Description

When I allowed LONG types in #81454 , I noticed some regressions. They were caused by the IR tree looking like this:
CAST(long <- CAST(int <- short <- long))

On x64, this would emit:

movsx    rdx, cx
movsxd   rdx, edx

The movsxd is redundant as movsx has already sign-extended to 8 bytes and we only ever emit movsx as an 8-byte operation; there will never be a case where we emit movsx edx, cx on 64-bit. Therefore, we can skip emitting movsxd if we know the previous instruction sign-extended the register all the way to 8 bytes.

Example diffs:

        imul     rdx, rdx, 60
        sub      rcx, rdx
        movsx    rdx, word  ptr [rbp-40H]
-       movsxd   rdx, edx
        imul     rdx, rdx, 60
        mov      r8, 0xD1FFAB1E
        cmp      rdx, r8
        jg       G_M18963_IG34
-						;; size=182 bbWeight=0.50 PerfScore 18.88
+						;; size=179 bbWeight=0.50 PerfScore 18.75

        movsx    rax, word  ptr [rcx]
-       cdqe     
        imul     rax, rax, 60
        mov      rdx, 0xD1FFAB1E
        cmp      rax, rdx
        jg       SHORT G_M26742_IG06
-						;; size=25 bbWeight=1.00 PerfScore 7.75
+						;; size=23 bbWeight=1.00 PerfScore 7.50

The cdqe instruction is simply an optimization for size when trying to emit movsxd rax, eax - so it's expected that this was also skipped.

Acceptance Criteria

• CI passes
• Disasm tests