[StackSlotColoring] Ignore non-spill objects in RemoveDeadStores. (#8…

…0242)

The stack slot coloring pass is concerned with optimizing spill
slots. If any change is a pass is made over the function to remove
stack stores that use the same register and stack slot as an
immediately preceding load.
    
The register check is too simple for constant registers like AArch64
and RISC-V's zero register. This register can be used as the result
of a load if we want to discard the result, but still have the memory
access performed. Like for a volatile or atomic load.
    
If the code sees a load from the zero register followed by a store
of the zero register at the same stack slot, the pass mistakenly
believes the store isn't needed.
    
Since the main stack coloring optimization is only concerned with
spill slots, it seems reasonable that RemoveDeadStores should
only be concerned with spills. Since we never generate a reload of
x0, this avoids the issue seen by RISC-V.
    
Test case concept is adapted from pr30821.mir from X86. That test
had to be updated to mark the stack slot as a spill slot.
    
Fixes #80052.

llvm/lib/CodeGen/StackSlotColoring.cpp

@@ -480,7 +480,7 @@ bool StackSlotColoring::RemoveDeadStores(MachineBasicBlock* MBB) {
     if (!(StoreReg = TII->isStoreToStackSlot(*NextMI, SecondSS, StoreSize)))
       continue;
     if (FirstSS != SecondSS || LoadReg != StoreReg || FirstSS == -1 ||
-        LoadSize != StoreSize)
+        LoadSize != StoreSize || !MFI->isSpillSlotObjectIndex(FirstSS))
       continue;
 
     ++NumDead;

llvm/test/CodeGen/RISCV/pr80052.mir

@@ -0,0 +1,116 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 4
+# RUN: llc %s -mtriple=riscv64 -run-pass=greedy,virtregrewriter,stack-slot-coloring -o - | FileCheck %s
+
+---
+name:            foo
+alignment:       4
+tracksRegLiveness: true
+frameInfo:
+  maxAlignment:    4
+  localFrameSize:  4
+stack:
+  - { id: 0, size: 1, alignment: 4, local-offset: -4 }
+machineFunctionInfo:
+  varArgsFrameIndex: 0
+  varArgsSaveSize: 0
+body:             |
+  bb.0.entry:
+    ; To trick stack-slot-colouring to run its dead-store-elimination phase,
+    ; which is at fault, we need the register allocator to run, and spill in two
+    ; places that can have their slots merged. Achieve this by volatile-loading
+    ; data into all allocatable GPRs except $x31. Then, volatile storing them
+    ; later, leaving regalloc only $x31 to play with in the middle.
+    $x1 = LW %stack.0, 0 :: (volatile load (s32))
+    $x5 = LW %stack.0, 0 :: (volatile load (s32))
+    $x6 = LW %stack.0, 0 :: (volatile load (s32))
+    $x7 = LW %stack.0, 0 :: (volatile load (s32))
+    $x8 = LW %stack.0, 0 :: (volatile load (s32))
+    $x9 = LW %stack.0, 0 :: (volatile load (s32))
+    $x10 = LW %stack.0, 0 :: (volatile load (s32))
+    $x11 = LW %stack.0, 0 :: (volatile load (s32))
+    $x12 = LW %stack.0, 0 :: (volatile load (s32))
+    $x13 = LW %stack.0, 0 :: (volatile load (s32))
+    $x14 = LW %stack.0, 0 :: (volatile load (s32))
+    $x15 = LW %stack.0, 0 :: (volatile load (s32))
+    $x16 = LW %stack.0, 0 :: (volatile load (s32))
+    $x17 = LW %stack.0, 0 :: (volatile load (s32))
+    $x18 = LW %stack.0, 0 :: (volatile load (s32))
+    $x19 = LW %stack.0, 0 :: (volatile load (s32))
+    $x20 = LW %stack.0, 0 :: (volatile load (s32))
+    $x21 = LW %stack.0, 0 :: (volatile load (s32))
+    $x22 = LW %stack.0, 0 :: (volatile load (s32))
+    $x23 = LW %stack.0, 0 :: (volatile load (s32))
+    $x24 = LW %stack.0, 0 :: (volatile load (s32))
+    $x25 = LW %stack.0, 0 :: (volatile load (s32))
+    $x26 = LW %stack.0, 0 :: (volatile load (s32))
+    $x27 = LW %stack.0, 0 :: (volatile load (s32))
+    $x28 = LW %stack.0, 0 :: (volatile load (s32))
+    $x29 = LW %stack.0, 0 :: (volatile load (s32))
+    $x30 = LW %stack.0, 0 :: (volatile load (s32))
+
+    ; Force the first spill.
+    %0:gpr = LW %stack.0, 0 :: (volatile load (s32))
+    %1:gpr = LW %stack.0, 0 :: (volatile load (s32))
+    SW %1, %stack.0, 0 :: (volatile store (s32))
+    SW %0, %stack.0, 0 :: (volatile store (s32))
+
+    ; CHECK: renamable $x31 = LW %stack.0, 0 :: (volatile load (s32))
+    ; CHECK-NEXT: SD killed renamable $x31, %stack.1, 0 :: (store (s64) into %stack.1)
+    ; CHECK-NEXT: renamable $x31 = LW %stack.0, 0 :: (volatile load (s32))
+    ; CHECK-NEXT: SW killed renamable $x31, %stack.0, 0 :: (volatile store (s32))
+    ; CHECK-NEXT: renamable $x31 = LD %stack.1, 0 :: (load (s64) from %stack.1)
+    ; CHECK-NEXT: SW killed renamable $x31, %stack.0, 0 :: (volatile store (s32))
+
+    ; stack-slot-coloring doesn't know that a write to $x0 is discarded.
+    dead $x0 = LW %stack.0, 0 :: (volatile load (s32))
+    ; This stores 0 rather than the result of the preceding load since $x0
+    ; is special.
+    ; We don't want this store to be deleted.
+    SW $x0, %stack.0, 0 :: (volatile store (s32))
+
+    ; CHECK-NEXT: dead $x0 = LW %stack.0, 0 :: (volatile load (s32))
+    ; CHECK-NEXT: SW $x0, %stack.0, 0 :: (volatile store (s32))
+
+    ; Force a second spill
+    %2:gpr = LW %stack.0, 0 :: (volatile load (s32))
+    %3:gpr = LW %stack.0, 0 :: (volatile load (s32))
+    SW %3, %stack.0, 0 :: (volatile store (s32))
+    SW %2, %stack.0, 0 :: (volatile store (s32))
+
+    ; CHECK-NEXT: renamable $x31 = LW %stack.0, 0 :: (volatile load (s32))
+    ; CHECK-NEXT: SD killed renamable $x31, %stack.1, 0 :: (store (s64) into %stack.1)
+    ; CHECK-NEXT: renamable $x31 = LW %stack.0, 0 :: (volatile load (s32))
+    ; CHECK-NEXT: SW killed renamable $x31, %stack.0, 0 :: (volatile store (s32))
+    ; CHECK-NEXT: renamable $x31 = LD %stack.1, 0 :: (load (s64) from %stack.1)
+    ; CHECK-NEXT: SW killed renamable $x31, %stack.0, 0 :: (volatile store (s32))
+
+    SW $x1, %stack.0, 0 :: (volatile store (s32))
+    SW $x5, %stack.0, 0 :: (volatile store (s32))
+    SW $x6, %stack.0, 0 :: (volatile store (s32))
+    SW $x7, %stack.0, 0 :: (volatile store (s32))
+    SW $x8, %stack.0, 0 :: (volatile store (s32))
+    SW $x9, %stack.0, 0 :: (volatile store (s32))
+    SW $x10, %stack.0, 0 :: (volatile store (s32))
+    SW $x11, %stack.0, 0 :: (volatile store (s32))
+    SW $x12, %stack.0, 0 :: (volatile store (s32))
+    SW $x13, %stack.0, 0 :: (volatile store (s32))
+    SW $x14, %stack.0, 0 :: (volatile store (s32))
+    SW $x15, %stack.0, 0 :: (volatile store (s32))
+    SW $x16, %stack.0, 0 :: (volatile store (s32))
+    SW $x17, %stack.0, 0 :: (volatile store (s32))
+    SW $x18, %stack.0, 0 :: (volatile store (s32))
+    SW $x19, %stack.0, 0 :: (volatile store (s32))
+    SW $x20, %stack.0, 0 :: (volatile store (s32))
+    SW $x21, %stack.0, 0 :: (volatile store (s32))
+    SW $x22, %stack.0, 0 :: (volatile store (s32))
+    SW $x23, %stack.0, 0 :: (volatile store (s32))
+    SW $x24, %stack.0, 0 :: (volatile store (s32))
+    SW $x25, %stack.0, 0 :: (volatile store (s32))
+    SW $x26, %stack.0, 0 :: (volatile store (s32))
+    SW $x27, %stack.0, 0 :: (volatile store (s32))
+    SW $x28, %stack.0, 0 :: (volatile store (s32))
+    SW $x29, %stack.0, 0 :: (volatile store (s32))
+    SW $x30, %stack.0, 0 :: (volatile store (s32))
+    PseudoRET
+
+...

llvm/test/CodeGen/X86/pr30821.mir

@@ -51,7 +51,7 @@ stack:
   - { id: 1, name: foxtrot, type: default, offset: 0, size: 16, alignment: 16,
       stack-id: default, callee-saved-register: '', callee-saved-restored: true,
       debug-info-variable: '', debug-info-expression: '', debug-info-location: '' }
-  - { id: 2, name: india, type: default, offset: 0, size: 16, alignment: 16,
+  - { id: 2, name: india, type: spill-slot, offset: 0, size: 16, alignment: 16,
       stack-id: default, callee-saved-register: '', callee-saved-restored: true,
       debug-info-variable: '', debug-info-expression: '',
       debug-info-location: '' }