[AMDGPU] Fix whole wavefront mode

[Flakebi]

Summary

We cannot move wwm over exec copies because the exec register needs an exact exec mask.

[arsenm]

llvm/lib/Target/AMDGPU/SIWholeQuadMode.cpp

538

Can't this whole function be replaced with SIInstrInfo::mayReadExec?

llvm/test/CodeGen/AMDGPU/wqm.mir

119–147

Should be able to remove all of this (except tracksRegLiveness: true)

[Flakebi]

Use mayReadEXEC as suggested by Matt.

[nhaehnle]

LGTM

llvm/lib/Target/AMDGPU/SIWholeQuadMode.cpp

 ///
 /// Scalar instructions must occasionally be marked WQM for correct propagation
 /// (e.g. thread masks leading up to branches), but when it comes to actual
 /// execution, they don't care about EXEC.
 bool SIWholeQuadMode::requiresCorrectState(const MachineInstr &MI) const {
   if (MI.isTerminator())
     return true;
 
   // Skip instructions that are not affected by EXEC
   if (TII->isScalarUnit(MI))
     return false;
 
   // Generic instructions such as COPY will either disappear by register
   // coalescing or be lowered to SALU or VALU instructions.
   if (MI.isTransient()) {
     if (MI.getNumExplicitOperands() >= 1) {
-      const MachineOperand &Op = MI.getOperand(0);
+      bool UsesNonSgpr = false;
+      for (unsigned I = 0; I < MI.getNumExplicitOperands(); I++) {
+        const MachineOperand &Op = MI.getOperand(I);
         if (Op.isReg()) {
-        if (TRI->isSGPRReg(*MRI, Op.getReg())) {
           // SGPR instructions are not affected by EXEC
-          return false;
+          if (!TRI->isSGPRReg(*MRI, Op.getReg())
+              || Op.getReg() == AMDGPU::EXEC
+              || Op.getReg() == AMDGPU::EXEC_LO
+              || Op.getReg() == AMDGPU::EXEC_HI) {
+            UsesNonSgpr = true;
+            break;
+          }
+        } else {
+          UsesNonSgpr = true;
+          break;
         }
       }
+      return UsesNonSgpr;
     }
   }
 
   return true;
 }
 
 MachineBasicBlock::iterator
 SIWholeQuadMode::saveSCC(MachineBasicBlock &MBB,

llvm/test/CodeGen/AMDGPU/wqm.mir

bb.0:
     %10:vgpr_32 = V_SET_INACTIVE_B32 %11, undef %12:sreg_32, implicit $exec
     %14:vgpr_32 = COPY %7
     %13:vgpr_32 = V_MOV_B32_dpp %14, killed %10, 323, 12, 15, 0, implicit $exec
     early-clobber %15:vgpr_32 = WWM killed %13, implicit $exec
     BUFFER_STORE_DWORD_OFFSET_exact killed %15, %6, %7, 4, 0, 0, 0, 0, 0, implicit $exec
     S_ENDPGM 0
 
 ...
+
+---
+# Ensure that wwm is not put around an EXEC copy
+#CHECK-LABEL: name: copy_exec
+#CHECK: %7:sreg_64 = COPY $exec
+#CHECK-NEXT: %15:sreg_64 = ENTER_WWM -1, implicit-def $exec, implicit-def $scc, implicit $exec
+#CHECK-NEXT: %8:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+#CHECK-NEXT: $exec = EXIT_WWM %15
+#CHECK-NEXT: %9:vgpr_32 = V_MBCNT_LO_U32_B32_e64 %7.sub0, 0, implicit $exec
+name:            copy_exec
+alignment:       1
+exposesReturnsTwice: false
+legalized:       false
+regBankSelected: false
+selected:        false
+failedISel:      false
+tracksRegLiveness: true
+hasWinCFI:       false
+registers:
+  - { id: 0, class: sgpr_32, preferred-register: '' }
+  - { id: 1, class: sgpr_32, preferred-register: '' }
+  - { id: 2, class: sgpr_32, preferred-register: '' }
+  - { id: 3, class: sgpr_32, preferred-register: '' }
+  - { id: 4, class: sgpr_128, preferred-register: '' }
+  - { id: 5, class: sreg_32, preferred-register: '' }
+  - { id: 7, class: vreg_64, preferred-register: '' }
+  - { id: 8, class: sreg_64, preferred-register: '' }
+  - { id: 9, class: vgpr_32, preferred-register: '' }
+  - { id: 10, class: vgpr_32, preferred-register: '' }
+  - { id: 11, class: vgpr_32, preferred-register: '' }
+  - { id: 12, class: sreg_32, preferred-register: '' }
+  - { id: 13, class: sreg_32, preferred-register: '' }
+  - { id: 14, class: vgpr_32, preferred-register: '' }
+liveins:
+  - { reg: '$sgpr0', virtual-reg: '%0' }
+  - { reg: '$sgpr1', virtual-reg: '%1' }
+  - { reg: '$sgpr2', virtual-reg: '%2' }
+  - { reg: '$sgpr3', virtual-reg: '%3' }
+body:             |
+  bb.0:
+    liveins: $sgpr0, $sgpr1, $sgpr2, $sgpr3
+
+    %3:sgpr_32 = COPY $sgpr3
+    %2:sgpr_32 = COPY $sgpr2
+    %1:sgpr_32 = COPY $sgpr1
+    %0:sgpr_32 = COPY $sgpr0
+    %4:sgpr_128 = REG_SEQUENCE %0, %subreg.sub0, %1, %subreg.sub1, %2, %subreg.sub2, %3, %subreg.sub3
+    %5:sreg_32 = S_MOV_B32 0
+    %6:vreg_64 = BUFFER_LOAD_DWORDX2_OFFSET %4, %5, 0, 0, 0, 0, 0, 0, implicit $exec
+
+    %8:sreg_64 = COPY $exec
+    %9:vgpr_32 = V_MOV_B32_e32 0, implicit $exec
+    %10:vgpr_32 = V_MBCNT_LO_U32_B32_e64 %8.sub0:sreg_64, 0, implicit $exec
+    %11:vgpr_32 = V_MOV_B32_dpp %9:vgpr_32, %10:vgpr_32, 312, 15, 15, 0, implicit $exec
+    %12:sreg_32 = V_READLANE_B32 %11:vgpr_32, 63
+    early-clobber %13:sreg_32 = WWM %9:vgpr_32, implicit $exec
+
+    %14:vgpr_32 = COPY %13
+    BUFFER_STORE_DWORD_OFFSET_exact killed %14, %4, %5, 4, 0, 0, 0, 0, 0, implicit $exec
+    S_ENDPGM 0
+
+...