Make bitcast, extractelement, and insertelement considered cheap for speculation in SimplifyCFG.

[arsenm]

Summary

This help fold more branches into selects.
On R600, vectors are cheap and anything that helps
remove branches is very good.

I don't know why this doesn't use the TTI for this cost
calculation or if it should.

[nadav]

Hi Matt,

I don’t know this code very well but your patch looks okay to me. Please benchmark the LLVM test suite and make sure that this change does not introduce new regressions.

SimplifyCFG should not use TTI because this pass is executed very early in the optimization pipe. TTI should only be used by “lowering” transformations in the late stages of the optimization pipe.

Thanks,
Nadav

[arsenm]

r209914

lib/Transforms/Utils/SimplifyCFG.cpp

static unsigned ComputeSpeculationCost(const User *I) {
   case Instruction::Xor:
   case Instruction::Shl:
   case Instruction::LShr:
   case Instruction::AShr:
   case Instruction::ICmp:
   case Instruction::Trunc:
   case Instruction::ZExt:
   case Instruction::SExt:
+  case Instruction::BitCast:
+  case Instruction::ExtractElement:
+  case Instruction::InsertElement:
     return 1; // These are all cheap.
 
   case Instruction::Call:
   case Instruction::Select:
     return 2;
   }
 }
 

test/Transforms/SimplifyCFG/speculate-vector-ops.ll

+; RUN: opt -S -simplifycfg < %s | FileCheck %s
+
+define i32 @speculate_vector_extract(i32 %d, <4 x i32> %v) #0 {
+; CHECK-LABEL: @speculate_vector_extract(
+; CHECK-NOT: br
+entry:
+  %conv = insertelement <4 x i32> undef, i32 %d, i32 0
+  %conv2 = insertelement <4 x i32> %conv, i32 %d, i32 1
+  %conv3 = insertelement <4 x i32> %conv2, i32 %d, i32 2
+  %conv4 = insertelement <4 x i32> %conv3, i32 %d, i32 3
+  %tmp6 = add nsw <4 x i32> %conv4, <i32 0, i32 -1, i32 -2, i32 -3>
+  %cmp = icmp eq <4 x i32> %tmp6, zeroinitializer
+  %cmp.ext = sext <4 x i1> %cmp to <4 x i32>
+  %tmp8 = extractelement <4 x i32> %cmp.ext, i32 0
+  %tobool = icmp eq i32 %tmp8, 0
+  br i1 %tobool, label %cond.else, label %cond.then
+
+return:                                           ; preds = %cond.end28
+  ret i32 %cond32
+
+cond.then:                                        ; preds = %entry
+  %tmp10 = extractelement <4 x i32> %v, i32 0
+  br label %cond.end
+
+cond.else:                                        ; preds = %entry
+  %tmp12 = extractelement <4 x i32> %v, i32 3
+  br label %cond.end
+
+cond.end:                                         ; preds = %cond.else, %cond.then
+  %cond = phi i32 [ %tmp10, %cond.then ], [ %tmp12, %cond.else ]
+  %tmp14 = extractelement <4 x i32> %cmp.ext, i32 1
+  %tobool15 = icmp eq i32 %tmp14, 0
+  br i1 %tobool15, label %cond.else17, label %cond.then16
+
+cond.then16:                                      ; preds = %cond.end
+  %tmp20 = extractelement <4 x i32> %v, i32 1
+  br label %cond.end18
+
+cond.else17:                                      ; preds = %cond.end
+  br label %cond.end18
+
+cond.end18:                                       ; preds = %cond.else17, %cond.then16
+  %cond22 = phi i32 [ %tmp20, %cond.then16 ], [ %cond, %cond.else17 ]
+  %tmp24 = extractelement <4 x i32> %cmp.ext, i32 2
+  %tobool25 = icmp eq i32 %tmp24, 0
+  br i1 %tobool25, label %cond.else27, label %cond.then26
+
+cond.then26:                                      ; preds = %cond.end18
+  %tmp30 = extractelement <4 x i32> %v, i32 2
+  br label %cond.end28
+
+cond.else27:                                      ; preds = %cond.end18
+  br label %cond.end28
+
+cond.end28:                                       ; preds = %cond.else27, %cond.then26
+  %cond32 = phi i32 [ %tmp30, %cond.then26 ], [ %cond22, %cond.else27 ]
+  br label %return
+}
+
+attributes #0 = { nounwind }