CHECK-GI-NEXT:    tbl v0.16b, { v0.16b, v1.16b }, v2.16b
 ; CHECK-GI-NEXT:    ret
@@ -166,19 +158,11 @@ entry:
 }
 
 define <4 x float> @v4f32(<4 x float> %a) {
-; CHECK-SD-LABEL: v4f32:
-; CHECK-SD:       // %bb.0: // %entry
-; CHECK-SD-NEXT:    rev64 v0.4s, v0.4s
-; CHECK-SD-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
-; CHECK-SD-NEXT:    ret
-;
-; CHECK-GI-LABEL: v4f32:
-; CHECK-GI:       // %bb.0: // %entry
-; CHECK-GI-NEXT:    adrp x8, .LCPI11_0
-; CHECK-GI-NEXT:    // kill: def $q0 killed $q0 def $q0_q1
-; CHECK-GI-NEXT:    ldr q2, [x8, :lo12:.LCPI11_0]
-; CHECK-GI-NEXT:    tbl v0.16b, { v0.16b, v1.16b }, v2.16b
-; CHECK-GI-NEXT:    ret
+; CHECK-LABEL: v4f32:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    rev64 v0.4s, v0.4s
+; CHECK-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    ret
 entry:
   %V128 = shufflevector <4 x float> %a, <4 x float> undef, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
   ret <4 x float> %V128
@@ -195,19 +179,11 @@ entry:
 }
 
 define <8 x half> @v8f16(<8 x half> %a) {
-; CHECK-SD-LABEL: v8f16:
-; CHECK-SD:       // %bb.0: // %entry
-; CHECK-SD-NEXT:    rev64 v0.8h, v0.8h
-; CHECK-SD-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
-; CHECK-SD-NEXT:    ret
-;
-; CHECK-GI-LABEL: v8f16:
-; CHECK-GI:       // %bb.0: // %entry
-; CHECK-GI-NEXT:    adrp x8, .LCPI13_0
-; CHECK-GI-NEXT:    // kill: def $q0 killed $q0 def $q0_q1
-; CHECK-GI-NEXT:    ldr q2, [x8, :lo12:.LCPI13_0]
-; CHECK-GI-NEXT:    tbl v0.16b, { v0.16b, v1.16b }, v2.16b
-; CHECK-GI-NEXT:    ret
+; CHECK-LABEL: v8f16:
+; CHECK:       // %bb.0: // %entry
+; CHECK-NEXT:    rev64 v0.8h, v0.8h
+; CHECK-NEXT:    ext v0.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    ret
 entry:
   %V128 = shufflevector <8 x half> %a, <8 x half> undef, <8 x i32> <i32 7, i32 6, i32 5, i32 4, i32 3, i32 2, i32 1, i32 0>
   ret <8 x half> %V128
