[AMDGPU] Fix SIFoldOperands when folding clamp into pseudo scalar instructions

Clamp is canonicaly a v_max* instruction with a VGPR dst. Folding clamp into
a pseudo scalar instruction can cause issue due to a change in regbank. We
fix this with a copy.

Author: mbrkusanin

llvm/lib/Target/AMDGPU/SIFoldOperands.cpp

@@ -1581,7 +1581,18 @@ bool SIFoldOperands::tryFoldClamp(MachineInstr &MI) {
 
   // Clamp is applied after omod, so it is OK if omod is set.
   DefClamp->setImm(1);
-  MRI->replaceRegWith(MI.getOperand(0).getReg(), Def->getOperand(0).getReg());
+
+  Register DefReg = Def->getOperand(0).getReg();
+  Register MIDstReg = MI.getOperand(0).getReg();
+  if (TRI->isSGPRReg(*MRI, DefReg)) {
+    // Psuedo scalar instructions have a SGPR for dst and clamp is a v_max*
+    // instruction with a VGPR dst.
+    BuildMI(*MI.getParent(), MI, MI.getDebugLoc(), TII->get(AMDGPU::COPY),
+            MIDstReg)
+        .addReg(DefReg);
+  } else {
+    MRI->replaceRegWith(MIDstReg, DefReg);
+  }
   MI.eraseFromParent();
 
   // Use of output modifiers forces VOP3 encoding for a VOP2 mac/fmac

llvm/test/CodeGen/AMDGPU/si-fold-scalar-clamp.mir

@@ -0,0 +1,24 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1200 -run-pass=si-fold-operands -verify-machineinstrs -o - %s | FileCheck %s
+---
+name:            test
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $sgpr0
+
+    ; CHECK-LABEL: name: test
+    ; CHECK: liveins: $sgpr0
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr_32 = COPY $sgpr0
+    ; CHECK-NEXT: [[V_S_RSQ_F32_e64_:%[0-9]+]]:sgpr_32 = nofpexcept V_S_RSQ_F32_e64 0, [[COPY]], 1, 0, implicit $mode, implicit $exec
+    ; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr_32 = COPY [[V_S_RSQ_F32_e64_]]
+    ; CHECK-NEXT: EXP_DONE 0, killed [[COPY1]], [[COPY1]], [[COPY1]], [[COPY1]], -1, 0, 15, implicit $exec
+    ; CHECK-NEXT: S_ENDPGM 0
+    %0:sgpr_32 = COPY $sgpr0
+    %1:sgpr_32 = nofpexcept V_S_RSQ_F32_e64 0, %0, 0, 0, implicit $mode, implicit $exec
+    %2:vgpr_32 = nofpexcept V_MAX_F32_e64 0, %1, 0, %1, -1, 0, implicit $mode, implicit $exec
+    EXP_DONE 0, killed %2, %2, %2, %2, -1, 0, 15, implicit $exec
+    S_ENDPGM 0
+
+...