AMDGPU: Add test showing bit operations that should be reducible (#141837)

v_xor_i64_known_i32_from_range_use_out_of_block demonstrates a regression
that appears in a future patch in the IR division expansion.

We could generalize splitBinaryBitConstantOp to use known bits. I'm
not sure if it's worth it in the original example, since the pattern
seems to disappear if I optimize the division expansion. We should
probably fix the divide expand to avoid this.

Author: arsenm

llvm/test/CodeGen/AMDGPU/bit-op-reduce-width-known-bits.ll

@@ -0,0 +1,162 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx900 < %s | FileCheck %s
+
+; Check for situations where we could reduce the width of bitwise
+; operations.
+
+
+; Should be able to reduce this to a 32-bit or plus a copy
+; https://alive2.llvm.org/ce/z/9LddFX
+define i64 @v_xor_i64_known_hi_i32_from_arg_range(i64 range(i64 0, 4294967296) %arg0, i64 %arg1) {
+; CHECK-LABEL: v_xor_i64_known_hi_i32_from_arg_range:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_xor_b32_e32 v1, v1, v3
+; CHECK-NEXT:    v_xor_b32_e32 v0, v0, v2
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %xor = xor i64 %arg0, %arg1
+  ret i64 %xor
+}
+
+; Should be able to reduce this to a 32-bit or plus a copy
+; https://alive2.llvm.org/ce/z/HaXnBJ
+define i64 @v_or_i64_known_hi_i32_from_arg_range(i64 range(i64 0, 4294967296) %arg0, i64 %arg1) {
+; CHECK-LABEL: v_or_i64_known_hi_i32_from_arg_range:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_or_b32_e32 v1, v1, v3
+; CHECK-NEXT:    v_or_b32_e32 v0, v0, v2
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %or = or i64 %arg0, %arg1
+  ret i64 %or
+}
+
+; https://alive2.llvm.org/ce/z/M96Ror
+; Should be able to reduce this to a 32-bit plus a copy
+define i64 @v_and_i64_known_i32_from_arg_range(i64 range(i64 -4294967296, 0) %arg0, i64 %arg1) {
+; CHECK-LABEL: v_and_i64_known_i32_from_arg_range:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_and_b32_e32 v1, v1, v3
+; CHECK-NEXT:    v_and_b32_e32 v0, v0, v2
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %and = and i64 %arg0, %arg1
+  ret i64 %and
+}
+
+define i64 @s_xor_i64_known_i32_from_arg_range(i64 range(i64 0, 65) inreg %arg) {
+; CHECK-LABEL: s_xor_i64_known_i32_from_arg_range:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_not_b64 s[4:5], s[16:17]
+; CHECK-NEXT:    v_mov_b32_e32 v0, s4
+; CHECK-NEXT:    v_mov_b32_e32 v1, s5
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %xor = xor i64 %arg, -1
+  ret i64 %xor
+}
+
+define i64 @v_xor_i64_known_i32_from_call_range() {
+; CHECK-LABEL: v_xor_i64_known_i32_from_call_range:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    ;;#ASMSTART
+; CHECK-NEXT:    ; def v[0:1]
+; CHECK-NEXT:    ;;#ASMEND
+; CHECK-NEXT:    v_not_b32_e32 v1, v1
+; CHECK-NEXT:    v_not_b32_e32 v0, v0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %call = call range(i64 0, 65) i64 asm "; def $0", "=v"()
+  %xor = xor i64 %call, -1
+  ret i64 %xor
+}
+
+define i64 @s_xor_i64_known_i32_from_call_range() {
+; CHECK-LABEL: s_xor_i64_known_i32_from_call_range:
+; CHECK:       ; %bb.0:
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    ;;#ASMSTART
+; CHECK-NEXT:    ; def s[4:5]
+; CHECK-NEXT:    ;;#ASMEND
+; CHECK-NEXT:    s_not_b64 s[4:5], s[4:5]
+; CHECK-NEXT:    v_mov_b32_e32 v0, s4
+; CHECK-NEXT:    v_mov_b32_e32 v1, s5
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+  %call = call range(i64 0, 65) i64 asm "; def $0", "=s"()
+  %xor = xor i64 %call, -1
+  ret i64 %xor
+}
+
+; Reduced from -amdgpu-codegenprepare-expand-div64 output, produces a
+; not of 0 which ideally would fold out.
+; FIXME: Produces not of constant 0
+define i64 @v_xor_i64_known_i32_from_range_use_out_of_block(i64 %x) {
+; CHECK-LABEL: v_xor_i64_known_i32_from_range_use_out_of_block:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    v_ffbh_u32_e32 v2, v0
+; CHECK-NEXT:    v_add_u32_e32 v2, 32, v2
+; CHECK-NEXT:    v_ffbh_u32_e32 v3, v1
+; CHECK-NEXT:    v_min_u32_e32 v4, v2, v3
+; CHECK-NEXT:    v_mov_b32_e32 v5, 0
+; CHECK-NEXT:    v_cmp_eq_u64_e32 vcc, v[4:5], v[0:1]
+; CHECK-NEXT:    v_mov_b32_e32 v2, 0
+; CHECK-NEXT:    v_mov_b32_e32 v3, 0
+; CHECK-NEXT:    s_and_saveexec_b64 s[4:5], vcc
+; CHECK-NEXT:  ; %bb.1: ; %inc
+; CHECK-NEXT:    v_not_b32_e32 v2, v4
+; CHECK-NEXT:    v_not_b32_e32 v3, 0
+; CHECK-NEXT:    v_add_co_u32_e32 v2, vcc, v0, v2
+; CHECK-NEXT:    v_addc_co_u32_e32 v3, vcc, v1, v3, vcc
+; CHECK-NEXT:  ; %bb.2: ; %UnifiedReturnBlock
+; CHECK-NEXT:    s_or_b64 exec, exec, s[4:5]
+; CHECK-NEXT:    v_mov_b32_e32 v0, v2
+; CHECK-NEXT:    v_mov_b32_e32 v1, v3
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+entry:
+  %ctlz = tail call range(i64 0, 65) i64 @llvm.ctlz.i64(i64 %x, i1 true)
+  %cmp.entry.not = icmp eq i64 %ctlz, %x
+  br i1 %cmp.entry.not, label %inc, label %ret
+
+inc:                                              ; preds = %entry
+  %i1 = xor i64 %ctlz, -1
+  %i2 = add i64 %x, %i1
+  ret i64 %i2
+
+ret:                                              ; preds = %loop, %entry
+  ret i64 0
+}
+
+define i64 @s_xor_i64_known_i32_from_range_use_out_of_block(i64 inreg %x) {
+; CHECK-LABEL: s_xor_i64_known_i32_from_range_use_out_of_block:
+; CHECK:       ; %bb.0: ; %entry
+; CHECK-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; CHECK-NEXT:    s_flbit_i32_b64 s4, s[16:17]
+; CHECK-NEXT:    s_mov_b32 s5, 0
+; CHECK-NEXT:    s_cmp_lg_u64 s[4:5], s[16:17]
+; CHECK-NEXT:    s_cbranch_scc1 .LBB7_2
+; CHECK-NEXT:  ; %bb.1: ; %inc
+; CHECK-NEXT:    s_not_b64 s[4:5], s[4:5]
+; CHECK-NEXT:    s_add_u32 s4, s16, s4
+; CHECK-NEXT:    s_addc_u32 s5, s17, s5
+; CHECK-NEXT:    v_mov_b32_e32 v0, s4
+; CHECK-NEXT:    v_mov_b32_e32 v1, s5
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+; CHECK-NEXT:  .LBB7_2: ; %ret
+; CHECK-NEXT:    v_mov_b32_e32 v0, 0
+; CHECK-NEXT:    v_mov_b32_e32 v1, 0
+; CHECK-NEXT:    s_setpc_b64 s[30:31]
+entry:
+  %ctlz = tail call range(i64 0, 65) i64 @llvm.ctlz.i64(i64 %x, i1 true)
+  %cmp.entry.not = icmp eq i64 %ctlz, %x
+  br i1 %cmp.entry.not, label %inc, label %ret
+
+inc:                                              ; preds = %entry
+  %i1 = xor i64 %ctlz, -1
+  %i2 = add i64 %x, %i1
+  ret i64 %i2
+
+ret:                                              ; preds = %loop, %entry
+  ret i64 0
+}
+