[X86] Tag PSLLDQ/PSRLDQ as WriteShuffle scheduler classes instead of shifts.

Although they are encoded similar to bit shifts, the byte shifts behave like shuffles from a scheduling point of view.

llvm-svn: 331253

6 files changed, 25 insertions, 39 deletions

diff --git a/llvm/lib/Target/X86/X86InstrAVX512.td b/llvm/lib/Target/X86/X86InstrAVX512.td
index 73dc780..e13258d 100644
--- a/llvm/lib/Target/X86/X86InstrAVX512.td
+++ b/llvm/lib/Target/X86/X86InstrAVX512.td
@@ -10044,6 +10044,7 @@ defm VSHUFPD: avx512_shufp<"vshufpd", avx512vl_i64_info, avx512vl_f64_info>, PD,
 // AVX-512 - Byte shift Left/Right
 //===----------------------------------------------------------------------===//
 
+// FIXME: The SSE/AVX names are PSLLDQri etc. - should we add the i here as well?
 multiclass avx512_shift_packed<bits<8> opc, SDNode OpNode, Format MRMr,
                                Format MRMm, string OpcodeStr,
                                X86FoldableSchedWrite sched, X86VectorVTInfo _>{
@@ -10063,24 +10064,23 @@ multiclass avx512_shift_packed<bits<8> opc, SDNode OpNode, Format MRMr,
 
 multiclass avx512_shift_packed_all<bits<8> opc, SDNode OpNode, Format MRMr,
                                    Format MRMm, string OpcodeStr,
-                                   X86FoldableSchedWrite sched, Predicate prd>{
+                                   X86SchedWriteWidths sched, Predicate prd>{
   let Predicates = [prd] in
-    defm Z : avx512_shift_packed<opc, OpNode, MRMr, MRMm,
-                                 OpcodeStr, sched, v64i8_info>, EVEX_V512;
+    defm Z : avx512_shift_packed<opc, OpNode, MRMr, MRMm, OpcodeStr,
+                                 sched.ZMM, v64i8_info>, EVEX_V512;
   let Predicates = [prd, HasVLX] in {
-    defm Z256 : avx512_shift_packed<opc, OpNode, MRMr, MRMm,
-                                    OpcodeStr, sched, v32i8x_info>, EVEX_V256;
-    defm Z128 : avx512_shift_packed<opc, OpNode, MRMr, MRMm,
-                                    OpcodeStr, sched, v16i8x_info>, EVEX_V128;
+    defm Z256 : avx512_shift_packed<opc, OpNode, MRMr, MRMm, OpcodeStr,
+                                    sched.YMM, v32i8x_info>, EVEX_V256;
+    defm Z128 : avx512_shift_packed<opc, OpNode, MRMr, MRMm, OpcodeStr,
+                                    sched.XMM, v16i8x_info>, EVEX_V128;
   }
 }
 defm VPSLLDQ : avx512_shift_packed_all<0x73, X86vshldq, MRM7r, MRM7m, "vpslldq",
-                                       WriteVecShift, HasBWI>, AVX512PDIi8Base,
-                                       EVEX_4V, VEX_WIG;
+                                       SchedWriteShuffle, HasBWI>,
+                                       AVX512PDIi8Base, EVEX_4V, VEX_WIG;
 defm VPSRLDQ : avx512_shift_packed_all<0x73, X86vshrdq, MRM3r, MRM3m, "vpsrldq",
-                                       WriteVecShift, HasBWI>, AVX512PDIi8Base,
-                                       EVEX_4V, VEX_WIG;
-
+                                       SchedWriteShuffle, HasBWI>,
+                                       AVX512PDIi8Base, EVEX_4V, VEX_WIG;
 
 multiclass avx512_psadbw_packed<bits<8> opc, SDNode OpNode,
                                 string OpcodeStr, X86FoldableSchedWrite sched,
diff --git a/llvm/lib/Target/X86/X86InstrSSE.td b/llvm/lib/Target/X86/X86InstrSSE.td
index 8da995c..e99f806 100644
--- a/llvm/lib/Target/X86/X86InstrSSE.td
+++ b/llvm/lib/Target/X86/X86InstrSSE.td
@@ -3446,16 +3446,17 @@ multiclass PDI_binop_ri<bits<8> opc, Format ImmForm, string OpcodeStr,
 }
 
 multiclass PDI_binop_ri_all<bits<8> opc, Format ImmForm, string OpcodeStr,
-                            SDNode OpNode, X86FoldableSchedWrite sched> {
+                            SDNode OpNode, X86SchedWriteWidths sched> {
 let Predicates = [HasAVX, NoVLX_Or_NoBWI] in
   defm V#NAME : PDI_binop_ri<opc, ImmForm, !strconcat("v", OpcodeStr), OpNode,
-                             VR128, v16i8, sched, 0>, VEX_4V, VEX_WIG;
+                             VR128, v16i8, sched.XMM, 0>, VEX_4V, VEX_WIG;
 let Predicates = [HasAVX2, NoVLX_Or_NoBWI] in
   defm V#NAME#Y : PDI_binop_ri<opc, ImmForm, !strconcat("v", OpcodeStr), OpNode,
-                               VR256, v32i8, sched, 0>,
+                               VR256, v32i8, sched.YMM, 0>,
                                VEX_4V, VEX_L, VEX_WIG;
 let Constraints = "$src1 = $dst" in
-  defm NAME : PDI_binop_ri<opc, ImmForm, OpcodeStr, OpNode, VR128, v16i8, sched>;
+  defm NAME : PDI_binop_ri<opc, ImmForm, OpcodeStr, OpNode, VR128, v16i8,
+                           sched.XMM>;
 }
 
 let ExeDomain = SSEPackedInt in {
@@ -3481,9 +3482,10 @@ let ExeDomain = SSEPackedInt in {
   defm PSRAD : PDI_binop_rmi_all<0xE2, 0x72, MRM4r, "psrad", X86vsra, X86vsrai,
                                  v4i32, v8i32, v4i32, WriteVecShift, NoVLX>;
 
-  defm PSLLDQ : PDI_binop_ri_all<0x73, MRM7r, "pslldq", X86vshldq, WriteVecShift>;
-  defm PSRLDQ : PDI_binop_ri_all<0x73, MRM3r, "psrldq", X86vshrdq, WriteVecShift>;
-  // PSRADQri doesn't exist in SSE[1-3].
+  defm PSLLDQ : PDI_binop_ri_all<0x73, MRM7r, "pslldq", X86vshldq,
+                                 SchedWriteShuffle>;
+  defm PSRLDQ : PDI_binop_ri_all<0x73, MRM3r, "psrldq", X86vshrdq,
+                                 SchedWriteShuffle>;
 } // ExeDomain = SSEPackedInt
 
 //===---------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/X86/X86SchedBroadwell.td b/llvm/lib/Target/X86/X86SchedBroadwell.td
index b9664ee..25a49cb 100755
--- a/llvm/lib/Target/X86/X86SchedBroadwell.td
+++ b/llvm/lib/Target/X86/X86SchedBroadwell.td
@@ -386,9 +386,7 @@ def: InstRW<[BWWriteResGroup3], (instregex "MMX_MOVD64rr",
                                            "MMX_MOVD64to64rr",
                                            "MMX_MOVQ2DQrr",
                                            "(V?)MOV64toPQIrr",
-                                           "(V?)MOVDI2PDIrr",
-                                           "(V?)PSLLDQ(Y?)ri",
-                                           "(V?)PSRLDQ(Y?)ri")>;
+                                           "(V?)MOVDI2PDIrr")>;
 
 def BWWriteResGroup4 : SchedWriteRes<[BWPort6]> {
   let Latency = 1;
diff --git a/llvm/lib/Target/X86/X86SchedHaswell.td b/llvm/lib/Target/X86/X86SchedHaswell.td
index 8682f2e0..d9b7be0 100644
--- a/llvm/lib/Target/X86/X86SchedHaswell.td
+++ b/llvm/lib/Target/X86/X86SchedHaswell.td
@@ -723,9 +723,7 @@ def: InstRW<[HWWriteResGroup4], (instregex "MMX_MOVD64rr",
                                            "MMX_MOVD64to64rr",
                                            "MMX_MOVQ2DQrr",
                                            "(V?)MOV64toPQIrr",
-                                           "(V?)MOVDI2PDIrr",
-                                           "(V?)PSLLDQ(Y?)ri",
-                                           "(V?)PSRLDQ(Y?)ri")>;
+                                           "(V?)MOVDI2PDIrr")>;
 
 def HWWriteResGroup5 : SchedWriteRes<[HWPort6]> {
   let Latency = 1;
diff --git a/llvm/lib/Target/X86/X86SchedSkylakeClient.td b/llvm/lib/Target/X86/X86SchedSkylakeClient.td
index b1a4714..5e02e43 100644
--- a/llvm/lib/Target/X86/X86SchedSkylakeClient.td
+++ b/llvm/lib/Target/X86/X86SchedSkylakeClient.td
@@ -397,9 +397,7 @@ def: InstRW<[SKLWriteResGroup3], (instregex "COMP_FST0r",
                                             "UCOM_FPr",
                                             "UCOM_Fr",
                                             "(V?)MOV64toPQIrr",
-                                            "(V?)MOVDI2PDIrr",
-                                            "(V?)PSLLDQ(Y?)ri",
-                                            "(V?)PSRLDQ(Y?)ri")>;
+                                            "(V?)MOVDI2PDIrr")>;
 
 def SKLWriteResGroup4 : SchedWriteRes<[SKLPort6]> {
   let Latency = 1;
diff --git a/llvm/lib/Target/X86/X86SchedSkylakeServer.td b/llvm/lib/Target/X86/X86SchedSkylakeServer.td
index 95b6776..a9693bf 100755
--- a/llvm/lib/Target/X86/X86SchedSkylakeServer.td
+++ b/llvm/lib/Target/X86/X86SchedSkylakeServer.td
@@ -445,17 +445,7 @@ def: InstRW<[SKXWriteResGroup3], (instregex "COMP_FST0r",
                                             "VMOV64toPQIZrr",
                                             "VMOV64toPQIrr",
                                             "VMOVDI2PDIZrr",
-                                            "VMOVDI2PDIrr",
-                                            "VPSLLDQYri",
-                                            "VPSLLDQZ128rr",
-                                            "VPSLLDQZ256rr",
-                                            "VPSLLDQZrr",
-                                            "(V?)PSLLDQri",
-                                            "VPSRLDQYri",
-                                            "VPSRLDQZ128rr",
-                                            "VPSRLDQZ256rr",
-                                            "VPSRLDQZrr",
-                                            "(V?)PSRLDQri")>;
+                                            "VMOVDI2PDIrr")>;
 
 def SKXWriteResGroup4 : SchedWriteRes<[SKXPort6]> {
   let Latency = 1;