llvm/test/CodeGen/RISCV/bswap-known-bits.ll

; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
; RUN: llc -mtriple=riscv32 -mattr=+zbb -verify-machineinstrs < %s \
; RUN:   | FileCheck %s -check-prefixes=RV32ZB
; RUN: llc -mtriple=riscv64 -mattr=+zbb -verify-machineinstrs < %s \
; RUN:   | FileCheck %s -check-prefixes=RV64ZB

; Source-side known-bits fold: when the bswap operand has at most one byte
; of possibly-nonzero bits at a known byte-aligned position, the bswap is
; equivalent to a shift moving that byte to the mirror byte.

declare i16 @llvm.bswap.i16(i16)
declare i32 @llvm.bswap.i32(i32)
declare i64 @llvm.bswap.i64(i64)

define i16 @bswap_src_and_lo_i16(i16 %x) {
; RV32ZB-LABEL: bswap_src_and_lo_i16:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    slli a0, a0, 8
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_src_and_lo_i16:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    slli a0, a0, 8
; RV64ZB-NEXT:    ret
  %m = and i16 %x, 255
  %b = call i16 @llvm.bswap.i16(i16 %m)
  ret i16 %b
}

define i16 @bswap_src_and_hi_i16(i16 %x) {
; RV32ZB-LABEL: bswap_src_and_hi_i16:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    slli a0, a0, 16
; RV32ZB-NEXT:    srli a0, a0, 24
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_src_and_hi_i16:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    slli a0, a0, 48
; RV64ZB-NEXT:    srli a0, a0, 56
; RV64ZB-NEXT:    ret
  %m = and i16 %x, 65280
  %b = call i16 @llvm.bswap.i16(i16 %m)
  ret i16 %b
}

define i16 @bswap_src_zext_i8_to_i16(i8 %x) {
; RV32ZB-LABEL: bswap_src_zext_i8_to_i16:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    slli a0, a0, 8
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_src_zext_i8_to_i16:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    slli a0, a0, 8
; RV64ZB-NEXT:    ret
  %z = zext i8 %x to i16
  %b = call i16 @llvm.bswap.i16(i16 %z)
  ret i16 %b
}

define i32 @bswap_src_and_byte0_i32(i32 %x) {
; RV32ZB-LABEL: bswap_src_and_byte0_i32:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    slli a0, a0, 24
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_src_and_byte0_i32:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    slliw a0, a0, 24
; RV64ZB-NEXT:    ret
  %m = and i32 %x, 255
  %b = call i32 @llvm.bswap.i32(i32 %m)
  ret i32 %b
}

define i32 @bswap_src_and_byte1_i32(i32 %x) {
; RV32ZB-LABEL: bswap_src_and_byte1_i32:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    lui a1, 16
; RV32ZB-NEXT:    addi a1, a1, -256
; RV32ZB-NEXT:    and a0, a0, a1
; RV32ZB-NEXT:    slli a0, a0, 8
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_src_and_byte1_i32:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    lui a1, 16
; RV64ZB-NEXT:    addi a1, a1, -256
; RV64ZB-NEXT:    and a0, a0, a1
; RV64ZB-NEXT:    slli a0, a0, 8
; RV64ZB-NEXT:    ret
  %m = and i32 %x, 65280
  %b = call i32 @llvm.bswap.i32(i32 %m)
  ret i32 %b
}

define i32 @bswap_src_and_byte2_i32(i32 %x) {
; RV32ZB-LABEL: bswap_src_and_byte2_i32:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    lui a1, 4080
; RV32ZB-NEXT:    and a0, a0, a1
; RV32ZB-NEXT:    srli a0, a0, 8
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_src_and_byte2_i32:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    lui a1, 4080
; RV64ZB-NEXT:    and a0, a0, a1
; RV64ZB-NEXT:    srli a0, a0, 8
; RV64ZB-NEXT:    ret
  %m = and i32 %x, 16711680
  %b = call i32 @llvm.bswap.i32(i32 %m)
  ret i32 %b
}

define i32 @bswap_src_and_byte3_i32(i32 %x) {
; RV32ZB-LABEL: bswap_src_and_byte3_i32:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    srli a0, a0, 24
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_src_and_byte3_i32:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    srliw a0, a0, 24
; RV64ZB-NEXT:    ret
  %m = and i32 %x, 4278190080
  %b = call i32 @llvm.bswap.i32(i32 %m)
  ret i32 %b
}

define i32 @bswap_src_zext_i8_to_i32(i8 %x) {
; RV32ZB-LABEL: bswap_src_zext_i8_to_i32:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    slli a0, a0, 24
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_src_zext_i8_to_i32:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    slliw a0, a0, 24
; RV64ZB-NEXT:    ret
  %z = zext i8 %x to i32
  %b = call i32 @llvm.bswap.i32(i32 %z)
  ret i32 %b
}

define i64 @bswap_src_zext_i8_to_i64(i8 %x) {
; RV32ZB-LABEL: bswap_src_zext_i8_to_i64:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    mv a1, a0
; RV32ZB-NEXT:    li a0, 0
; RV32ZB-NEXT:    slli a1, a1, 24
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_src_zext_i8_to_i64:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    slli a0, a0, 56
; RV64ZB-NEXT:    ret
  %z = zext i8 %x to i64
  %b = call i64 @llvm.bswap.i64(i64 %z)
  ret i64 %b
}

define i64 @bswap_src_zext_i16_to_i64(i16 %x) {
; RV32ZB-LABEL: bswap_src_zext_i16_to_i64:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    mv a1, a0
; RV32ZB-NEXT:    li a0, 0
; RV32ZB-NEXT:    zext.h a1, a1
; RV32ZB-NEXT:    rev8 a1, a1
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_src_zext_i16_to_i64:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    zext.h a0, a0
; RV64ZB-NEXT:    rev8 a0, a0
; RV64ZB-NEXT:    ret
  %z = zext i16 %x to i64
  %b = call i64 @llvm.bswap.i64(i64 %z)
  ret i64 %b
}

; Regression: the producer-side combine must not tag the produced shl
; with nsw. For bswap(and X, 0xFF) on i16 we shift byte 0 into the top
; byte, which becomes the sign bit. Bit 7 of the input freely flips the
; result's sign, so nsw would be unsound and make `icmp slt 0` fold to
; false. Here the comparison must remain: the result depends on bit 7
; of the input.
define i1 @bswap_lo_byte_sign_i16(i16 %x) {
; RV32ZB-LABEL: bswap_lo_byte_sign_i16:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    slli a0, a0, 8
; RV32ZB-NEXT:    sext.h a0, a0
; RV32ZB-NEXT:    srli a0, a0, 31
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_lo_byte_sign_i16:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    slli a0, a0, 8
; RV64ZB-NEXT:    sext.h a0, a0
; RV64ZB-NEXT:    srli a0, a0, 63
; RV64ZB-NEXT:    ret
  %m = and i16 %x, 255
  %b = call i16 @llvm.bswap.i16(i16 %m)
  %neg = icmp slt i16 %b, 0
  ret i1 %neg
}

; Regression test for the known-zero fold: two disjoint byte masks ANDed
; together are always zero, but the DAG combiner doesn't structurally
; simplify to a constant before visiting the bswap. computeKnownBits
; correctly proves the operand is zero, and the combine must fold
; bswap(known-zero) to 0.
define i32 @bswap_nested_and_disjoint_i32(i32 %x, i32 %y) {
; RV32ZB-LABEL: bswap_nested_and_disjoint_i32:
; RV32ZB:       # %bb.0:
; RV32ZB-NEXT:    li a0, 0
; RV32ZB-NEXT:    ret
;
; RV64ZB-LABEL: bswap_nested_and_disjoint_i32:
; RV64ZB:       # %bb.0:
; RV64ZB-NEXT:    li a0, 0
; RV64ZB-NEXT:    ret
  %m1 = and i32 %x, 255
  %m2 = and i32 %y, 65280
  %m3 = and i32 %m1, %m2
  %b = call i32 @llvm.bswap.i32(i32 %m3)
  ret i32 %b
}