[llvm] [AMDGPU] Marking super-reg as implicit-def in first spill instruction (PR #114773)

[Christudasan Devadasan via llvm-commits]

================
@@ -0,0 +1,72 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=amdgcn -mcpu=gfx908 %s -o - -run-pass prologepilog,machine-cp -verify-machineinstrs | FileCheck -check-prefix=GFX908 %s
+
+# When VGPRs are available for spilling, prologepilog marks the tuple implicit-def as well as implicit in the first spill instruction.
+# As a consequence, machine-cp would NOT delete agpr2 copy here.
+
+---
+name:  agpr-spill-to-vgpr-machine-cp
+tracksRegLiveness: true
+stack:
+  - { id: 0, name: '', type: spill-slot, offset: 0, size: 128, alignment: 4 }
+machineFunctionInfo:
+  scratchRSrcReg:  $sgpr0_sgpr1_sgpr2_sgpr3
+  stackPtrOffsetReg: '$sgpr32'
+  hasSpilledVGPRs: true
+body: |
+  bb.0:
+    successors:
+    liveins: $vgpr0, $vgpr1
+
+    ; GFX908-LABEL: name: agpr-spill-to-vgpr-machine-cp
+    ; GFX908: liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3, $vgpr4, $vgpr5, $vgpr6, $vgpr7, $vgpr8, $vgpr9, $vgpr10, $vgpr11, $vgpr12, $vgpr13, $vgpr14, $vgpr15, $vgpr16, $vgpr17, $vgpr18, $vgpr19, $vgpr20, $vgpr21, $vgpr22, $vgpr23, $vgpr24, $vgpr25, $vgpr26, $vgpr27, $vgpr28, $vgpr29, $vgpr30, $vgpr31, $vgpr32, $vgpr33
+    ; GFX908-NEXT: {{  $}}
+    ; GFX908-NEXT: renamable $agpr0 = COPY renamable $vgpr0, implicit $exec
+    ; GFX908-NEXT: renamable $agpr2 = COPY renamable $vgpr1, implicit $exec
+    ; GFX908-NEXT: $vgpr33 = V_ACCVGPR_READ_B32_e64 $agpr0, implicit $exec, implicit-def $agpr0_agpr1_agpr2, implicit $agpr0_agpr1_agpr2
+    ; GFX908-NEXT: $vgpr32 = V_ACCVGPR_READ_B32_e64 $agpr1, implicit $exec
+    ; GFX908-NEXT: $vgpr31 = V_ACCVGPR_READ_B32_e64 $agpr2, implicit $exec, implicit $agpr0_agpr1_agpr2
+    ; GFX908-NEXT: S_ENDPGM 0
+    renamable $agpr0 = COPY renamable $vgpr0, implicit $exec
+    renamable $agpr2 = COPY renamable $vgpr1, implicit $exec
+    SI_SPILL_AV96_SAVE $agpr0_agpr1_agpr2, %stack.0, $sgpr32, 0, implicit $exec :: (store (s96) into %stack.0, align 4, addrspace 5)
+    S_ENDPGM 0
+...
+
+# When VGPRs are NOT available for spilling (stack is used), prologepilog marks the tuple implicit-def only and NOT implicit.
----------------
cdevadas wrote:

> Is this representative of the real testcase where you encountered it? I thought we were reserving a VGPR on 908 for this case. We should never need the emergency stack slot spill to handle the copy, it should trivially use the reserved register

The concern here is not the VGPR but the AGPR tuple and its liveness representation when we peel off its single-spill pseudo into individual spill stores for its subregs. Yes, we do have the vgprForAGPRCopy field introduced for reserving a VGPR for the same and its serialized version is available too.

https://github.com/llvm/llvm-project/pull/114773