[llvm] r301199 - InstCombine: Fix assert when reassociating fsub with undef

[Matt Arsenault via llvm-commits]

Author: arsenm
Date: Mon Apr 24 12:24:37 2017
New Revision: 301199

URL: http://llvm.org/viewvc/llvm-project?rev=301199&view=rev
Log:
InstCombine: Fix assert when reassociating fsub with undef

There is logic to track the expected number of instructions
produced. It thought in this case an instruction would
be necessary to negate the result, but here it folded
into a ConstantExpr fneg when the non-undef value operand
was cancelled out by the second fsub.

I'm not sure why we don't fold constant FP ops with undef currently,
but I think that would also avoid this problem.

Modified:
    llvm/trunk/lib/Transforms/InstCombine/InstCombineAddSub.cpp
    llvm/trunk/test/Transforms/InstCombine/fsub.ll

Modified: llvm/trunk/lib/Transforms/InstCombine/InstCombineAddSub.cpp
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/InstCombine/InstCombineAddSub.cpp?rev=301199&r1=301198&r2=301199&view=diff
==============================================================================
--- llvm/trunk/lib/Transforms/InstCombine/InstCombineAddSub.cpp (original)
+++ llvm/trunk/lib/Transforms/InstCombine/InstCombineAddSub.cpp Mon Apr 24 12:24:37 2017
@@ -794,6 +794,11 @@ unsigned FAddCombine::calcInstrNumber(co
     if (Opnd->isConstant())
       continue;
 
+    // The constant check above is really for a few special constant
+    // coefficients.
+    if (isa<UndefValue>(Opnd->getSymVal()))
+      continue;
+
     const FAddendCoef &CE = Opnd->getCoef();
     if (CE.isMinusOne() || CE.isMinusTwo())
       NegOpndNum++;

Modified: llvm/trunk/test/Transforms/InstCombine/fsub.ll
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/InstCombine/fsub.ll?rev=301199&r1=301198&r2=301199&view=diff
==============================================================================
--- llvm/trunk/test/Transforms/InstCombine/fsub.ll (original)
+++ llvm/trunk/test/Transforms/InstCombine/fsub.ll Mon Apr 24 12:24:37 2017
@@ -21,3 +21,47 @@ define double @test2(double %x, double %
 
   ret double %t2
 }
+
+; CHECK-LABEL: @fsub_undef(
+; CHECK: %sub = fsub float %val, undef
+define float @fsub_undef(float %val) {
+bb:
+  %sub = fsub float %val, undef
+  ret float %sub
+}
+
+; XXX - Why doesn't this fold to undef?
+; CHECK-LABEL: @fsub_fast_undef(
+; CHCK: %sub = fsub fast float %val, undef
+define float @fsub_fast_undef(float %val) {
+bb:
+  %sub = fsub fast float %val, undef
+  ret float %sub
+}
+
+; CHECK-LABEL: @fneg_undef(
+; CHECK: ret float fsub (float -0.000000e+00, float undef)
+define float @fneg_undef(float %val) {
+bb:
+  %sub = fsub float -0.0, undef
+  ret float %sub
+}
+
+; CHECK-LABEL: @fneg_fast_undef(
+; CHECK: ret float fsub (float -0.000000e+00, float undef)
+define float @fneg_fast_undef(float %val) {
+bb:
+  %sub = fsub fast float -0.0, undef
+  ret float %sub
+}
+
+; This folds to a constant expression, which produced 0 instructions
+; contrary to the expected one for negation.
+; CHECK-LABEL: @inconsistent_numbers_fsub_undef(
+; CHECK: ret float fsub (float -0.000000e+00, float undef)
+define float @inconsistent_numbers_fsub_undef(float %val) {
+bb:
+  %sub0 = fsub fast float %val, undef
+  %sub1 = fsub fast float %sub0, %val
+  ret float %sub1
+}