InstCombine: Handle shrinking fneg used by fptrunc

[arsenm]

Summary

This would have regressed when the fsub special case was dropped

[foad]

This would have regressed when the fsub special case was dropped

What happened? Got a link to a patch?

[arsenm]

In D150074#4329515, @foad wrote:

This would have regressed when the fsub special case was dropped

What happened? Got a link to a patch?

fneg used to not exist and we had fsub -0, x = "fneg"

[foad]

Right. LGTM. Should probably handle fabs too?

[arsenm]

In D150074#4329525, @foad wrote:

Right. LGTM. Should probably handle fabs too?

It's already handled under the unary intrinsic case

[foad]

llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp

1744

This comment seems misplaced.

1753

Now I'm confused. What does this code handle and how is it different from what you added?

[arsenm]

llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp

1753

I'm confused too, I saw these not fold in a test but the tests I added work without this

llvm/lib/Transforms/InstCombine/InstCombineCasts.cpp

switch (BO->getOpcode()) {
            LHS = Builder.CreateFPTrunc(BO->getOperand(0), RHSMinType);
            RHS = Builder.CreateFPTrunc(BO->getOperand(1), RHSMinType);
         }
 
         Value *ExactResult = Builder.CreateFRemFMF(LHS, RHS, BO);
         return CastInst::CreateFPCast(ExactResult, Ty);
       }
     }
+  } else if (auto *UO = dyn_cast<UnaryOperator>(FPT.getOperand(0));
+             UO && UO->hasOneUse()) {
+    switch (UO->getOpcode()) {
+    case Instruction::FNeg: {
+        Value *TruncSrc = Builder.CreateFPTrunc(UO->getOperand(0), Ty);
+        return UnaryOperator::CreateFNegFMF(TruncSrc, UO);
+    }
+    default:
+        break;
+    }
   }
 
   // (fptrunc (fneg x)) -> (fneg (fptrunc x))
   Value *X;
   Instruction *Op = dyn_cast<Instruction>(FPT.getOperand(0));
   if (Op && Op->hasOneUse()) {
     // FIXME: The FMF should propagate from the fptrunc, not the source op.
     IRBuilder<>::FastMathFlagGuard FMFG(Builder);
     if (isa<FPMathOperator>(Op))
       Builder.setFastMathFlags(Op->getFastMathFlags());
 
     if (match(Op, m_FNeg(m_Value(X)))) {
       Value *InnerTrunc = Builder.CreateFPTrunc(X, Ty);
 
       return UnaryOperator::CreateFNegFMF(InnerTrunc, Op);
     }
 
     // If we are truncating a select that has an extended operand, we can
     // narrow the other operand and do the select as a narrow op.
     Value *Cond, *X, *Y;

llvm/test/Transforms/InstCombine/double-float-shrink-2.ll

 ;
   %D = fpext half %C to double
   store volatile double %D, ptr undef
   %E = call double @llvm.fabs.f64(double %D)
   %F = fptrunc double %E to float
   ret float %F
 }
 
+define float @test_shrink_intrin_fneg_double_src(double %D) {
+; CHECK-LABEL: @test_shrink_intrin_fneg_double_src(
+; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc double [[D:%.*]] to float
+; CHECK-NEXT:    [[F:%.*]] = fneg float [[TMP1]]
+; CHECK-NEXT:    ret float [[F]]
+;
+  %E = fneg double %D
+  %F = fptrunc double %E to float
+  ret float %F
+}
+
+define <2 x float> @test_shrink_intrin_fneg_v2f64_src(<2 x double> %D) {
+; CHECK-LABEL: @test_shrink_intrin_fneg_v2f64_src(
+; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc <2 x double> [[D:%.*]] to <2 x float>
+; CHECK-NEXT:    [[F:%.*]] = fneg <2 x float> [[TMP1]]
+; CHECK-NEXT:    ret <2 x float> [[F]]
+;
+  %E = fneg <2 x double> %D
+  %F = fptrunc <2 x double> %E to <2 x float>
+  ret <2 x float> %F
+}
+
+; Make sure fast math flags are preserved
+define float @test_shrink_intrin_fneg_nnan_double_src(double %D) {
+; CHECK-LABEL: @test_shrink_intrin_fneg_nnan_double_src(
+; CHECK-NEXT:    [[TMP1:%.*]] = fptrunc double [[D:%.*]] to float
+; CHECK-NEXT:    [[F:%.*]] = fneg nnan float [[TMP1]]
+; CHECK-NEXT:    ret float [[F]]
+;
+  %E = fneg nnan double %D
+  %F = fptrunc double %E to float
+  ret float %F
+}
+
+define float @test_shrink_intrin_fneg_double_src_multi_use(double %D, ptr %ptr) {
+; CHECK-LABEL: @test_shrink_intrin_fneg_double_src_multi_use(
+; CHECK-NEXT:    [[E:%.*]] = fneg double [[D:%.*]]
+; CHECK-NEXT:    store double [[E]], ptr [[PTR:%.*]], align 8
+; CHECK-NEXT:    [[F:%.*]] = fptrunc double [[E]] to float
+; CHECK-NEXT:    ret float [[F]]
+;
+  %E = fneg double %D
+  store double %E, ptr %ptr
+  %F = fptrunc double %E to float
+  ret float %F
+}
+
 ; DBG-VALID: CheckModuleDebugify: PASS