AMDGPU/R600: Assert on infinite loop in EmitClauseMarkers

[arsenm]

Needs a test?

[jvesely]

In D29792#672667, @arsenm wrote:

Needs a test?

I don't know how the test would look like. Instructions that trigger this should be fixed in either isALU() or MakeALUClause().

[jvesely]

add test. simplified version of gromacs failure

[jvesely]

remove leftover metadata

[arsenm]

test/CodeGen/AMDGPU/r600.alu-limits.ll

2 ↗

(On Diff #88656)

Why is it XFAILed?

[jvesely]

test/CodeGen/AMDGPU/r600.alu-limits.ll

2 ↗

(On Diff #88656)

Because it hits the newly added assertion.
This patch changes infinite loop into assertion failure, fix to the underlying cause will enable this test.

[arsenm]

LGTM

[chapuni]

Tweaked in r295591.

llvm/trunk/lib/Target/AMDGPU/R600EmitClauseMarkers.cpp

326

assert() doesn't do anything with NDEBUG.

llvm/trunk/test/CodeGen/AMDGPU/r600.alu-limits.ll

2

This assumes THE INFINITE LOOP WOULD STOP.

[jvesely]

In D29792#681112, @chapuni wrote:

Tweaked in r295591.

thanks

llvm/trunk/lib/Target/AMDGPU/R600EmitClauseMarkers.cpp

bool runOnMachineFunction(MachineFunction &MF) override {
 
     for (MachineFunction::iterator BB = MF.begin(), BB_E = MF.end();
                                                     BB != BB_E; ++BB) {
       MachineBasicBlock &MBB = *BB;
       MachineBasicBlock::iterator I = MBB.begin();
       if (I != MBB.end() && I->getOpcode() == AMDGPU::CF_ALU)
         continue; // BB was already parsed
       for (MachineBasicBlock::iterator E = MBB.end(); I != E;) {
-        if (isALU(*I))
-          I = MakeALUClause(MBB, I);
-        else
+        if (isALU(*I)) {
+          auto next = MakeALUClause(MBB, I);
+          assert(next != I);
+          I = next;
+        } else
           ++I;
       }
     }
     return false;
   }
 
   StringRef getPassName() const override {
     return "R600 Emit Clause Markers Pass";

llvm/trunk/test/CodeGen/AMDGPU/r600.alu-limits.ll

+; RUN: opt -loop-unroll -unroll-threshold=2000 -S < %s | llc -march=r600 -mcpu=cypress | FileCheck %s
+; XFAIL: *
+
+; CHECK: {{^}}@alu_limits:
+
+%struct.foo = type {i32, i32, i32}
+
+define void @alu_limits(i32 addrspace(1)* %out, %struct.foo* %in, i32 %offset) {
+entry:
+  %ptr = getelementptr inbounds %struct.foo, %struct.foo* %in, i32 1, i32 2
+  %x = load i32, i32 *%ptr, align 4
+  br label %loop
+loop:
+  %i = phi i32 [ 100, %entry ], [ %nexti, %loop ]
+  %val = phi i32 [ 1, %entry ], [ %nextval, %loop ]
+
+  %nexti = sub i32 %i, 1
+
+  %y = xor i32 %x, %i
+  %nextval = mul i32 %val, %y
+
+  %cond = icmp ne i32 %nexti, 0
+  br i1 %cond, label %loop, label %end
+end:
+  %out_val = add i32 %nextval, 4
+  store i32 %out_val, i32 addrspace(1)* %out, align 4
+  ret void
+}