[PATCH] D21151: AMDGPU: Preserve undef flag on vcc when shrinking v_cndmask_b32

[Matt Arsenault via llvm-commits]

arsenm updated this revision to Diff 60093.
arsenm added a comment.

Don't preserve kill


http://reviews.llvm.org/D21151

Files:
  lib/Target/AMDGPU/SIShrinkInstructions.cpp
  test/CodeGen/AMDGPU/cndmask-no-def-vcc.ll

Index: test/CodeGen/AMDGPU/cndmask-no-def-vcc.ll
===================================================================
--- test/CodeGen/AMDGPU/cndmask-no-def-vcc.ll
+++ test/CodeGen/AMDGPU/cndmask-no-def-vcc.ll
@@ -25,3 +25,29 @@
 bb2:
   ret void
 }
+
+; The undef flag on the condition src must be preserved on the
+; implicit vcc use to avoid verifier errors.
+
+; GCN-LABEL: {{^}}preserve_condition_undef_flag:
+; GCN-NOT: vcc
+; GCN: v_cndmask_b32_e32 v{{[0-9]+}}, 1.0, v{{[0-9]+}}, vcc
+; GCN: v_cndmask_b32_e64 v1, 0, 1, s{{\[[0-9]+:[0-9]+\]}}
+define void @preserve_condition_undef_flag(float %arg, i32 %arg1, float %arg2) {
+bb0:
+  %tmp = icmp sgt i32 %arg1, 4
+  %tmp4 = select i1 undef, float %arg, float 1.000000e+00
+  %tmp5 = fcmp ogt float %arg2, 0.000000e+00
+  %tmp6 = fcmp olt float %arg2, 1.000000e+00
+  %tmp7 = fcmp olt float %arg, %tmp4
+  %tmp8 = and i1 %tmp5, %tmp6
+  %tmp9 = and i1 %tmp8, %tmp7
+  br i1 %tmp9, label %bb1, label %bb2
+
+bb1:
+  store volatile i32 0, i32 addrspace(1)* undef
+  br label %bb2
+
+bb2:
+  ret void
+}
Index: lib/Target/AMDGPU/SIShrinkInstructions.cpp
===================================================================
--- lib/Target/AMDGPU/SIShrinkInstructions.cpp
+++ lib/Target/AMDGPU/SIShrinkInstructions.cpp
@@ -184,18 +184,15 @@
 }
 
 // Copy MachineOperand with all flags except setting it as implicit.
-static MachineOperand copyRegOperandAsImplicit(const MachineOperand &Orig) {
-  assert(!Orig.isImplicit());
-  return MachineOperand::CreateReg(Orig.getReg(),
-                                   Orig.isDef(),
-                                   true,
-                                   Orig.isKill(),
-                                   Orig.isDead(),
-                                   Orig.isUndef(),
-                                   Orig.isEarlyClobber(),
-                                   Orig.getSubReg(),
-                                   Orig.isDebug(),
-                                   Orig.isInternalRead());
+static void copyFlagsToImplicitVCC(MachineInstr &MI,
+                                   const MachineOperand &Orig) {
+
+  for (MachineOperand &Use : MI.implicit_operands()) {
+    if (Use.getReg() == AMDGPU::VCC) {
+      Use.setIsUndef(Orig.isUndef());
+      return;
+    }
+  }
 }
 
 static bool isKImmOperand(const SIInstrInfo *TII, const MachineOperand &Src) {
@@ -398,10 +395,9 @@
           Inst32.addOperand(*Src2);
         } else {
           // In the case of V_CNDMASK_B32_e32, the explicit operand src2 is
-          // replaced with an implicit read of vcc.
-          assert(Src2->getReg() == AMDGPU::VCC &&
-                 "Unexpected missing register operand");
-          Inst32.addOperand(copyRegOperandAsImplicit(*Src2));
+          // replaced with an implicit read of vcc. This was already added
+          // during the initial BuildMI, so find it to preserve the flags.
+          copyFlagsToImplicitVCC(*Inst32, *Src2);
         }
       }
 


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D21151.60093.patch
Type: text/x-patch
Size: 2968 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160608/c54349e9/attachment.bin>