[AMDGPU] Use fmac_f64 in "if (cond) a -= c"

[vpykhtin]

PSDB only.

[z1-cciauto]

PSDB Link: https://compiler-ci.amd.com/job/compiler-psdb-amd-staging/2614

[michaelselehov]

5% improvement in 450.md on MI350 (as expected)

[z1-cciauto]

PSDB Link: https://compiler-ci.amd.com/job/compiler-psdb-amd-staging/2693

[michaelselehov]

LGTM. Conditional approve after QA performance testing.
See minor comments in the code.

[michaelselehov]

Optional: Consider adding a comment explaining why we skip loop-invariant selects:

// Check if select and the add/sub are in same loop context.
// Even for loop-invariant select (which might get hoisted), we skip
// the optimization because it wouldn't provide benefit in the loop body
// (same 1 instruction, but worse register pressure: 2 vs 4+ registers).
if (MLI->getLoopFor(MI.getParent()) != MLI->getLoopFor(SelMI->getParent()))
  return false;

This will help future maintainers understand the reasoning.

[michaelselehov]

Consider updating the comment to reflect that constant support is deferred:

// Check if TrueVal is not constant.
// TODO: Support constants in a follow-up patch. Currently disabled due to
// codegen issues that need investigation.
if (TrueConst)
  return false;

[michaelselehov]

Optional: Add validation after LiveIntervals updates to catch potential issues early:

// Validate LiveIntervals after revert transformation
assert(LIS->verify() && "LiveIntervals verification failed after FMA revert");

Note: Check if LIS has verify() or similar method. The -verify-machineinstrs tests should catch issues anyway, but an explicit assert helps with debugging.