[AMDGPU] Add GFX11 tests for fcmp and ballot. NFC

[foad]

Seems obviously fine.

llvm/test/CodeGen/AMDGPU/v_cmp_gfx11.ll

15–16

It's a bit nicer to use named values instead of numbered.

37

Same.

[arsenm]

llvm/test/CodeGen/AMDGPU/v_cmp_gfx11.ll

3–4

global-isel=1 and =0 lines?

[Joe_Nash]

llvm/test/CodeGen/AMDGPU/v_cmp_gfx11.ll

3–4

It appears that global-isel selection for 16 bit types in llvm.amdgcn.icmp and even 32 bit types for llvm.amdgcn.fcmp is incomplete. Therefore I will file a ticket but not fix it in this patch. https://github.com/llvm/llvm-project/issues/58326

llvm/test/CodeGen/AMDGPU/v_cmp_gfx11.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefixes=CHECK %s
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mattr=-wavefrontsize32,+wavefrontsize64 --global-isel=0 -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefixes=CHECK %s
 
-define amdgpu_kernel void @test() {
-; CHECK-LABEL: test:
+define amdgpu_kernel void @icmp_test() {
+; CHECK-LABEL: icmp_test:
 ; CHECK:       ; %bb.0: ; %entry
-; CHECK-NEXT:    v_cmp_eq_u16_e64 s0, 0, 0
+; CHECK-NEXT:    v_cmp_eq_u16_e64 s[0:1], 0, 0
 ; CHECK-NEXT:    v_mov_b32_e32 v1, 0
 ; CHECK-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_1) | instid1(SALU_CYCLE_1)
-; CHECK-NEXT:    s_cmp_eq_u32 s0, 0
-; CHECK-NEXT:    s_cselect_b32 s0, -1, 0
-; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s0
+; CHECK-NEXT:    s_cmp_eq_u64 s[0:1], 0
+; CHECK-NEXT:    s_cselect_b64 s[0:1], -1, 0
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[0:1]
 ; CHECK-NEXT:    ds_store_b32 v1, v0
 ; CHECK-NEXT:    s_endpgm
 entry:
-  %0 = tail call i64 @llvm.amdgcn.icmp.i64.i16(i16 0, i16 0, i32 32)
-  %cmp0 = icmp eq i64 %0, 0
-  %add0 = zext i1 %cmp0 to i32
-  store i32 %add0, ptr addrspace(3) null, align 2147483648
+  %icmp.intr = tail call i64 @llvm.amdgcn.icmp.i64.i16(i16 0, i16 0, i32 32)
+  %cmp0 = icmp eq i64 %icmp.intr, 0
+  %zext0 = zext i1 %cmp0 to i32
+  store i32 %zext0, ptr addrspace(3) null, align 2147483648
+  ret void
+}
+
+define amdgpu_kernel void @fcmp_test(half %x, half %y) {
+; CHECK-LABEL: fcmp_test:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    s_load_b32 s0, s[0:1], 0x0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    s_lshr_b32 s1, s0, 16
+; CHECK-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; CHECK-NEXT:    v_cmp_le_f16_e64 s[0:1], s0, s1
+; CHECK-NEXT:    s_cmp_eq_u64 s[0:1], 0
+; CHECK-NEXT:    s_cselect_b64 s[0:1], -1, 0
+; CHECK-NEXT:    s_delay_alu instid0(SALU_CYCLE_1)
+; CHECK-NEXT:    v_cndmask_b32_e64 v0, 0, 1, s[0:1]
+; CHECK-NEXT:    ds_store_b32 v1, v0
+; CHECK-NEXT:    s_endpgm
+entry:
+  %fcmp.intr = tail call i64 @llvm.amdgcn.fcmp.i64.f16(half %x, half %y, i32 5)
+  %cmp0 = icmp eq i64 %fcmp.intr, 0
+  %zext0 = zext i1 %cmp0 to i32
+  store i32 %zext0, ptr addrspace(3) null, align 2147483648
+  ret void
+}
+
+define amdgpu_kernel void @ballot_test(half %x, half %y) {
+; CHECK-LABEL: ballot_test:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_load_b32 s0, s[0:1], 0x0
+; CHECK-NEXT:    v_mov_b32_e32 v2, 0
+; CHECK-NEXT:    s_waitcnt lgkmcnt(0)
+; CHECK-NEXT:    s_lshr_b32 s1, s0, 16
+; CHECK-NEXT:    s_delay_alu instid0(SALU_CYCLE_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; CHECK-NEXT:    v_cmp_eq_f16_e64 s[0:1], s0, s1
+; CHECK-NEXT:    v_mov_b32_e32 v0, s0
+; CHECK-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; CHECK-NEXT:    v_mov_b32_e32 v1, s1
+; CHECK-NEXT:    ds_store_b64 v2, v[0:1]
+; CHECK-NEXT:    s_endpgm
+  %cmp = fcmp oeq half %x, %y
+  %ballot = tail call i64 @llvm.amdgcn.ballot.i64(i1 %cmp)
+  store i64 %ballot, ptr addrspace(3) null, align 2147483648
   ret void
 }
 
-; Function Attrs: convergent nounwind readnone willreturn
 declare i64 @llvm.amdgcn.icmp.i64.i16(i16, i16, i32 immarg) #0
 
+declare i64 @llvm.amdgcn.fcmp.i64.f16(half, half, i32 immarg) #0
+
+declare i64 @llvm.amdgcn.ballot.i64(i1) #0
+
 attributes #0 = { convergent nounwind readnone willreturn }