[llvm] r235993 - Fixed crash of variable shift inst on AVX2

Elena Demikhovsky elena.demikhovsky at intel.com
Tue Apr 28 07:46:35 PDT 2015


Author: delena
Date: Tue Apr 28 09:46:35 2015
New Revision: 235993

URL: http://llvm.org/viewvc/llvm-project?rev=235993&view=rev
Log:
Fixed crash of variable shift inst on AVX2

https://llvm.org/bugs/show_bug.cgi?id=22955


Added:
    llvm/trunk/test/CodeGen/X86/shift-avx2-crash.ll   (with props)
Modified:
    llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

Modified: llvm/trunk/lib/Target/X86/X86ISelLowering.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/X86ISelLowering.cpp?rev=235993&r1=235992&r2=235993&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/X86ISelLowering.cpp (original)
+++ llvm/trunk/lib/Target/X86/X86ISelLowering.cpp Tue Apr 28 09:46:35 2015
@@ -16495,9 +16495,7 @@ static SDValue LowerScalarVariableShift(
   }
 
   // Special case in 32-bit mode, where i64 is expanded into high and low parts.
-  if (!Subtarget->is64Bit() &&
-      (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64) ||
-      (Subtarget->hasAVX512() && VT == MVT::v8i64)) &&
+  if (!Subtarget->is64Bit() && VT == MVT::v2i64  &&
       Amt.getOpcode() == ISD::BITCAST &&
       Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
     Amt = Amt.getOperand(0);

Added: llvm/trunk/test/CodeGen/X86/shift-avx2-crash.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/shift-avx2-crash.ll?rev=235993&view=auto
==============================================================================
--- llvm/trunk/test/CodeGen/X86/shift-avx2-crash.ll (added)
+++ llvm/trunk/test/CodeGen/X86/shift-avx2-crash.ll Tue Apr 28 09:46:35 2015
@@ -0,0 +1,38 @@
+; RUN: llc < %s  -mcpu=core-avx2 > /dev/null
+; This test crashed on variable shift creation on AVX2
+
+target datalayout = "e-m:e-p:32:32-f64:32:64-f80:32-n8:16:32-S128"
+target triple = "i386-unknown-linux-gnu"
+
+define void @f_f(float* noalias nocapture %RET, float %aFOO, i32 %div)  {
+allocas:
+  %__idiv_table_u32_offset10_offset_load.i = add i64 0, -2
+  br label %if_then18.i
+
+if_then18.i:
+  %aFOO_load_to_uint32 = fptoui float %aFOO to i32
+  %aFOO_load_to_uint32_broadcast_init = insertelement <8 x i32> undef, i32 %aFOO_load_to_uint32, i32 0
+  %aFOO_load_to_uint32_broadcast = shufflevector <8 x i32> %aFOO_load_to_uint32_broadcast_init, <8 x i32> undef, <8 x i32> zeroinitializer
+
+  %multiplier_load_broadcast_init.i = insertelement <8 x i64> undef, i64 2, i32 0
+  %multiplier_load_broadcast.i = shufflevector <8 x i64> %multiplier_load_broadcast_init.i, <8 x i64> undef, <8 x i32> zeroinitializer
+  %numerator_load_to_uint64.i = zext <8 x i32> %aFOO_load_to_uint32_broadcast to <8 x i64>
+
+  ;if replace '%__idiv_table_u32_offset10_offset_load.i' with '-2' or remove 'if_then18.i' label the error disappears
+  %add__shift_load21.i = add i64 %__idiv_table_u32_offset10_offset_load.i, 32
+  %add__shift_load21_broadcast_init.i = insertelement <8 x i64> undef, i64 %add__shift_load21.i, i32 0
+  %add__shift_load21_broadcast.i = shufflevector <8 x i64> %add__shift_load21_broadcast_init.i, <8 x i64> undef, <8 x i32> zeroinitializer
+
+  %mul_val_load_mult_load.i = mul <8 x i64> %numerator_load_to_uint64.i, %multiplier_load_broadcast.i
+  %bitop22.i = lshr <8 x i64> %mul_val_load_mult_load.i, %add__shift_load21_broadcast.i
+  %bitop22_to_uint32.i = trunc <8 x i64> %bitop22.i to <8 x i32>
+  br label %__fast_idiv___UM_vyuunu.exit
+
+
+__fast_idiv___UM_vyuunu.exit:                   
+  %calltmp_to_float = uitofp <8 x i32> %bitop22_to_uint32.i to <8 x float>
+  %ptrcast = bitcast float* %RET to <8 x float>*
+  store <8 x float> %calltmp_to_float, <8 x float>* %ptrcast, align 4
+  ret void
+}
+

Propchange: llvm/trunk/test/CodeGen/X86/shift-avx2-crash.ll
------------------------------------------------------------------------------
    svn:executable = *





More information about the llvm-commits mailing list