[AMDGPU] Remove GDS and GWS for GFX12 (#76148)

9 files changed, 45 insertions, 11 deletions

diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index 060fb66d38f7..d2a325d5ad89 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -1100,8 +1100,8 @@ def FeatureGFX12 : GCNSubtargetFeatureGeneration<"GFX12",
    FeatureVOP3Literal, FeatureDPP8,
    FeatureNoDataDepHazard, FeaturePkFmacF16Inst,
    FeatureA16, FeatureFastDenormalF32, FeatureG16,
-   FeatureUnalignedBufferAccess, FeatureUnalignedDSAccess, FeatureGDS,
-   FeatureGWS, FeatureTrue16BitInsts
+   FeatureUnalignedBufferAccess, FeatureUnalignedDSAccess,
+   FeatureTrue16BitInsts
   ]
 >;
 
diff --git a/llvm/lib/Target/AMDGPU/DSInstructions.td b/llvm/lib/Target/AMDGPU/DSInstructions.td
index 3a895923fa4b..bc9049b4ef33 100644
--- a/llvm/lib/Target/AMDGPU/DSInstructions.td
+++ b/llvm/lib/Target/AMDGPU/DSInstructions.td
@@ -1147,7 +1147,8 @@ def : GCNPat <
 >;
 } // End SubtargetPredicate = HasAtomicDsPkAdd16Insts
 
-def : Pat <
+let OtherPredicates = [HasGDS] in
+def : GCNPat <
   (SIds_ordered_count i32:$value, i16:$offset),
   (DS_ORDERED_COUNT $value, (as_i16imm $offset))
 >;
@@ -1189,7 +1190,8 @@ def : GCNPat <
 //===----------------------------------------------------------------------===//
 
 class Base_DS_Real_gfx6_gfx7_gfx10_gfx11_gfx12<bits<8> op, DS_Pseudo ps, int ef,
-                                               string opName = ps.Mnemonic>
+                                               string opName = ps.Mnemonic,
+                                               bit hasGFX12Enc = 0>
     : DS_Real<ps, opName>, SIMCInstr <ps.Mnemonic, ef> {
 
   let Inst{7-0}   = !if(ps.has_offset0, offset0, 0);
@@ -1201,6 +1203,8 @@ class Base_DS_Real_gfx6_gfx7_gfx10_gfx11_gfx12<bits<8> op, DS_Pseudo ps, int ef,
   let Inst{47-40} = !if(ps.has_data0, data0{7-0}, 0);
   let Inst{55-48} = !if(ps.has_data1, data1{7-0}, 0);
   let Inst{63-56} = !if(ps.has_vdst, vdst{7-0}, 0);
+
+  let gds = !if(hasGFX12Enc, 0, ?);
 }
 
 //===----------------------------------------------------------------------===//
@@ -1212,7 +1216,7 @@ let AssemblerPredicate = isGFX12Plus, DecoderNamespace = "GFX12" in {
     defvar ps = !cast<DS_Pseudo>(NAME);
     def _gfx12 :
       Base_DS_Real_gfx6_gfx7_gfx10_gfx11_gfx12<op, ps, SIEncodingFamily.GFX12,
-                                               ps.Mnemonic>;
+                                               ps.Mnemonic, 1>;
   }
 
   multiclass DS_Real_Renamed_gfx12<bits<8> op, DS_Pseudo backing_pseudo,
@@ -1220,7 +1224,7 @@ let AssemblerPredicate = isGFX12Plus, DecoderNamespace = "GFX12" in {
     def _gfx12 :
       Base_DS_Real_gfx6_gfx7_gfx10_gfx11_gfx12<op, backing_pseudo,
                                                SIEncodingFamily.GFX12,
-                                               real_name>,
+                                               real_name, 1>,
       MnemonicAlias<backing_pseudo.Mnemonic, real_name>,
       Requires<[isGFX12Plus]>;
   }
diff --git a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
index ed2e7e4f189e..7939d0036568 100644
--- a/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
+++ b/llvm/lib/Target/AMDGPU/Disassembler/AMDGPUDisassembler.cpp
@@ -702,6 +702,11 @@ DecodeStatus AMDGPUDisassembler::getInstruction(MCInst &MI, uint64_t &Size,
                          AMDGPU::OpName::src2_modifiers);
   }
 
+  if (Res && (MCII->get(MI.getOpcode()).TSFlags & SIInstrFlags::DS) &&
+      !AMDGPU::hasGDS(STI)) {
+    insertNamedMCOperand(MI, MCOperand::createImm(0), AMDGPU::OpName::gds);
+  }
+
   if (Res && (MCII->get(MI.getOpcode()).TSFlags &
           (SIInstrFlags::MUBUF | SIInstrFlags::FLAT | SIInstrFlags::SMRD))) {
     int CPolPos = AMDGPU::getNamedOperandIdx(MI.getOpcode(),
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index e599f23101c8..29ac08b6895e 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -4983,6 +4983,14 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
     }
   }
 
+  if (isDS(MI) && !ST.hasGDS()) {
+    const MachineOperand *GDSOp = getNamedOperand(MI, AMDGPU::OpName::gds);
+    if (GDSOp && GDSOp->getImm() != 0) {
+      ErrInfo = "GDS is not supported on this subtarget";
+      return false;
+    }
+  }
+
   if (isImage(MI)) {
     const MachineOperand *DimOp = getNamedOperand(MI, AMDGPU::OpName::dim);
     if (DimOp) {
diff --git a/llvm/test/CodeGen/AMDGPU/gds-unsupported.ll b/llvm/test/CodeGen/AMDGPU/gds-unsupported.ll
index 174de0d58295..d35b9f79e23a 100644
--- a/llvm/test/CodeGen/AMDGPU/gds-unsupported.ll
+++ b/llvm/test/CodeGen/AMDGPU/gds-unsupported.ll
@@ -1,4 +1,5 @@
 ; RUN: not --crash llc -march=amdgcn -mcpu=gfx90a < %s 2>&1 | FileCheck %s
+; RUN: not --crash llc -march=amdgcn -mcpu=gfx1200 < %s 2>&1 | FileCheck %s
 
 ; GDS is not supported on GFX90A+
 ; CHECK: LLVM ERROR: Cannot select: {{.*}} AtomicLoadAdd
diff --git a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add.gfx11.ll b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add.gfx11.ll
index a9c2c2790389..3e977c054ec2 100644
--- a/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add.gfx11.ll
+++ b/llvm/test/CodeGen/AMDGPU/llvm.amdgcn.ds.ordered.add.gfx11.ll
@@ -1,5 +1,8 @@
 ; RUN: llc -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,FUNC %s
 ; RUN: llc -global-isel -march=amdgcn -mcpu=gfx1100 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s | FileCheck -check-prefixes=GCN,FUNC %s
+; RUN: not --crash llc -march=amdgcn -mcpu=gfx1200 -amdgpu-enable-vopd=0 -verify-machineinstrs < %s 2>&1 | FileCheck -check-prefix=GFX12-ERR %s
+
+; GFX12-ERR: LLVM ERROR: Cannot select: {{.*}} = DS_ORDERED_COUNT
 
 ; FUNC-LABEL: {{^}}ds_ordered_add:
 ; GCN-DAG: v_mov_b32_e32 v[[INCR:[0-9]+]], 31
diff --git a/llvm/test/CodeGen/AMDGPU/verify-gfx12-gds.mir b/llvm/test/CodeGen/AMDGPU/verify-gfx12-gds.mir
new file mode 100644
index 000000000000..a2182aa8d6ef
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/verify-gfx12-gds.mir
@@ -0,0 +1,10 @@
+# RUN: not --crash llc -march=amdgcn -mcpu=gfx1200 -run-pass=none -o /dev/null %s 2>&1 | FileCheck -check-prefix=GFX12 %s
+
+---
+name: gds
+body: |
+  bb.0:
+    ; GFX12: *** Bad machine code: GDS is not supported on this subtarget ***
+    ; GFX12: - instruction: DS_ADD_U32 %0:vgpr_32, %1:vgpr_32, 0, 1, implicit $m0, implicit $exec :: (load store acq_rel (s32), addrspace 2)
+    DS_ADD_U32 %0:vgpr_32, %2:vgpr_32, 0, 1, implicit $m0, implicit $exec :: (load store acq_rel (s32), addrspace 2)
+...
diff --git a/llvm/test/MC/Disassembler/AMDGPU/decode-err.txt b/llvm/test/MC/Disassembler/AMDGPU/decode-err.txt
index 24716b1226e4..e1bb7ad51171 100644
--- a/llvm/test/MC/Disassembler/AMDGPU/decode-err.txt
+++ b/llvm/test/MC/Disassembler/AMDGPU/decode-err.txt
@@ -34,6 +34,14 @@
 # W64: v_wmma_f32_16x16x16_f16 v[16:19], v[0:7], v[8:15], s[0:3]/*Invalid register, operand has 'VReg_128' register class*/ ; encoding: [0x10,0x40,0x40,0xcc,0x00,0x11,0x02,0x18]
 0x10,0x40,0x40,0xcc,0x00,0x11,0x02,0x18 # src2 sgpr0
 
+# this is ds_add_f32 with gds bit which is not valid on gfx12+
+# GFX12: [[@LINE+1]]:1: warning: invalid instruction encoding
+0x00,0x00,0x56,0xd8,0x00,0x01,0x00,0x00
+
 # this is image_msaa_load where samp field for gfx12 VSAMPLE is not all zeros
 # GFX12: [[@LINE+1]]:1: warning: invalid instruction encoding
 0x06,0x00,0x46,0xe4,0x01,0x10,0x80,0x00,0x05,0x06,0x07,0x00
+
+# This is ds_read_b32 with gds bit which is not valid on gfx90a.
+# GFX90A: [[@LINE+1]]:1: warning: invalid instruction encoding
+0x00,0x00,0x6d,0xd8,0x01,0x00,0x00,0x00
diff --git a/llvm/test/MC/Disassembler/AMDGPU/gfx90a_features.txt b/llvm/test/MC/Disassembler/AMDGPU/gfx90a_features.txt
index 8746ee79c8f5..b348e8b5ef01 100644
--- a/llvm/test/MC/Disassembler/AMDGPU/gfx90a_features.txt
+++ b/llvm/test/MC/Disassembler/AMDGPU/gfx90a_features.txt
@@ -773,8 +773,3 @@
 
 # GFX90A: flat_atomic_min_f64 v[0:1], v[0:1], v[2:3] glc ; encoding: [0x00,0x00,0x41,0xdd,0x00,0x02,0x00,0x00]
 0x00,0x00,0x41,0xdd,0x00,0x02,0x00,0x00
-
-# Disassembler still decodes the gds modifier even though the assembler does
-# not accept it.
-# GFX90A: ds_read_b32 v0, v1 gds ; encoding: [0x00,0x00,0x6d,0xd8,0x01,0x00,0x00,0x00]
-0x00,0x00,0x6d,0xd8,0x01,0x00,0x00,0x00