[AMDGPU] Do not clause NSA instructions

[critson]

Summary

To ensure correct behaviour NSA instructions should not be claused.

[critson]

• Rebase on pre-commit tests.

[foad]

As I understand it, this is only needed for GFX10.1, not for GFX10.3. So I think it probably deserves its own target feature which would be included in GFX10_1_Bugs.

llvm/lib/Target/AMDGPU/SIInsertHardClauses.cpp

80

Maybe sink this down to just before we return HARDCLAUSE_VMEM?

[critson]

In D102211#2750312, @foad wrote:

As I understand it, this is only needed for GFX10.1, not for GFX10.3. So I think it probably deserves its own target feature which would be included in GFX10_1_Bugs.

You're right, this is only needed for GFX10.1.

[critson]

• Rebase
• Contain in GFX10.1 specific bug feature

[foad]

LGTM, thanks!

llvm/lib/Target/AMDGPU/SIInsertHardClauses.cpp

public:
   SIInsertHardClauses() : MachineFunctionPass(ID) {}
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.setPreservesCFG();
     MachineFunctionPass::getAnalysisUsage(AU);
   }
 
   HardClauseType getHardClauseType(const MachineInstr &MI) {
+    // Do not clause NSA instructions on GFX10.
+    if (AMDGPU::isGFX10(*ST) && ST->hasFeature(AMDGPU::FeatureNSAEncoding)) {
+      const AMDGPU::MIMGInfo *Info = AMDGPU::getMIMGInfo(MI.getOpcode());
+      if (Info && Info->MIMGEncoding == AMDGPU::MIMGEncGfx10NSA)
+        return HARDCLAUSE_ILLEGAL;
+    }
+
     // On current architectures we only get a benefit from clausing loads.
     if (MI.mayLoad()) {
       if (SIInstrInfo::isVMEM(MI) || SIInstrInfo::isSegmentSpecificFLAT(MI))
         return HARDCLAUSE_VMEM;
       if (SIInstrInfo::isFLAT(MI))
         return HARDCLAUSE_FLAT;
       // TODO: LDS
       if (SIInstrInfo::isSMRD(MI))

llvm/test/CodeGen/AMDGPU/cluster_stores.ll

 ; GFX10-NEXT:    v_cvt_f32_i32_e32 v2, v0
 ; GFX10-NEXT:    v_cvt_f32_i32_e32 v3, v1
 ; GFX10-NEXT:    v_mov_b32_e32 v13, 0
 ; GFX10-NEXT:    v_mov_b32_e32 v10, 1.0
 ; GFX10-NEXT:    v_add_f32_e32 v11, 1.0, v2
 ; GFX10-NEXT:    v_add_f32_e32 v12, 1.0, v3
 ; GFX10-NEXT:    v_add_f32_e32 v14, 2.0, v2
 ; GFX10-NEXT:    v_add_f32_e32 v15, 2.0, v3
-; GFX10-NEXT:    s_clause 0x1
 ; GFX10-NEXT:    image_sample_d v[2:5], [v11, v12, v13, v13, v13, v13], s[0:7], s[8:11] dmask:0xf dim:SQ_RSRC_IMG_2D
 ; GFX10-NEXT:    image_sample_d v[6:9], [v14, v15, v10, v10, v10, v10], s[0:7], s[8:11] dmask:0xf dim:SQ_RSRC_IMG_2D
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    v_add_f32_e32 v5, v5, v9
 ; GFX10-NEXT:    v_add_f32_e32 v4, v4, v8
 ; GFX10-NEXT:    v_add_f32_e32 v3, v3, v7
 ; GFX10-NEXT:    v_add_f32_e32 v2, v2, v6
 ; GFX10-NEXT:    image_store v[2:5], v[0:1], s[12:19] dmask:0xf dim:SQ_RSRC_IMG_2D unorm

llvm/test/CodeGen/AMDGPU/hard-clauses.mir

bb.0:
     $vgpr74 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, 0, 296, 0, 0, 0, implicit $exec
     $vgpr75 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, 0, 300, 0, 0, 0, implicit $exec
     $vgpr76 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, 0, 304, 0, 0, 0, implicit $exec
     $vgpr77 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, 0, 308, 0, 0, 0, implicit $exec
     $vgpr78 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, 0, 312, 0, 0, 0, implicit $exec
     $vgpr79 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, 0, 316, 0, 0, 0, implicit $exec
     $vgpr80 = BUFFER_LOAD_DWORD_OFFEN $vgpr0, $sgpr0_sgpr1_sgpr2_sgpr3, 0, 320, 0, 0, 0, implicit $exec
 ...
+
+---
+name: mimg_nsa
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
+    ; CHECK-LABEL: name: mimg_nsa
+    ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
+    ; CHECK: $vgpr10_vgpr11_vgpr12_vgpr13 = IMAGE_SAMPLE_D_V4_V9_nsa_gfx10 $vgpr3, $vgpr8, $vgpr7, $vgpr5, $vgpr4, $vgpr6, $vgpr0, $vgpr2, $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 15, 2, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (load 16)
+    ; CHECK: $vgpr20_vgpr21_vgpr22_vgpr23 = IMAGE_SAMPLE_D_V4_V9_nsa_gfx10 $vgpr3, $vgpr8, $vgpr7, $vgpr5, $vgpr4, $vgpr6, $vgpr0, $vgpr2, $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 15, 2, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (load 16)
+    $vgpr10_vgpr11_vgpr12_vgpr13 = IMAGE_SAMPLE_D_V4_V9_nsa_gfx10 $vgpr3, $vgpr8, $vgpr7, $vgpr5, $vgpr4, $vgpr6, $vgpr0, $vgpr2, $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 15, 2, 0, 0, 0, 0, 0, 0, 0, implicit $exec  :: (load 16)
+    $vgpr20_vgpr21_vgpr22_vgpr23 = IMAGE_SAMPLE_D_V4_V9_nsa_gfx10 $vgpr3, $vgpr8, $vgpr7, $vgpr5, $vgpr4, $vgpr6, $vgpr0, $vgpr2, $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 15, 2, 0, 0, 0, 0, 0, 0, 0, implicit $exec  :: (load 16)
+...
+
+---
+name: mimg_nsa_mixed
+tracksRegLiveness: true
+body: |
+  bb.0:
+    liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
+    ; CHECK-LABEL: name: mimg_nsa_mixed
+    ; CHECK: liveins: $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8
+    ; CHECK: $vgpr10_vgpr11_vgpr12_vgpr13 = IMAGE_SAMPLE_D_V4_V9_nsa_gfx10 $vgpr3, $vgpr8, $vgpr7, $vgpr5, $vgpr4, $vgpr6, $vgpr0, $vgpr2, $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 15, 2, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (load 16)
+    ; CHECK: $vgpr14 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 $vgpr5_vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 1, 1, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
+    ; CHECK: $vgpr20_vgpr21_vgpr22_vgpr23 = IMAGE_SAMPLE_D_V4_V9_nsa_gfx10 $vgpr3, $vgpr8, $vgpr7, $vgpr5, $vgpr4, $vgpr6, $vgpr0, $vgpr2, $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 15, 2, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (load 16)
+    $vgpr10_vgpr11_vgpr12_vgpr13 = IMAGE_SAMPLE_D_V4_V9_nsa_gfx10 $vgpr3, $vgpr8, $vgpr7, $vgpr5, $vgpr4, $vgpr6, $vgpr0, $vgpr2, $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 15, 2, 0, 0, 0, 0, 0, 0, 0, implicit $exec  :: (load 16)
+    $vgpr14 = IMAGE_SAMPLE_LZ_V1_V2_gfx10 $vgpr5_vgpr6, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 1, 1, 0, 0, 0, 0, 0, 0, 0, implicit $exec :: (dereferenceable load 16 from custom "ImageResource")
+    $vgpr20_vgpr21_vgpr22_vgpr23 = IMAGE_SAMPLE_D_V4_V9_nsa_gfx10 $vgpr3, $vgpr8, $vgpr7, $vgpr5, $vgpr4, $vgpr6, $vgpr0, $vgpr2, $vgpr2, $sgpr0_sgpr1_sgpr2_sgpr3_sgpr4_sgpr5_sgpr6_sgpr7, $sgpr8_sgpr9_sgpr10_sgpr11, 15, 2, 0, 0, 0, 0, 0, 0, 0, implicit $exec  :: (load 16)
+...