[llvm] r366132 - AMDGPU: Use standalone MUBUF load patterns

[Matt Arsenault via llvm-commits]

Author: arsenm
Date: Mon Jul 15 14:41:44 2019
New Revision: 366132

URL: http://llvm.org/viewvc/llvm-project?rev=366132&view=rev
Log:
AMDGPU: Use standalone MUBUF load patterns

We already do this for the flat and DS instructions, although it is
certainly uglier and more verbose.

This will allow using separate pattern definitions for extload and
zextload. Currently we get away with using a single PatFrag with
custom predicate code to check if the extension type is a zextload or
anyextload. The generic mechanism the global isel emitter understands
treats these as mutually exclusive. I was considering making the
pattern emitter accept zextload or sextload extensions for anyextload
patterns, but in global isel, the different extending loads have
distinct opcodes, and there is currently no mechanism for an opcode
matcher to try multiple (and there probably is very little need for
one beyond this case).

Modified:
    llvm/trunk/lib/Target/AMDGPU/BUFInstructions.td

Modified: llvm/trunk/lib/Target/AMDGPU/BUFInstructions.td
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/AMDGPU/BUFInstructions.td?rev=366132&r1=366131&r2=366132&view=diff
==============================================================================
--- llvm/trunk/lib/Target/AMDGPU/BUFInstructions.td (original)
+++ llvm/trunk/lib/Target/AMDGPU/BUFInstructions.td Mon Jul 15 14:41:44 2019
@@ -470,6 +470,24 @@ class MUBUF_Load_Pseudo <string opName,
   let dwords = getMUBUFDwords<vdataClass>.ret;
 }
 
+class MUBUF_Offset_Load_Pat <Instruction inst, ValueType load_vt = i32, SDPatternOperator ld = null_frag> : Pat <
+  (load_vt (ld (MUBUFOffset v4i32:$srsrc, i32:$soffset, i16:$offset, i1:$glc, i1:$slc, i1:$tfe, i1:$dlc))),
+  (load_vt (inst v4i32:$srsrc, i32:$soffset, i16:$offset, i1:$glc, i1:$slc, i1:$tfe, i1:$dlc))
+>;
+
+class MUBUF_Addr64_Load_Pat <Instruction inst,
+                            ValueType load_vt = i32,
+                            SDPatternOperator ld = null_frag> : Pat <
+  (load_vt (ld (MUBUFAddr64 v4i32:$srsrc, i64:$vaddr, i32:$soffset, i16:$offset, i1:$glc, i1:$slc, i1:$tfe, i1:$dlc))),
+  (load_vt (inst i64:$vaddr, v4i32:$srsrc, i32:$soffset, i16:$offset, i1:$glc, i1:$slc, i1:$tfe, i1:$dlc))
+>;
+
+multiclass MUBUF_Pseudo_Load_Pats<string BaseInst, ValueType load_vt = i32, SDPatternOperator ld = null_frag> {
+  def : MUBUF_Offset_Load_Pat<!cast<Instruction>(BaseInst#"_OFFSET"), load_vt, ld>;
+  def : MUBUF_Addr64_Load_Pat<!cast<Instruction>(BaseInst#"_ADDR64"), load_vt, ld>;
+}
+
+
 // FIXME: tfe can't be an operand because it requires a separate
 // opcode because it needs an N+1 register class dest register.
 multiclass MUBUF_Pseudo_Loads<string opName, RegisterClass vdataClass,
@@ -478,20 +496,10 @@ multiclass MUBUF_Pseudo_Loads<string opN
                               bit TiedDest = 0,
                               bit isLds = 0> {
 
-  def _OFFSET : MUBUF_Load_Pseudo <opName, BUFAddrKind.Offset, vdataClass,
-    TiedDest, isLds,
-    !if(isLds,
-        [],
-        [(set load_vt:$vdata,
-         (ld (MUBUFOffset v4i32:$srsrc, i32:$soffset, i16:$offset, i1:$glc, i1:$slc, i1:$tfe, i1:$dlc)))])>,
+  def _OFFSET : MUBUF_Load_Pseudo <opName, BUFAddrKind.Offset, vdataClass, TiedDest, isLds>,
     MUBUFAddr64Table<0, NAME # !if(isLds, "_LDS", "")>;
 
-  def _ADDR64 : MUBUF_Load_Pseudo <opName, BUFAddrKind.Addr64, vdataClass,
-    TiedDest, isLds,
-    !if(isLds,
-        [],
-        [(set load_vt:$vdata,
-         (ld (MUBUFAddr64 v4i32:$srsrc, i64:$vaddr, i32:$soffset, i16:$offset, i1:$glc, i1:$slc, i1:$tfe, i1:$dlc)))])>,
+  def _ADDR64 : MUBUF_Load_Pseudo <opName, BUFAddrKind.Addr64, vdataClass, TiedDest, isLds>,
     MUBUFAddr64Table<1, NAME # !if(isLds, "_LDS", "")>;
 
   def _OFFEN  : MUBUF_Load_Pseudo <opName, BUFAddrKind.OffEn, vdataClass, TiedDest, isLds>;
@@ -819,30 +827,39 @@ let SubtargetPredicate = HasPackedD16VMe
 } // End HasPackedD16VMem.
 
 defm BUFFER_LOAD_UBYTE : MUBUF_Pseudo_Loads_Lds <
-  "buffer_load_ubyte", VGPR_32, i32, az_extloadi8_global
+  "buffer_load_ubyte", VGPR_32, i32
 >;
 defm BUFFER_LOAD_SBYTE : MUBUF_Pseudo_Loads_Lds <
-  "buffer_load_sbyte", VGPR_32, i32, sextloadi8_global
+  "buffer_load_sbyte", VGPR_32, i32
 >;
 defm BUFFER_LOAD_USHORT : MUBUF_Pseudo_Loads_Lds <
-  "buffer_load_ushort", VGPR_32, i32, az_extloadi16_global
+  "buffer_load_ushort", VGPR_32, i32
 >;
 defm BUFFER_LOAD_SSHORT : MUBUF_Pseudo_Loads_Lds <
-  "buffer_load_sshort", VGPR_32, i32, sextloadi16_global
+  "buffer_load_sshort", VGPR_32, i32
 >;
 defm BUFFER_LOAD_DWORD : MUBUF_Pseudo_Loads_Lds <
-  "buffer_load_dword", VGPR_32, i32, load_global
+  "buffer_load_dword", VGPR_32, i32
 >;
 defm BUFFER_LOAD_DWORDX2 : MUBUF_Pseudo_Loads <
-  "buffer_load_dwordx2", VReg_64, v2i32, load_global
+  "buffer_load_dwordx2", VReg_64, v2i32
 >;
 defm BUFFER_LOAD_DWORDX3 : MUBUF_Pseudo_Loads <
-  "buffer_load_dwordx3", VReg_96, v3i32, load_global
+  "buffer_load_dwordx3", VReg_96, v3i32
 >;
 defm BUFFER_LOAD_DWORDX4 : MUBUF_Pseudo_Loads <
-  "buffer_load_dwordx4", VReg_128, v4i32, load_global
+  "buffer_load_dwordx4", VReg_128, v4i32
 >;
 
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_UBYTE", i32, az_extloadi8_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_SBYTE", i32, sextloadi8_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_USHORT", i32, az_extloadi16_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_SSHORT", i32, sextloadi16_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_DWORD", i32, load_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_DWORDX2", v2i32, load_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_DWORDX3", v3i32, load_global>;
+defm : MUBUF_Pseudo_Load_Pats<"BUFFER_LOAD_DWORDX4", v4i32, load_global>;
+
 // This is not described in AMD documentation,
 // but 'lds' versions of these opcodes are available
 // in at least GFX8+ chips. See Bug 37653.