[PATCH] D125459: [FastISel] Fix load folding for registers with fixups

[Nikita Popov via Phabricator via llvm-commits]

nikic created this revision.
nikic added reviewers: ab, RKSimon, craig.topper.
Herald added subscribers: StephenFan, pengfei, hiraditya.
Herald added a project: All.
nikic requested review of this revision.
Herald added a project: LLVM.
Herald added a subscriber: llvm-commits.

FastISel tries to fold loads into the single using instruction. However, if the register has fixups, then there may be additional uses through an alias of the register.

In particular, this fixes the problem reported at https://reviews.llvm.org/D119432#3507087. The load register is (at the time of load folding) only used in a single call instruction. However, selection of the bitcast has added a fixup between the load register and the cross-BB register of the bitcast result. After fixups are applied, there would now be two uses of the load register, so load folding is not legal.


https://reviews.llvm.org/D125459

Files:
  llvm/lib/CodeGen/SelectionDAG/FastISel.cpp
  llvm/test/CodeGen/X86/fast-isel-load-bitcast-fold.ll


Index: llvm/test/CodeGen/X86/fast-isel-load-bitcast-fold.ll
===================================================================
--- /dev/null
+++ llvm/test/CodeGen/X86/fast-isel-load-bitcast-fold.ll
@@ -0,0 +1,36 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -O0 -mtriple=x86_64-- -verify-machineinstrs < %s | FileCheck %s
+
+define void @repro(i8** %a0, i1 %a1) nounwind {
+; CHECK-LABEL: repro:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    subq $24, %rsp
+; CHECK-NEXT:    movb %sil, %al
+; CHECK-NEXT:    movb %al, {{[-0-9]+}}(%r{{[sb]}}p) # 1-byte Spill
+; CHECK-NEXT:    movq (%rdi), %rax
+; CHECK-NEXT:    movq %rax, {{[-0-9]+}}(%r{{[sb]}}p) # 8-byte Spill
+; CHECK-NEXT:    callq *%rax
+; CHECK-NEXT:    movb {{[-0-9]+}}(%r{{[sb]}}p), %al # 1-byte Reload
+; CHECK-NEXT:    testb $1, %al
+; CHECK-NEXT:    jne .LBB0_1
+; CHECK-NEXT:    jmp .LBB0_2
+; CHECK-NEXT:  .LBB0_1: # %bb1
+; CHECK-NEXT:    movq {{[-0-9]+}}(%r{{[sb]}}p), %rax # 8-byte Reload
+; CHECK-NEXT:    callq *%rax
+; CHECK-NEXT:    addq $24, %rsp
+; CHECK-NEXT:    retq
+; CHECK-NEXT:  .LBB0_2: # %bb2
+; CHECK-NEXT:    addq $24, %rsp
+; CHECK-NEXT:    retq
+  %tmp0 = load i8*, i8** %a0
+  %tmp1 = bitcast i8* %tmp0 to void ()*
+  call void %tmp1()
+  br i1 %a1, label %bb1, label %bb2
+
+bb1:
+  call void %tmp1()
+  ret void
+
+bb2:
+  ret void
+}
Index: llvm/lib/CodeGen/SelectionDAG/FastISel.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/FastISel.cpp
+++ llvm/lib/CodeGen/SelectionDAG/FastISel.cpp
@@ -2230,6 +2230,11 @@
   if (!MRI.hasOneUse(LoadReg))
     return false;
 
+  // If the register has fixups, there may be additional uses through a
+  // different alias of the register.
+  if (FuncInfo.RegsWithFixups.contains(LoadReg))
+    return false;
+
   MachineRegisterInfo::reg_iterator RI = MRI.reg_begin(LoadReg);
   MachineInstr *User = RI->getParent();
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D125459.428901.patch
Type: text/x-patch
Size: 1949 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220512/6acf0496/attachment.bin>