[PATCH] D19208: AMDGPU: Guard VOPC instructions against incorrect commute

[Nicolai Hähnle via llvm-commits]

nhaehnle created this revision.
nhaehnle added reviewers: arsenm, tstellarAMD.
nhaehnle added a subscriber: llvm-commits.
Herald added a subscriber: arsenm.

The added testcase, which triggered this, was derived from a shader-db case
via bugpoint. A separate question is why scalar branching wasn't used.

http://reviews.llvm.org/D19208

Files:
  lib/Target/AMDGPU/SIInstrInfo.cpp
  test/CodeGen/AMDGPU/bug-vopc-commute.ll

Index: test/CodeGen/AMDGPU/bug-vopc-commute.ll
===================================================================
--- /dev/null
+++ test/CodeGen/AMDGPU/bug-vopc-commute.ll
@@ -0,0 +1,49 @@
+; RUN: llc < %s -march=amdgcn -mcpu=verde -verify-machineinstrs | FileCheck %s
+; RUN: llc < %s -march=amdgcn -mcpu=tonga -verify-machineinstrs | FileCheck %s
+
+target triple = "amdgcn--"
+
+; CHECK-LABEL: {{^}}main:
+;
+; Test for compilation only. This generated an invalid machine instruction
+; by trying to commute the operands of a V_CMP_EQ_i32_e32 instruction, both
+; of which were in SGPRs.
+define amdgpu_vs float @main(i32 %v) {
+main_body:
+  %d1 = call float @llvm.SI.load.const(<16 x i8> undef, i32 960)
+  %d2 = call float @llvm.SI.load.const(<16 x i8> undef, i32 976)
+  br i1 undef, label %ENDIF56, label %IF57
+
+IF57:                                             ; preds = %ENDIF
+  %v.1 = mul i32 %v, 2
+  br label %ENDIF56
+
+ENDIF56:                                          ; preds = %IF57, %ENDIF
+  %v.2 = phi i32 [ %v, %main_body ], [ %v.1, %IF57 ]
+  %d1.i = bitcast float %d1 to i32
+  %cc1 = icmp eq i32 %d1.i, 0
+  br i1 %cc1, label %ENDIF59, label %IF60
+
+IF60:                                             ; preds = %ENDIF56
+  %v.3 = mul i32 %v.2, 2
+  br label %ENDIF59
+
+ENDIF59:                                          ; preds = %IF60, %ENDIF56
+  %v.4 = phi i32 [ %v.2, %ENDIF56 ], [ %v.3, %IF60 ]
+  %d2.i = bitcast float %d2 to i32
+  %cc2 = icmp eq i32 %d2.i, 0
+  br i1 %cc2, label %ENDIF62, label %IF63
+
+IF63:                                             ; preds = %ENDIF59
+  unreachable
+
+ENDIF62:                                          ; preds = %ENDIF59
+  %r = bitcast i32 %v.4 to float
+  ret float %r
+}
+
+; Function Attrs: nounwind readnone
+declare float @llvm.SI.load.const(<16 x i8>, i32) #0
+
+attributes #0 = { nounwind readnone }
+attributes #1 = { readnone }
Index: lib/Target/AMDGPU/SIInstrInfo.cpp
===================================================================
--- lib/Target/AMDGPU/SIInstrInfo.cpp
+++ lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -944,10 +944,10 @@
   MachineOperand &Src1 = MI->getOperand(Src1Idx);
 
 
-  if (isVOP2(*MI)) {
+  if (isVOP2(*MI) /*|| isVOPC(*MI)*/) {
     const MCInstrDesc &InstrDesc = MI->getDesc();
-    // For VOP2 instructions, any operand type is valid to use for src0.  Make
-    // sure we can use the src1 as src0.
+    // For VOP2 and VOPC instructions, any operand type is valid to use for
+    // src0.  Make sure we can use the src0 as src1.
     //
     // We could be stricter here and only allow commuting if there is a reason
     // to do so. i.e. if both operands are VGPRs there is no real benefit,


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D19208.54023.patch
Type: text/x-patch
Size: 2707 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20160418/0536eaf0/attachment.bin>