[InstCombine] Missed optimization in math expression: simplify calls exp functions

[Quolyk]

Summary

This patch enables folding following expressions under -ffast-math flag: exp(X) * exp(Y) -> exp(X + Y), exp2(X) * exp2(Y) -> exp2(X + Y). Motivation: https://bugs.llvm.org/show_bug.cgi?id=35594

[hfinkel]

lib/Transforms/InstCombine/InstCombineMulDivRem.cpp

737

Needs

BuilderTy::FastMathFlagGuard Guard(Builder);

first (so that the flags will get unset in the builder as required).

739

Line is too long.

[hfinkel]

lib/Transforms/InstCombine/InstCombineMulDivRem.cpp

736

Also, ->hasOneUse() checks here.

[Quolyk]

Since I'm new to llvm, I looked though other test files, and noticed that it isn't necessary to write full test-check, instead people write just one instruction to make sure test is ok, while consequent instructions doesn't matter. I wonder what style is preferred?

[spatel]

In D41342#960505, @Quolyk wrote:

Since I'm new to llvm, I looked though other test files, and noticed that it isn't necessary to write full test-check, instead people write just one instruction to make sure test is ok, while consequent instructions doesn't matter. I wonder what style is preferred?

Checks used to be mostly written by hand (and had a lot of shortcomings and bugs because of that). Now that we have scripts to auto-generate checks, I prefer that you use those scripts when possible. It should make subsequent changes or additions to the test file easier for other contributors.

[lebedev.ri]

test/Transforms/InstCombine/fmul-exp.ll

16

Please cleanup testcases, name all the variables (-instnamer will help)
(When committing, please commit tests first, so the code change shows the test change)

31–32

I think this can still be transformed, as long as at least one of them has only one use.
This would be transformed into

%2 = call fast double @llvm.exp.f64(double %b)
call void @use(double %2)
%t1 = add fast double %a, %b
%t2 = call fast double @llvm.exp.f64(double %t2)
ret double %t2

which has the same number of instruction,
And add a test where both of them are two-use, which won't be transformed.

test/Transforms/InstCombine/fmul-exp2.ll

68

Please add newlines

[Quolyk]

Update tests.

[spatel]

test/Transforms/InstCombine/fmul-exp.ll

63

We don't need the full "fast" to enable these transforms. Can you adjust the tests to show the minimal amount of FMF necessary to allow the optimizations? We need "reassoc" + anything else?

Do you have commit access? If so, please commit these files to trunk with the current (without this code patch) output.

[Quolyk]

Update tests.

[Quolyk]

test/Transforms/InstCombine/fmul-exp.ll

63

I have commit access, but I don't want to commit without patch approval. Revision is fixed according to your comments.

[spatel]

test/Transforms/InstCombine/fmul-exp.ll

31–32

This comment has not been addressed. I agree with Roman - if there's only 1 extra use, it makes sense to reduce the dependency chain and reduce fmul to fadd.

63

Thanks. I think the patch is close to approval, so you should commit the tests now. There are 2 benefits to committing the tests first:

1. You and the reviewers can see the actual diffs that will occur with the code change. Ie, if something in a test is not already in canonical form, you can correct or show that in the baseline test commit.
2. If the code change must be reverted for some reason, we should not lose the regression test coverage because of that revert.

[Quolyk]

Update tests.

[lebedev.ri]

LGTM, but please wait for @spatel.

test/Transforms/InstCombine/fmul-exp.ll

63

We need "reassoc" + anything else?

Any further thoughts here? Just "reassoc" is enough?
(that is what is currently being done)

[lebedev.ri]

test/Transforms/InstCombine/fmul-exp2.ll

74

Hm that is interesting, somehow i have never seen that before.
So the FMF can be set on such intrinsic calls too.
Here, we only need reassoc on the original fmul, not the exp?

[Quolyk]

test/Transforms/InstCombine/fmul-exp2.ll

74

I've been experimenting with this. As @spatel mentioned we need minimum reassoc calls here. So as it appears we need all fmul reassoc, however call double doesn't need reassoc.

[spatel]

LGTM

test/Transforms/InstCombine/fmul-exp2.ll

74

Yes, FMF can be set for any call. IIRC, we first needed this for llvm.sqrt, but it can be used with any FP call.

And yes, FMF is only required on the fmul here because that value has loosened strictness, so we assume that intermediate values leading up to it may also use that loosened strictness to compute the final result. We should probably make that clearer in the LangRef.

The flag requirement that trips me up most often on these is 'nsz', but I think we're safe here:
exp2(+/-0.0) * exp2(+/-0.0) --> 1 * 1 --> 1
exp2((+-0.0) + (+/-0.0)) --> exp2(+/- 0.0) --> 1

[Quolyk]

@spatel @lebedev.ri thank you for review and helping me get through.

lib/Transforms/InstCombine/InstCombineMulDivRem.cpp

if (OpX && OpY) {
       Log2->setArgOperand(0, OpY);
       Value *FMulVal = Builder.CreateFMul(OpX, Log2);
       Value *FSub = Builder.CreateFSub(FMulVal, OpX);
       FSub->takeName(&I);
       return replaceInstUsesWith(I, FSub);
     }
   }
 
+  if (AllowReassociate) {
+    Value *Opnd0 = nullptr;
+    Value *Opnd1 = nullptr;
+    if (Op0->hasOneUse() && Op1->hasOneUse()) {
+      BuilderTy::FastMathFlagGuard Guard(Builder);
+      Builder.setFastMathFlags(I.getFastMathFlags());
+      
+      // exp(a) * exp(b) -> exp(a + b)
+      if (match(Op0, m_Intrinsic<Intrinsic::exp>(m_Value(Opnd0))) &&
+          match(Op1, m_Intrinsic<Intrinsic::exp>(m_Value(Opnd1)))) {
+        Value *FAddVal = Builder.CreateFAdd(Opnd0, Opnd1);
+        Value *Exp =
+          Intrinsic::getDeclaration(I.getModule(), Intrinsic::exp, I.getType());
+        Value *ExpCall = Builder.CreateCall(Exp, FAddVal);
+        return replaceInstUsesWith(I, ExpCall);
+      }
+      
+      // exp2(a) * exp2(b) -> exp2(a + b)
+      if (match(Op0, m_Intrinsic<Intrinsic::exp2>(m_Value(Opnd0))) &&
+          match(Op1, m_Intrinsic<Intrinsic::exp2>(m_Value(Opnd1)))) {
+        Value *FAddVal = Builder.CreateFAdd(Opnd0, Opnd1);
+        Value *Exp =
+          Intrinsic::getDeclaration(I.getModule(), Intrinsic::exp2, I.getType());
+        Value *ExpCall = Builder.CreateCall(Exp, FAddVal);
+        return replaceInstUsesWith(I, ExpCall);
+      }
+    }
+  }
+
   // Handle symmetric situation in a 2-iteration loop
   Value *Opnd0 = Op0;
   Value *Opnd1 = Op1;
   for (int i = 0; i < 2; i++) {
     bool IgnoreZeroSign = I.hasNoSignedZeros();
     if (BinaryOperator::isFNeg(Opnd0, IgnoreZeroSign)) {
       BuilderTy::FastMathFlagGuard Guard(Builder);
       Builder.setFastMathFlags(I.getFastMathFlags());

test/Transforms/InstCombine/fmul-exp.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -instcombine < %s | FileCheck %s
+
+declare double @llvm.exp.f64(double) nounwind readnone speculatable
+declare void @use(double)
+
+; exp(a) * exp(b) no math flags
+define double @exp_a_exp_b(double %a, double %b) {
+; CHECK-LABEL: @exp_a_exp_b(
+; CHECK-NEXT:    [[TMP1:%.*]] = call double @llvm.exp.f64(double [[A:%.*]])
+; CHECK-NEXT:    [[TMP2:%.*]] = call double @llvm.exp.f64(double [[B:%.*]])
+; CHECK-NEXT:    [[MUL:%.*]] = fmul double [[TMP1]], [[TMP2]]
+; CHECK-NEXT:    ret double [[MUL]]
+;
+  %1 = call double @llvm.exp.f64(double %a)
+  %2 = call double @llvm.exp.f64(double %b)
+  %mul = fmul double %1, %2
+  ret double %mul
+}
+
+; exp(a) * exp(b) fast-math, multiple uses
+define double @exp_a_exp_b_multiple_uses(double %a, double %b) {
+; CHECK-LABEL: @exp_a_exp_b_multiple_uses(
+; CHECK-NEXT:    [[TMP1:%.*]] = call fast double @llvm.exp.f64(double [[A:%.*]])
+; CHECK-NEXT:    [[TMP2:%.*]] = call fast double @llvm.exp.f64(double [[B:%.*]])
+; CHECK-NEXT:    [[MUL:%.*]] = fmul fast double [[TMP1]], [[TMP2]]
+; CHECK-NEXT:    call void @use(double [[TMP2]])
+; CHECK-NEXT:    ret double [[MUL]]
+;
+  %1 = call fast double @llvm.exp.f64(double %a)
+  %2 = call fast double @llvm.exp.f64(double %b)
+  %mul = fmul fast double %1, %2
+  call void @use(double %2)
+  ret double %mul
+}
+
+; exp(a) * exp(b) => exp(a+b) with fast-math
+define double @exp_a_exp_b_fast(double %a, double %b) {
+; CHECK-LABEL: @exp_a_exp_b_fast(
+; CHECK-NEXT:    [[TMP1:%.*]] = fadd fast double [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call fast double @llvm.exp.f64(double [[TMP1]])
+; CHECK-NEXT:    ret double [[TMP2]]
+;
+  %1 = call fast double @llvm.exp.f64(double %a)
+  %2 = call fast double @llvm.exp.f64(double %b)
+  %mul = fmul fast double %1, %2
+  ret double %mul
+}
+
+; exp(a) * exp(b) * exp(c) * exp(d) => exp(a+b+c+d) with fast-math
+define double @exp_a_exp_b_exp_c_exp_d_fast(double %a, double %b, double %c, double %d) {
+; CHECK-LABEL: @exp_a_exp_b_exp_c_exp_d_fast(
+; CHECK-NEXT:    [[TMP1:%.*]] = fadd fast double [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = fadd fast double [[TMP1]], [[C:%.*]]
+; CHECK-NEXT:    [[TMP3:%.*]] = fadd fast double [[TMP2]], [[D:%.*]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call fast double @llvm.exp.f64(double [[TMP3]])
+; CHECK-NEXT:    ret double [[TMP4]]
+;
+  %1 = call fast double @llvm.exp.f64(double %a)
+  %2 = call fast double @llvm.exp.f64(double %b)
+  %mul = fmul fast double %1, %2
+  %3 = call fast double @llvm.exp.f64(double %c)
+  %mul1 = fmul fast double %mul, %3
+  %4 = call fast double @llvm.exp.f64(double %d)
+  %mul2 = fmul fast double %mul1, %4
+  ret double %mul2
+}

test/Transforms/InstCombine/fmul-exp2.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -instcombine < %s | FileCheck %s
+
+declare double @llvm.exp2.f64(double) nounwind readnone speculatable
+declare void @use(double)
+
+; exp2(a) * exp2(b) no math flags
+define double @exp2_a_exp2_b(double %a, double %b) {
+; CHECK-LABEL: @exp2_a_exp2_b(
+; CHECK-NEXT:    [[TMP1:%.*]] = call double @llvm.exp2.f64(double [[A:%.*]])
+; CHECK-NEXT:    [[TMP2:%.*]] = call double @llvm.exp2.f64(double [[B:%.*]])
+; CHECK-NEXT:    [[MUL:%.*]] = fmul double [[TMP1]], [[TMP2]]
+; CHECK-NEXT:    ret double [[MUL]]
+;
+  %1 = call double @llvm.exp2.f64(double %a)
+  %2 = call double @llvm.exp2.f64(double %b)
+  %mul = fmul double %1, %2
+  ret double %mul
+}
+
+; exp2(a) * exp2(b) fast-math, multiple uses
+define double @exp2_a_exp2_b_multiple_uses(double %a, double %b) {
+; CHECK-LABEL: @exp2_a_exp2_b_multiple_uses(
+; CHECK-NEXT:    [[TMP1:%.*]] = call fast double @llvm.exp2.f64(double [[A:%.*]])
+; CHECK-NEXT:    [[TMP2:%.*]] = call fast double @llvm.exp2.f64(double [[B:%.*]])
+; CHECK-NEXT:    [[MUL:%.*]] = fmul fast double [[TMP1]], [[TMP2]]
+; CHECK-NEXT:    call void @use(double [[TMP2]])
+; CHECK-NEXT:    ret double [[MUL]]
+;
+  %1 = call fast double @llvm.exp2.f64(double %a)
+  %2 = call fast double @llvm.exp2.f64(double %b)
+  %mul = fmul fast double %1, %2
+  call void @use(double %2)
+  ret double %mul
+}
+
+; exp2(a) * exp2(b) => exp2(a+b) with fast-math
+define double @exp2_a_exp2_b_fast(double %a, double %b) {
+; CHECK-LABEL: @exp2_a_exp2_b_fast(
+; CHECK-NEXT:    [[TMP1:%.*]] = fadd fast double [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = call fast double @llvm.exp2.f64(double [[TMP1]])
+; CHECK-NEXT:    ret double [[TMP2]]
+;
+  %1 = call fast double @llvm.exp2.f64(double %a)
+  %2 = call fast double @llvm.exp2.f64(double %b)
+  %mul = fmul fast double %1, %2
+  ret double %mul
+}
+
+; exp2(a) * exp2(b) * exp2(c) * exp2(d) => exp2(a+b+c+d) with fast-math
+define double @exp2_a_exp2_b_exp2_c_exp2_d(double %a, double %b, double %c, double %d) {
+; CHECK-LABEL: @exp2_a_exp2_b_exp2_c_exp2_d(
+; CHECK-NEXT:    [[TMP1:%.*]] = fadd fast double [[A:%.*]], [[B:%.*]]
+; CHECK-NEXT:    [[TMP2:%.*]] = fadd fast double [[TMP1]], [[C:%.*]]
+; CHECK-NEXT:    [[TMP3:%.*]] = fadd fast double [[TMP2]], [[D:%.*]]
+; CHECK-NEXT:    [[TMP4:%.*]] = call fast double @llvm.exp2.f64(double [[TMP3]])
+; CHECK-NEXT:    ret double [[TMP4]]
+;
+  %1 = call fast double @llvm.exp2.f64(double %a)
+  %2 = call fast double @llvm.exp2.f64(double %b)
+  %mul = fmul fast double %1, %2
+  %3 = call fast double @llvm.exp2.f64(double %c)
+  %mul1 = fmul fast double %mul, %3
+  %4 = call fast double @llvm.exp2.f64(double %d)
+  %mul2 = fmul fast double %mul1, %4
+  ret double %mul2
+}