[llvm] [NVPTX] Optimize v2x16 BUILD_VECTORs to PRMT (PR #116675)

[Artem Belevich via llvm-commits]

================
@@ -197,8 +196,7 @@ define <2 x bfloat> @test_faddx2(<2 x bfloat> %a, <2 x bfloat> %b) #0 {
 ; SM70-NEXT:    setp.nan.f32 %p2, %f6, %f6;
 ; SM70-NEXT:    or.b32 %r21, %r17, 4194304;
 ; SM70-NEXT:    selp.b32 %r22, %r21, %r20, %p2;
-; SM70-NEXT:    { .reg .b16 tmp; mov.b32 {tmp, %rs11}, %r22; }
-; SM70-NEXT:    mov.b32 %r23, {%rs11, %rs7};
+; SM70-NEXT:    prmt.b32 %r23, %r22, %r12, 0x7632U;
----------------
Artem-B wrote:

> @!PT SHFL.IDX PT, RZ, RZ, RZ, RZ

This appears to be a fancy `NOP` as the `@!PT` predicate will always be false.

> LDC RZ, c[0x0][R7+0x160]

The `LDC RZ` instruction appears to be part of some sort of calling convention magic that's not directly related to the calculations we do in the function -- it loads something from the constants/parameters area, and, apparently ignores the result. We can ignore that.

The short version -- the code didn't change.


https://github.com/llvm/llvm-project/pull/116675