Kulagin Aleksandr / lab 4 / var 2

[mahbhlddnhakkh]

Has features:

• A support for d = a - b * c and d = b * c - a.
• If the result of a fadd or fsub operation is not used anywhere else, we can safely delete it.

Issues:

• Works correctly with -O1+. With -O0 I cannot bypass store operation because I can't catch it with getDefiningOp. I can catch load operation, but it's operand is not a result of fadd. It's a result of alloca, so it's useless. I hope it's okay.

[mahbhlddnhakkh]

The code has been refactored and I added requested test case f7. The result of this looks good enough for me. Basically:

double c = a * b;
double d = fma(a, b, c);

[m-ly4]

The code has been refactored and I added requested test case f7. The result of this looks good enough for me. Basically:

double c = a * b;
double d = fma(a, b, c);

Well, in "optimized" case we perform 2 multiplications and one addition, while initially there was 1 multiplication and 1 addition. So, I'd prefer to leave the code as is here.

[mahbhlddnhakkh]

The code has been refactored and I added requested test case f7. The result of this looks good enough for me. Basically:

double c = a * b;
double d = fma(a, b, c);

Well, in "optimized" case we perform 2 multiplications and one addition, while initially there was 1 multiplication and 1 addition. So, I'd prefer to leave the code as is here.

But the initial task was "Write a pass that merges multiplication and addition into a single math.fma". In that case, we don't need to use fma if the result of addition operation is used anywhere else because fma will always just replace fadd. So it means that f6 test should also not expect fma for 'optimization'. Am I correct?

[m-ly4]

The code has been refactored and I added requested test case f7. The result of this looks good enough for me. Basically:

double c = a * b;
double d = fma(a, b, c);

Well, in "optimized" case we perform 2 multiplications and one addition, while initially there was 1 multiplication and 1 addition. So, I'd prefer to leave the code as is here.

But the initial task was "Write a pass that merges multiplication and addition into a single math.fma". In that case, we don't need to use fma if the result of addition operation is used anywhere else because fma will always just replace fadd. So it means that f6 test should also not expect fma for 'optimization'. Am I correct?

Strictly speaking, yes, it has small sense if we cannot remove both mul and add operations after replacement. The aim of this replacement is to reduce potential program execution time by replacing two operations with one.

[mahbhlddnhakkh]

The code has been refactored and I added requested test case f7. The result of this looks good enough for me. Basically:

double c = a * b;
double d = fma(a, b, c);

Well, in "optimized" case we perform 2 multiplications and one addition, while initially there was 1 multiplication and 1 addition. So, I'd prefer to leave the code as is here.

But the initial task was "Write a pass that merges multiplication and addition into a single math.fma". In that case, we don't need to use fma if the result of addition operation is used anywhere else because fma will always just replace fadd. So it means that f6 test should also not expect fma for 'optimization'. Am I correct?

Strictly speaking, yes, it has small sense if we cannot remove both mul and add operations after replacement. The aim of this replacement is to reduce potential program execution time by replacing two operations with one.

Done