Volodin Evgeniy. Lab №3 Option №1.

[evldn]

In this lab, a pass was implemented that replaces the argument-dependent mul and add instructions with one - muladd.
The source code:

#include <immintrin.h>

__m128d simple_test(__m128d a, __m128d b, __m128d c, __m128d d) {
	__m128d temp = a * b;
	d = temp + c;
	return d;
}

LLVM MIR code before applying the pass:

body:             |
  bb.0.entry:
    liveins: $xmm0, $xmm1, $xmm2
  
    renamable $xmm0 = nofpexcept MULPDrr killed renamable $xmm0, killed renamable $xmm1, implicit $mxcsr
    renamable $xmm0 = nofpexcept ADDPDrr killed renamable $xmm0, killed renamable $xmm2, implicit $mxcsr
    RET64 $xmm0

LLVM MIR code after applying the pass:

body:             |
  bb.0.entry:
    liveins: $xmm0, $xmm1, $xmm2
  
    $xmm0 = VFMADD213PDr $xmm0, $xmm1, $xmm2, implicit $mxcsr
    RET64 $xmm0

[m-ly4]

Please rewrite it as shared lib plugin.

[m-ly4]

Suggested change

	if (result == end(MIvector))
	return false;
	else
	return true;
	return (result != end(MIvector));

[m-ly4]

Suggested change

	if (iterator->getOpcode() == X86::MULPDrr) {
	if (iterator->getOpcode() != X86::MULPDrr) {
	continue;
	}

[m-ly4]

Please add also:

# __m128d another_test(__m128d a, __m128d b, __m128d c, __m128d d) {
# 	__m128d t = a * b;
# 	__m128d t2 = c * d;
# 	t2 = t + t2;
# 	t = t + c;
# 	return t * t2;
# }

[evldn]

1. I rewrote pass as shared lib plugin.
2. I added the test you suggested, which was performed incorrectly on the old version of the pass. I fixed this by adding additional checks to the pass that track the mutability of operands between the mul and add instructions.