[llvm] [GlobalISel] Combine G_UNMERGE_VALUES with anyext and build vector (PR #112370)

[Amara Emerson via llvm-commits]

Thorsten =?utf-8?q?Schütt?= <schuett at gmail.com>,
Thorsten =?utf-8?q?Schütt?= <schuett at gmail.com>,
Thorsten =?utf-8?q?Schütt?= <schuett at gmail.com>,
Thorsten =?utf-8?q?Schütt?= <schuett at gmail.com>,
Thorsten =?utf-8?q?Schütt?= <schuett at gmail.com>,
Thorsten =?utf-8?q?Schütt?= <schuett at gmail.com>,
Thorsten =?utf-8?q?Schütt?= <schuett at gmail.com>
Message-ID:
In-Reply-To: <llvm.org/llvm/llvm-project/pull/112370 at github.com>


================
@@ -498,21 +500,23 @@ define void @or_v4i8(ptr %p1, ptr %p2) {
 ; CHECK-GI-NEXT:    fmov s0, w8
 ; CHECK-GI-NEXT:    fmov s1, w9
 ; CHECK-GI-NEXT:    mov b2, v0.b[1]
-; CHECK-GI-NEXT:    mov v3.b[0], v0.b[0]
-; CHECK-GI-NEXT:    mov b4, v1.b[1]
-; CHECK-GI-NEXT:    mov v5.b[0], v1.b[0]
-; CHECK-GI-NEXT:    mov v3.b[1], v2.b[0]
-; CHECK-GI-NEXT:    mov b2, v0.b[2]
-; CHECK-GI-NEXT:    mov b0, v0.b[3]
-; CHECK-GI-NEXT:    mov v5.b[1], v4.b[0]
-; CHECK-GI-NEXT:    mov b4, v1.b[2]
-; CHECK-GI-NEXT:    mov b1, v1.b[3]
-; CHECK-GI-NEXT:    mov v3.b[2], v2.b[0]
-; CHECK-GI-NEXT:    mov v5.b[2], v4.b[0]
-; CHECK-GI-NEXT:    mov v3.b[3], v0.b[0]
-; CHECK-GI-NEXT:    mov v5.b[3], v1.b[0]
-; CHECK-GI-NEXT:    ushll v0.8h, v3.8b, #0
-; CHECK-GI-NEXT:    ushll v1.8h, v5.8b, #0
+; CHECK-GI-NEXT:    mov b3, v1.b[1]
+; CHECK-GI-NEXT:    mov b4, v0.b[2]
+; CHECK-GI-NEXT:    mov b5, v0.b[3]
+; CHECK-GI-NEXT:    fmov w8, s2
+; CHECK-GI-NEXT:    mov b2, v1.b[2]
+; CHECK-GI-NEXT:    fmov w9, s3
+; CHECK-GI-NEXT:    mov b3, v1.b[3]
+; CHECK-GI-NEXT:    mov v0.h[1], w8
+; CHECK-GI-NEXT:    mov v1.h[1], w9
+; CHECK-GI-NEXT:    fmov w8, s4
+; CHECK-GI-NEXT:    fmov w9, s2
+; CHECK-GI-NEXT:    mov v0.h[2], w8
+; CHECK-GI-NEXT:    mov v1.h[2], w9
+; CHECK-GI-NEXT:    fmov w8, s5
+; CHECK-GI-NEXT:    fmov w9, s3
+; CHECK-GI-NEXT:    mov v0.h[3], w8
+; CHECK-GI-NEXT:    mov v1.h[3], w9
----------------
aemerson wrote:

Not sure what's happened here, but in general on large modern cores, copies between the GPR and FPR register banks are expensive and this code sequence will likely be slower than the original. Do you know what happened here?

https://github.com/llvm/llvm-project/pull/112370