[InstCombine] Propagate fast math flags through selects

[foad]

Summary

In SimplifySelectsFeedingBinaryOp, propagate fast math flags from the
outer op into both arms of the new select, to take advantage of
simplifications that require fast math flags.

[spatel]

The code change seems fine, but the lone test does not provide enough coverage. Please have a look and rebase:
rL368028

[foad]

Rebase.

[foad]

In D65658#1616800, @spatel wrote:

The code change seems fine, but the lone test does not provide enough coverage. Please have a look and rebase:
rL368028

Done. Thanks for the test cases! Note that the fdiv case is not affected because InstCombiner::visitFDiv does not call SimplifySelectsFeedingBinaryOp. Perhaps it should.

[spatel]

In D65658#1618618, @foad wrote:

In D65658#1616800, @spatel wrote:

The code change seems fine, but the lone test does not provide enough coverage. Please have a look and rebase:
rL368028

Done. Thanks for the test cases! Note that the fdiv case is not affected because InstCombiner::visitFDiv does not call SimplifySelectsFeedingBinaryOp. Perhaps it should.

Yes - I noticed that test wasn't changing, but I didn't track down why. If you want to mark that test with a TODO comment, that would be good. Or you could fix the fdiv logic first, then update this patch again to show the improvement for fdiv.

LGTM

llvm/trunk/lib/Transforms/InstCombine/InstructionCombining.cpp

Value *InstCombiner::SimplifySelectsFeedingBinaryOp(BinaryOperator &I,
   Instruction::BinaryOps Opcode = I.getOpcode();
   // (op (select (a, b, c)), (select (a, d, e))) -> (select (a, (op b, d), (op
   // c, e)))
   Value *A, *B, *C, *D, *E;
   Value *SI = nullptr;
   if (match(LHS, m_Select(m_Value(A), m_Value(B), m_Value(C))) &&
       match(RHS, m_Select(m_Specific(A), m_Value(D), m_Value(E)))) {
     bool SelectsHaveOneUse = LHS->hasOneUse() && RHS->hasOneUse();
+
+    FastMathFlags FMF;
     BuilderTy::FastMathFlagGuard Guard(Builder);
-    if (isa<FPMathOperator>(&I))
-      Builder.setFastMathFlags(I.getFastMathFlags());
+    if (isa<FPMathOperator>(&I)) {
+      FMF = I.getFastMathFlags();
+      Builder.setFastMathFlags(FMF);
+    }
 
-    Value *V1 = SimplifyBinOp(Opcode, C, E, SQ.getWithInstruction(&I));
-    Value *V2 = SimplifyBinOp(Opcode, B, D, SQ.getWithInstruction(&I));
+    Value *V1 = SimplifyBinOp(Opcode, C, E, FMF, SQ.getWithInstruction(&I));
+    Value *V2 = SimplifyBinOp(Opcode, B, D, FMF, SQ.getWithInstruction(&I));
     if (V1 && V2)
       SI = Builder.CreateSelect(A, V2, V1);
     else if (V2 && SelectsHaveOneUse)
       SI = Builder.CreateSelect(A, V2, Builder.CreateBinOp(Opcode, C, E));
     else if (V1 && SelectsHaveOneUse)
       SI = Builder.CreateSelect(A, Builder.CreateBinOp(Opcode, B, D), V1);
 
     if (SI)

llvm/trunk/test/Transforms/InstCombine/select_arithmetic.ll

;
   %tmp = select i1 %arg, float %div, float 5.000000e+00
   %mul = fmul contract float %tmp, %tmp
   call void @use_float(float %tmp)
   ret float %mul
 }
 
 define float @fmul_nnan_nsz(i1 %cond, float %val) {
 ; CHECK-LABEL: @fmul_nnan_nsz(
-; CHECK-NEXT:    [[LHS:%.*]] = select i1 [[COND:%.*]], float [[VAL:%.*]], float 0.000000e+00
-; CHECK-NEXT:    [[RHS:%.*]] = select i1 [[COND]], float -0.000000e+00, float [[VAL]]
-; CHECK-NEXT:    [[MUL:%.*]] = fmul nnan nsz float [[LHS]], [[RHS]]
-; CHECK-NEXT:    ret float [[MUL]]
+; CHECK-NEXT:    ret float 0.000000e+00
 ;
   %lhs = select i1 %cond, float %val, float +0.0
   %rhs = select i1 %cond, float -0.0, float %val
   %mul = fmul nnan nsz float %lhs, %rhs
   ret float %mul
 }
 
 define <2 x float> @fadd_nsz(<2 x i1> %cond, <2 x float> %val) {
 ; CHECK-LABEL: @fadd_nsz(
-; CHECK-NEXT:    [[LHS:%.*]] = select <2 x i1> [[COND:%.*]], <2 x float> [[VAL:%.*]], <2 x float> zeroinitializer
-; CHECK-NEXT:    [[RHS:%.*]] = select <2 x i1> [[COND]], <2 x float> zeroinitializer, <2 x float> [[VAL]]
-; CHECK-NEXT:    [[ADD:%.*]] = fadd nsz <2 x float> [[LHS]], [[RHS]]
-; CHECK-NEXT:    ret <2 x float> [[ADD]]
+; CHECK-NEXT:    ret <2 x float> [[VAL:%.*]]
 ;
   %lhs = select <2 x i1> %cond, <2 x float> %val, <2 x float> <float +0.0, float +0.0>
   %rhs = select <2 x i1> %cond, <2 x float> <float +0.0, float +0.0>, <2 x float> %val
   %add = fadd nsz <2 x float> %lhs, %rhs
   ret <2 x float> %add
 }
 
 define double @fsub_nnan(i1 %cond, double %val, double %val2) {
 ; CHECK-LABEL: @fsub_nnan(
-; CHECK-NEXT:    [[LHS:%.*]] = select i1 [[COND:%.*]], double [[VAL:%.*]], double [[VAL2:%.*]]
-; CHECK-NEXT:    [[RHS:%.*]] = select i1 [[COND]], double [[VAL]], double 7.000000e+00
-; CHECK-NEXT:    [[ADD:%.*]] = fsub nnan double [[LHS]], [[RHS]]
+; CHECK-NEXT:    [[TMP1:%.*]] = fadd nnan double [[VAL2:%.*]], -7.000000e+00
+; CHECK-NEXT:    [[ADD:%.*]] = select nnan i1 [[COND:%.*]], double 0.000000e+00, double [[TMP1]]
 ; CHECK-NEXT:    ret double [[ADD]]
 ;
   %lhs = select i1 %cond, double %val, double %val2
   %rhs = select i1 %cond, double %val, double 7.0
   %add = fsub nnan double %lhs, %rhs
   ret double %add
 }