[RISCV] Restrict performANY_EXTENDCombine to prevent an infinite loop.

[craig.topper]

Summary

The sign_extend we insert here can get turned into a zero_extend if
the sign bit is known zero. This can enable a setcc combine that
shrinks compares with zero_extend. This reduces the use count of
the zero_extend allowing other combines to turn it back into an
any_extend.

This restricts the combine to only cases where the result is used
by a CopyToReg. This works for my original motivating case. I
hope the CopyToReg use will prevent any converted extends from
turning back into an any_extend.

[nathanchance]

I can confirm that an ARCH=riscv allmodconfig Linux kernel can now successfully build.

[frasercrmck]

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

5820

I'd probably suggest

if (!any_of(N->uses(),
            [](SDNode *User) { return User->getOpcode() == ISD::CopyToReg; }))
  return SDValue();

which I believe is equivalent.

[craig.topper]

Use none_of

[luismarques]

I don't really know if this is the appropriate solution or not for the long term. It at least sounds reasonable for now.

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

5817

Nit: grammar. Missing period?

[nathanchance]

Can this be cherry-picked to release/13.x?

[craig.topper]

In D106754#2910649, @nathanchance wrote:

Can this be cherry-picked to release/13.x?

Filed in bugzilla here https://bugs.llvm.org/show_bug.cgi?id=51206

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

case ISD::MUL:
     if (!Subtarget.hasStdExtM())
       return SDValue();
     LLVM_FALLTHROUGH;
   case ISD::ADD:
   case ISD::SUB:
     break;
   }
 
+  // Only handle cases where the result is used by a CopyToReg that likely
+  // means the value is a liveout of the basic block. This helps prevent
+  // infinite combine loops like PR51206.
+  if (none_of(N->uses(),
+              [](SDNode *User) { return User->getOpcode() == ISD::CopyToReg; }))
+    return SDValue();
+
   SmallVector<SDNode *, 4> SetCCs;
   for (SDNode::use_iterator UI = Src.getNode()->use_begin(),
                             UE = Src.getNode()->use_end();
        UI != UE; ++UI) {
     SDNode *User = *UI;
     if (User == N)
       continue;
     if (UI.getUse().getResNo() != Src.getResNo())

llvm/test/CodeGen/RISCV/pr51206.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+;RUN: llc < %s -mtriple=riscv64-unknown-linux-gnu -mattr=+m | FileCheck %s
+
+; This test used to cause an infinite loop.
+
+@global = global i8 0, align 1
+@global.1 = global i32 0, align 4
+@global.2 = global i8 0, align 1
+@global.3 = global i32 0, align 4
+
+define signext i32 @wobble() nounwind {
+; CHECK-LABEL: wobble:
+; CHECK:       # %bb.0: # %bb
+; CHECK-NEXT:    addi sp, sp, -16
+; CHECK-NEXT:    sd ra, 8(sp) # 8-byte Folded Spill
+; CHECK-NEXT:    lui a0, %hi(global)
+; CHECK-NEXT:    lbu a0, %lo(global)(a0)
+; CHECK-NEXT:    lui a1, %hi(global.2)
+; CHECK-NEXT:    lbu a1, %lo(global.2)(a1)
+; CHECK-NEXT:    addi a0, a0, 1
+; CHECK-NEXT:    lui a2, %hi(global.1)
+; CHECK-NEXT:    sw a0, %lo(global.1)(a2)
+; CHECK-NEXT:    mul a0, a0, a1
+; CHECK-NEXT:    lui a1, 16
+; CHECK-NEXT:    addiw a1, a1, -1
+; CHECK-NEXT:    and a1, a0, a1
+; CHECK-NEXT:    lui a2, 13
+; CHECK-NEXT:    addiw a2, a2, -819
+; CHECK-NEXT:    mul a1, a1, a2
+; CHECK-NEXT:    srli a1, a1, 18
+; CHECK-NEXT:    lui a2, %hi(global.3)
+; CHECK-NEXT:    addi a3, zero, 5
+; CHECK-NEXT:    sw a1, %lo(global.3)(a2)
+; CHECK-NEXT:    bltu a0, a3, .LBB0_2
+; CHECK-NEXT:  # %bb.1: # %bb10
+; CHECK-NEXT:    call quux@plt
+; CHECK-NEXT:  .LBB0_2: # %bb12
+; CHECK-NEXT:    mv a0, zero
+; CHECK-NEXT:    ld ra, 8(sp) # 8-byte Folded Reload
+; CHECK-NEXT:    addi sp, sp, 16
+; CHECK-NEXT:    ret
+bb:
+  %tmp = load i8, i8* @global, align 1
+  %tmp1 = zext i8 %tmp to i32
+  %tmp2 = add nuw nsw i32 %tmp1, 1
+  store i32 %tmp2, i32* @global.1, align 4
+  %tmp3 = load i8, i8* @global.2, align 1
+  %tmp4 = zext i8 %tmp3 to i32
+  %tmp5 = mul nuw nsw i32 %tmp2, %tmp4
+  %tmp6 = trunc i32 %tmp5 to i16
+  %tmp7 = udiv i16 %tmp6, 5
+  %tmp8 = zext i16 %tmp7 to i32
+  store i32 %tmp8, i32* @global.3, align 4
+  %tmp9 = icmp ult i32 %tmp5, 5
+  br i1 %tmp9, label %bb12, label %bb10
+
+bb10:                                             ; preds = %bb
+  %tmp11 = tail call signext i32 bitcast (i32 (...)* @quux to i32 ()*)()
+  br label %bb12
+
+bb12:                                             ; preds = %bb10, %bb
+  ret i32 undef
+}
+
+declare signext i32 @quux(...)