[llvm] [AArch64][GlobalISel] Legalize 128-bit types for FABS (PR #104753)

[Thorsten Schütt via llvm-commits]

================
@@ -0,0 +1,82 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
+; RUN: llc -mtriple=aarch64-unknown-linux-gnu -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-SD
+; RUN: llc -mtriple=aarch64-unknown-linux-gnu -global-isel=1 -verify-machineinstrs %s -o - | FileCheck %s --check-prefixes=CHECK,CHECK-GI
+
+define fp128 @fabs_f128(fp128 %a) {
+; CHECK-SD-LABEL: fabs_f128:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    str q0, [sp, #-16]!
+; CHECK-SD-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-SD-NEXT:    ldrb w8, [sp, #15]
+; CHECK-SD-NEXT:    and w8, w8, #0x7f
+; CHECK-SD-NEXT:    strb w8, [sp, #15]
+; CHECK-SD-NEXT:    ldr q0, [sp], #16
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fabs_f128:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    mov x8, v0.d[1]
+; CHECK-GI-NEXT:    mov v0.d[0], v0.d[0]
+; CHECK-GI-NEXT:    and x8, x8, #0x7fffffffffffffff
+; CHECK-GI-NEXT:    mov v0.d[1], x8
+; CHECK-GI-NEXT:    ret
+entry:
+  %c = call fp128 @llvm.fabs.f128(fp128 %a)
+  ret fp128 %c
+}
+
+define <1 x fp128> @fabs_v1f128(<1 x fp128> %a) {
+; CHECK-SD-LABEL: fabs_v1f128:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    str q0, [sp, #-16]!
+; CHECK-SD-NEXT:    .cfi_def_cfa_offset 16
+; CHECK-SD-NEXT:    ldrb w8, [sp, #15]
+; CHECK-SD-NEXT:    and w8, w8, #0x7f
+; CHECK-SD-NEXT:    strb w8, [sp, #15]
+; CHECK-SD-NEXT:    ldr q0, [sp], #16
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fabs_v1f128:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    mov x8, v0.d[1]
+; CHECK-GI-NEXT:    mov v0.d[0], v0.d[0]
+; CHECK-GI-NEXT:    and x8, x8, #0x7fffffffffffffff
+; CHECK-GI-NEXT:    mov v0.d[1], x8
+; CHECK-GI-NEXT:    ret
+entry:
+  %c = call <1 x fp128> @llvm.fabs.v1f128(<1 x fp128> %a)
+  ret <1 x fp128> %c
+}
+
+define <2 x fp128> @fabs_v2f128(<2 x fp128> %a) {
+; CHECK-SD-LABEL: fabs_v2f128:
+; CHECK-SD:       // %bb.0: // %entry
+; CHECK-SD-NEXT:    stp q0, q1, [sp, #-32]!
+; CHECK-SD-NEXT:    .cfi_def_cfa_offset 32
+; CHECK-SD-NEXT:    ldrb w8, [sp, #15]
+; CHECK-SD-NEXT:    and w8, w8, #0x7f
+; CHECK-SD-NEXT:    strb w8, [sp, #15]
+; CHECK-SD-NEXT:    ldrb w8, [sp, #31]
+; CHECK-SD-NEXT:    and w8, w8, #0x7f
+; CHECK-SD-NEXT:    strb w8, [sp, #31]
+; CHECK-SD-NEXT:    ldp q0, q1, [sp], #32
+; CHECK-SD-NEXT:    ret
+;
+; CHECK-GI-LABEL: fabs_v2f128:
+; CHECK-GI:       // %bb.0: // %entry
+; CHECK-GI-NEXT:    mov x8, v0.d[1]
+; CHECK-GI-NEXT:    mov x9, v1.d[1]
+; CHECK-GI-NEXT:    mov v0.d[0], v0.d[0]
+; CHECK-GI-NEXT:    mov v1.d[0], v1.d[0]
+; CHECK-GI-NEXT:    and x8, x8, #0x7fffffffffffffff
+; CHECK-GI-NEXT:    and x9, x9, #0x7fffffffffffffff
+; CHECK-GI-NEXT:    mov v0.d[1], x8
+; CHECK-GI-NEXT:    mov v1.d[1], x9
+; CHECK-GI-NEXT:    ret
+entry:
+  %c = call <2 x fp128> @llvm.fabs.v2f128(<2 x fp128> %a)
----------------
tschuett wrote:

One moment later, non-power of two would be even nicer, e.g., `<3 x fp128>`.

https://github.com/llvm/llvm-project/pull/104753