# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
# RUN: llc -debugify-and-strip-all-safe -mtriple aarch64 -O0 -run-pass=aarch64-prelegalizer-combiner --aarch64prelegalizercombinerhelper-only-enable-rule="load_or_combine" -global-isel -verify-machineinstrs %s -o - | FileCheck %s --check-prefix=LITTLE
# RUN: llc -debugify-and-strip-all-safe -mtriple arm64eb -O0 -run-pass=aarch64-prelegalizer-combiner --aarch64prelegalizercombinerhelper-only-enable-rule="load_or_combine" -global-isel -verify-machineinstrs %s -o - | FileCheck %s --check-prefix=BIG
# REQUIRES: asserts
# Test that we can combine patterns like
#
# s8* x = ...
# s32 y = (x[0] | (x[1] << 8)) | ((x[2] << 16) | (x[3] << 24))
#
# Into either a load, or a load with a bswap.
...
---
name: s8_loads_to_s32_little_endian_pat
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; s8* x = ...
; s32 y = (x[0] | (x[1] << 8)) | ((x[2] << 16) | (x[3] << 24))
;
; -> Little endian: Load from x[0]
; -> Big endian: Load from x[0] + BSWAP
; LITTLE-LABEL: name: s8_loads_to_s32_little_endian_pat
; LITTLE: liveins: $x0, $x1
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %full_load:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 1)
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: s8_loads_to_s32_little_endian_pat
; BIG: liveins: $x0, $x1
; BIG: %ptr:_(p0) = COPY $x1
; BIG: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 1)
; BIG: %full_load:_(s32) = G_BSWAP [[LOAD]]
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s32) = G_CONSTANT i32 1
%cst_2:_(s32) = G_CONSTANT i32 2
%cst_3:_(s32) = G_CONSTANT i32 3
%cst_8:_(s32) = G_CONSTANT i32 8
%cst_16:_(s32) = G_CONSTANT i32 16
%cst_24:_(s32) = G_CONSTANT i32 24
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
%ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s32)
%ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s32)
%byte0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
%elt2:_(s32) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s8))
%elt3:_(s32) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s8))
%byte1:_(s32) = nuw G_SHL %elt1, %cst_8(s32)
%byte2:_(s32) = nuw G_SHL %elt2, %cst_16(s32)
%byte3:_(s32) = nuw G_SHL %elt3, %cst_24(s32)
; Note the shape of the tree:
;
; byte byte byte byte
; \ / \ /
; OR OR
; \ /
; \ /
; OR
%or1:_(s32) = G_OR %byte0, %byte1
%or2:_(s32) = G_OR %byte2, %byte3
%full_load:_(s32) = G_OR %or1, %or2
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: s8_loads_to_s32_big_endian_pat
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; s8* x = ...
; s32 y = (x[0] << 24 | (x[1] << 16)) | ((x[2] << 8) | x[3]))
;
; -> Little endian: Load from x[0] + BSWAP
; -> Big endian: Load from x[0]
; LITTLE-LABEL: name: s8_loads_to_s32_big_endian_pat
; LITTLE: liveins: $x0, $x1
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 1)
; LITTLE: %full_load:_(s32) = G_BSWAP [[LOAD]]
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: s8_loads_to_s32_big_endian_pat
; BIG: liveins: $x0, $x1
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %full_load:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 1)
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s32) = G_CONSTANT i32 1
%cst_2:_(s32) = G_CONSTANT i32 2
%cst_3:_(s32) = G_CONSTANT i32 3
%cst_8:_(s32) = G_CONSTANT i32 8
%cst_16:_(s32) = G_CONSTANT i32 16
%cst_24:_(s32) = G_CONSTANT i32 24
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
%ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s32)
%ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s32)
%elt0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
%elt2:_(s32) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s8))
%byte0:_(s32) = nuw G_SHL %elt0, %cst_24(s32)
%byte1:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
%byte2:_(s32) = nuw G_SHL %elt2, %cst_8(s32)
%byte3:_(s32) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s8))
%or1:_(s32) = G_OR %byte0, %byte1
%or2:_(s32) = G_OR %byte2, %byte3
%full_load:_(s32) = G_OR %or1, %or2
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: different_or_pattern
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; Slightly different OR tree.
;
; s8* x = ...
; s32 y = (((x[0] | (x[1] << 8)) | (x[2] << 16)) | (x[3] << 24))
;
; -> Little endian: Load from x[0]
; -> Big endian: Load from x[0] + BSWAP
; LITTLE-LABEL: name: different_or_pattern
; LITTLE: liveins: $x0, $x1
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %full_load:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 1)
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: different_or_pattern
; BIG: liveins: $x0, $x1
; BIG: %ptr:_(p0) = COPY $x1
; BIG: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 1)
; BIG: %full_load:_(s32) = G_BSWAP [[LOAD]]
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s32) = G_CONSTANT i32 1
%cst_2:_(s32) = G_CONSTANT i32 2
%cst_3:_(s32) = G_CONSTANT i32 3
%cst_8:_(s32) = G_CONSTANT i32 8
%cst_16:_(s32) = G_CONSTANT i32 16
%cst_24:_(s32) = G_CONSTANT i32 24
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
%ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s32)
%ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s32)
%byte0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
%elt2:_(s32) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s8))
%elt3:_(s32) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s8))
%byte1:_(s32) = nuw G_SHL %elt1, %cst_8(s32)
%byte2:_(s32) = nuw G_SHL %elt2, %cst_16(s32)
%byte3:_(s32) = nuw G_SHL %elt3, %cst_24(s32)
; Note the shape of the tree:
;
; byte byte
; \ /
; OR_1 byte
; \ /
; OR_2
; \
; ...
%or1:_(s32) = G_OR %byte0, %byte1
%or2:_(s32) = G_OR %or1, %byte2
%full_load:_(s32) = G_OR %or2, %byte3
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: s16_loads_to_s32_little_endian_pat
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; s16* x = ...
; s32 y = x[0] | (x[1] << 16)
;
; -> Little endian: Load from x[0]
; -> Big endian: Load from x[0] + BSWAP
; LITTLE-LABEL: name: s16_loads_to_s32_little_endian_pat
; LITTLE: liveins: $x0, $x1
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %full_load:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 2)
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: s16_loads_to_s32_little_endian_pat
; BIG: liveins: $x0, $x1
; BIG: %ptr:_(p0) = COPY $x1
; BIG: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 2)
; BIG: %full_load:_(s32) = G_BSWAP [[LOAD]]
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_16:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: s16_loads_to_s32_big_endian_pat
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; s16 *x = ...
; s32 y = x[1] | (x[0] << 16)
;
; -> Little endian: Load from x[0] + BSWAP
; -> Big endian: Load from x[0]
; LITTLE-LABEL: name: s16_loads_to_s32_big_endian_pat
; LITTLE: liveins: $x0, $x1
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 2)
; LITTLE: %full_load:_(s32) = G_BSWAP [[LOAD]]
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: s16_loads_to_s32_big_endian_pat
; BIG: liveins: $x0, $x1
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %full_load:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 2)
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_16:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%elt0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%high_half:_(s32) = nuw G_SHL %elt0, %cst_16(s32)
%low_half:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%full_load:_(s32) = G_OR %low_half, %high_half
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: s16_loads_to_s64_little_endian_pat
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; s16 *x = ...
; s32 y = (x[0] | (x[1] << 16)) | ((x[2] << 32) | (x[3] << 48))
;
; -> Little endian: Load from x[0]
; -> Big endian: Load from x[0] + BSWAP
; LITTLE-LABEL: name: s16_loads_to_s64_little_endian_pat
; LITTLE: liveins: $x0, $x1
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %full_load:_(s64) = G_LOAD %ptr(p0) :: (load (s64), align 2)
; LITTLE: $x1 = COPY %full_load(s64)
; LITTLE: RET_ReallyLR implicit $x1
; BIG-LABEL: name: s16_loads_to_s64_little_endian_pat
; BIG: liveins: $x0, $x1
; BIG: %ptr:_(p0) = COPY $x1
; BIG: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD %ptr(p0) :: (load (s64), align 2)
; BIG: %full_load:_(s64) = G_BSWAP [[LOAD]]
; BIG: $x1 = COPY %full_load(s64)
; BIG: RET_ReallyLR implicit $x1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_2:_(s64) = G_CONSTANT i64 2
%cst_3:_(s64) = G_CONSTANT i64 3
%cst_16:_(s64) = G_CONSTANT i64 16
%cst_32:_(s64) = G_CONSTANT i64 32
%cst_48:_(s64) = G_CONSTANT i64 48
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s64)
%ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s64)
%byte0_byte1:_(s64) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%elt1:_(s64) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%elt2:_(s64) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s16))
%elt3:_(s64) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s16))
%byte2_byte3:_(s64) = nuw G_SHL %elt1, %cst_16(s64)
%byte4_byte5:_(s64) = nuw G_SHL %elt2, %cst_32(s64)
%byte6_byte7:_(s64) = nuw G_SHL %elt3, %cst_48(s64)
%or1:_(s64) = G_OR %byte0_byte1, %byte2_byte3
%or2:_(s64) = G_OR %byte4_byte5, %byte6_byte7
%full_load:_(s64) = G_OR %or1, %or2
$x1 = COPY %full_load(s64)
RET_ReallyLR implicit $x1
...
---
name: s16_loads_to_s64_big_endian_pat
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; s16 *x = ...
; s64 y = (x[3] | (x[2] << 16)) | ((x[1] << 32) | (x[0] << 48))
;
; -> Little endian: Load from x[0] + BSWAP
; -> Big endian: Load from x[0]
; LITTLE-LABEL: name: s16_loads_to_s64_big_endian_pat
; LITTLE: liveins: $x0, $x1
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: [[LOAD:%[0-9]+]]:_(s64) = G_LOAD %ptr(p0) :: (load (s64), align 2)
; LITTLE: %full_load:_(s64) = G_BSWAP [[LOAD]]
; LITTLE: $x1 = COPY %full_load(s64)
; LITTLE: RET_ReallyLR implicit $x1
; BIG-LABEL: name: s16_loads_to_s64_big_endian_pat
; BIG: liveins: $x0, $x1
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %full_load:_(s64) = G_LOAD %ptr(p0) :: (load (s64), align 2)
; BIG: $x1 = COPY %full_load(s64)
; BIG: RET_ReallyLR implicit $x1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_2:_(s64) = G_CONSTANT i64 2
%cst_3:_(s64) = G_CONSTANT i64 3
%cst_16:_(s64) = G_CONSTANT i64 16
%cst_32:_(s64) = G_CONSTANT i64 32
%cst_48:_(s64) = G_CONSTANT i64 48
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s64)
%ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s64)
%elt0:_(s64) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%elt1:_(s64) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%elt2:_(s64) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s16))
%byte0_byte1:_(s64) = nuw G_SHL %elt0, %cst_48(s64)
%byte2_byte3:_(s64) = nuw G_SHL %elt1, %cst_32(s64)
%byte4_byte5:_(s64) = nuw G_SHL %elt2, %cst_16(s64)
%byte6_byte7:_(s64) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s16))
%or1:_(s64) = G_OR %byte0_byte1, %byte2_byte3
%or2:_(s64) = G_OR %byte4_byte5, %byte6_byte7
%full_load:_(s64) = G_OR %or1, %or2
$x1 = COPY %full_load(s64)
RET_ReallyLR implicit $x1
...
---
name: nonzero_start_idx_positive_little_endian_pat
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; s8* x = ...
; s32 y = (x[1] | (x[2] << 8)) | ((x[3] << 16) | (x[4] << 24))
;
; -> Little endian: Load from x[1]
; -> Big endian: Load from x[1] + BSWAP
; LITTLE-LABEL: name: nonzero_start_idx_positive_little_endian_pat
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s32) = G_CONSTANT i32 1
; LITTLE: %ptr:_(p0) = COPY $x0
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
; LITTLE: %full_load:_(s32) = G_LOAD %ptr_elt_1(p0) :: (load (s32), align 1)
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: nonzero_start_idx_positive_little_endian_pat
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s32) = G_CONSTANT i32 1
; BIG: %ptr:_(p0) = COPY $x0
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
; BIG: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD %ptr_elt_1(p0) :: (load (s32), align 1)
; BIG: %full_load:_(s32) = G_BSWAP [[LOAD]]
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s32) = G_CONSTANT i32 1
%cst_2:_(s32) = G_CONSTANT i32 2
%cst_3:_(s32) = G_CONSTANT i32 3
%cst_4:_(s32) = G_CONSTANT i32 4
%cst_8:_(s32) = G_CONSTANT i32 8
%cst_16:_(s32) = G_CONSTANT i32 16
%cst_24:_(s32) = G_CONSTANT i32 24
%ptr:_(p0) = COPY $x0
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
%ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s32)
%ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s32)
%ptr_elt_4:_(p0) = G_PTR_ADD %ptr, %cst_4(s32)
%elt2:_(s32) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s8))
%elt3:_(s32) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s8))
%elt4:_(s32) = G_ZEXTLOAD %ptr_elt_4(p0) :: (load (s8))
%byte0:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
%byte1:_(s32) = nuw G_SHL %elt2, %cst_8(s32)
%byte2:_(s32) = nuw G_SHL %elt3, %cst_16(s32)
%byte3:_(s32) = nuw G_SHL %elt4, %cst_24(s32)
%or1:_(s32) = G_OR %byte0, %byte1
%or2:_(s32) = G_OR %byte2, %byte3
%full_load:_(s32) = G_OR %or1, %or2
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: nonzero_start_idx_positive_big_endian_pat
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; s8* x = ...
; s32 y = (x[4] | (x[3] << 8)) | ((x[2] << 16) | (x[1] << 24))
;
; -> Little endian: Load from x[1] + BSWAP
; -> Big endian: Load from x[1]
; LITTLE-LABEL: name: nonzero_start_idx_positive_big_endian_pat
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s32) = G_CONSTANT i32 1
; LITTLE: %ptr:_(p0) = COPY $x0
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
; LITTLE: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD %ptr_elt_1(p0) :: (load (s32), align 1)
; LITTLE: %full_load:_(s32) = G_BSWAP [[LOAD]]
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: nonzero_start_idx_positive_big_endian_pat
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s32) = G_CONSTANT i32 1
; BIG: %ptr:_(p0) = COPY $x0
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
; BIG: %full_load:_(s32) = G_LOAD %ptr_elt_1(p0) :: (load (s32), align 1)
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s32) = G_CONSTANT i32 1
%cst_2:_(s32) = G_CONSTANT i32 2
%cst_3:_(s32) = G_CONSTANT i32 3
%cst_4:_(s32) = G_CONSTANT i32 4
%cst_8:_(s32) = G_CONSTANT i32 8
%cst_16:_(s32) = G_CONSTANT i32 16
%cst_24:_(s32) = G_CONSTANT i32 24
%ptr:_(p0) = COPY $x0
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
%ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s32)
%ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s32)
%ptr_elt_4:_(p0) = G_PTR_ADD %ptr, %cst_4(s32)
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
%elt2:_(s32) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s8))
%elt3:_(s32) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s8))
%byte0:_(s32) = G_ZEXTLOAD %ptr_elt_4(p0) :: (load (s8))
%byte1:_(s32) = nuw G_SHL %elt3, %cst_8(s32)
%byte2:_(s32) = nuw G_SHL %elt2, %cst_16(s32)
%byte3:_(s32) = nuw G_SHL %elt1, %cst_24(s32)
%or1:_(s32) = G_OR %byte0, %byte1
%or2:_(s32) = G_OR %byte2, %byte3
%full_load:_(s32) = G_OR %or1, %or2
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: nonzero_start_idx_negative_little_endian_pat
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; s8* x = ...
; s32 y = (x[-3] | (x[-2] << 8)) | ((x[-1] << 16) | (x[0] << 24))
;
; -> Little endian: Load from x[-3]
; -> Big endian: Load from x[-3] + BSWAP
; LITTLE-LABEL: name: nonzero_start_idx_negative_little_endian_pat
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_neg_3:_(s32) = G_CONSTANT i32 -3
; LITTLE: %ptr:_(p0) = COPY $x0
; LITTLE: %ptr_elt_neg_3:_(p0) = G_PTR_ADD %ptr, %cst_neg_3(s32)
; LITTLE: %full_load:_(s32) = G_LOAD %ptr_elt_neg_3(p0) :: (load (s32), align 1)
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: nonzero_start_idx_negative_little_endian_pat
; BIG: liveins: $x0, $x1
; BIG: %cst_neg_3:_(s32) = G_CONSTANT i32 -3
; BIG: %ptr:_(p0) = COPY $x0
; BIG: %ptr_elt_neg_3:_(p0) = G_PTR_ADD %ptr, %cst_neg_3(s32)
; BIG: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD %ptr_elt_neg_3(p0) :: (load (s32), align 1)
; BIG: %full_load:_(s32) = G_BSWAP [[LOAD]]
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_neg_1:_(s32) = G_CONSTANT i32 -1
%cst_neg_2:_(s32) = G_CONSTANT i32 -2
%cst_neg_3:_(s32) = G_CONSTANT i32 -3
%cst_8:_(s32) = G_CONSTANT i32 8
%cst_16:_(s32) = G_CONSTANT i32 16
%cst_24:_(s32) = G_CONSTANT i32 24
%ptr:_(p0) = COPY $x0
%ptr_elt_neg_3:_(p0) = G_PTR_ADD %ptr, %cst_neg_3(s32)
%ptr_elt_neg_2:_(p0) = G_PTR_ADD %ptr, %cst_neg_2(s32)
%ptr_elt_neg_1:_(p0) = G_PTR_ADD %ptr, %cst_neg_1(s32)
%elt_neg_2:_(s32) = G_ZEXTLOAD %ptr_elt_neg_2(p0) :: (load (s8))
%elt_neg_1:_(s32) = G_ZEXTLOAD %ptr_elt_neg_1(p0) :: (load (s8))
%elt_0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
%byte0:_(s32) = G_ZEXTLOAD %ptr_elt_neg_3(p0) :: (load (s8))
%byte1:_(s32) = nuw G_SHL %elt_neg_2, %cst_8(s32)
%byte2:_(s32) = nuw G_SHL %elt_neg_1, %cst_16(s32)
%byte3:_(s32) = nuw G_SHL %elt_0, %cst_24(s32)
%or1:_(s32) = G_OR %byte0, %byte1
%or2:_(s32) = G_OR %byte2, %byte3
%full_load:_(s32) = G_OR %or1, %or2
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: nonzero_start_idx_negative_big_endian_pat
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; s8* x = ...
; s32 y = (x[0] | (x[-1] << 8)) | ((x[-2] << 16) | (x[-3] << 24))
;
; -> Little endian: Load from x[-3] + BSWAP
; -> Big endian: Load from x[-3]
; LITTLE-LABEL: name: nonzero_start_idx_negative_big_endian_pat
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_neg_3:_(s32) = G_CONSTANT i32 -3
; LITTLE: %ptr:_(p0) = COPY $x0
; LITTLE: %ptr_elt_neg_3:_(p0) = G_PTR_ADD %ptr, %cst_neg_3(s32)
; LITTLE: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD %ptr_elt_neg_3(p0) :: (load (s32), align 1)
; LITTLE: %full_load:_(s32) = G_BSWAP [[LOAD]]
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: nonzero_start_idx_negative_big_endian_pat
; BIG: liveins: $x0, $x1
; BIG: %cst_neg_3:_(s32) = G_CONSTANT i32 -3
; BIG: %ptr:_(p0) = COPY $x0
; BIG: %ptr_elt_neg_3:_(p0) = G_PTR_ADD %ptr, %cst_neg_3(s32)
; BIG: %full_load:_(s32) = G_LOAD %ptr_elt_neg_3(p0) :: (load (s32), align 1)
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_neg_1:_(s32) = G_CONSTANT i32 -1
%cst_neg_2:_(s32) = G_CONSTANT i32 -2
%cst_neg_3:_(s32) = G_CONSTANT i32 -3
%cst_8:_(s32) = G_CONSTANT i32 8
%cst_16:_(s32) = G_CONSTANT i32 16
%cst_24:_(s32) = G_CONSTANT i32 24
%ptr:_(p0) = COPY $x0
%ptr_elt_neg_3:_(p0) = G_PTR_ADD %ptr, %cst_neg_3(s32)
%ptr_elt_neg_2:_(p0) = G_PTR_ADD %ptr, %cst_neg_2(s32)
%ptr_elt_neg_1:_(p0) = G_PTR_ADD %ptr, %cst_neg_1(s32)
%elt_neg_3:_(s32) = G_ZEXTLOAD %ptr_elt_neg_3(p0) :: (load (s8))
%elt_neg_2:_(s32) = G_ZEXTLOAD %ptr_elt_neg_2(p0) :: (load (s8))
%elt_neg_1:_(s32) = G_ZEXTLOAD %ptr_elt_neg_1(p0) :: (load (s8))
%elt_0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
%byte0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
%byte1:_(s32) = nuw G_SHL %elt_neg_1, %cst_8(s32)
%byte2:_(s32) = nuw G_SHL %elt_neg_2, %cst_16(s32)
%byte3:_(s32) = nuw G_SHL %elt_neg_3, %cst_24(s32)
%or1:_(s32) = G_OR %byte0, %byte1
%or2:_(s32) = G_OR %byte2, %byte3
%full_load:_(s32) = G_OR %or1, %or2
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_combine_volatile
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; Combine should only happen with unordered loads.
; LITTLE-LABEL: name: dont_combine_volatile
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s64) = G_CONSTANT i64 1
; LITTLE: %cst_16:_(s32) = G_CONSTANT i32 16
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; LITTLE: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; LITTLE: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (volatile load (s16))
; LITTLE: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; LITTLE: %full_load:_(s32) = G_OR %low_half, %high_half
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_combine_volatile
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s64) = G_CONSTANT i64 1
; BIG: %cst_16:_(s32) = G_CONSTANT i32 16
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; BIG: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; BIG: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (volatile load (s16))
; BIG: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; BIG: %full_load:_(s32) = G_OR %low_half, %high_half
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_16:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (volatile load (s16))
%high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_wrong_memop_size
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; Combine should only happen when the loads load the same size.
; LITTLE-LABEL: name: dont_wrong_memop_size
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s64) = G_CONSTANT i64 1
; LITTLE: %cst_16:_(s32) = G_CONSTANT i32 16
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; LITTLE: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; LITTLE: %wrong_size_load:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
; LITTLE: %high_half:_(s32) = nuw G_SHL %wrong_size_load, %cst_16(s32)
; LITTLE: %full_load:_(s32) = G_OR %low_half, %high_half
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_wrong_memop_size
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s64) = G_CONSTANT i64 1
; BIG: %cst_16:_(s32) = G_CONSTANT i32 16
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; BIG: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; BIG: %wrong_size_load:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
; BIG: %high_half:_(s32) = nuw G_SHL %wrong_size_load, %cst_16(s32)
; BIG: %full_load:_(s32) = G_OR %low_half, %high_half
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_16:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%wrong_size_load:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
%high_half:_(s32) = nuw G_SHL %wrong_size_load, %cst_16(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_combine_wrong_offset
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; This is not equivalent to a 32-bit load with/without a BSWAP:
;
; s16 *x = ...
; s32 y = x[0] | (x[1] << 24)
; LITTLE-LABEL: name: dont_combine_wrong_offset
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s64) = G_CONSTANT i64 1
; LITTLE: %cst_24:_(s32) = G_CONSTANT i32 24
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; LITTLE: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; LITTLE: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; LITTLE: %high_half:_(s32) = nuw G_SHL %elt1, %cst_24(s32)
; LITTLE: %full_load:_(s32) = G_OR %low_half, %high_half
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_combine_wrong_offset
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s64) = G_CONSTANT i64 1
; BIG: %cst_24:_(s32) = G_CONSTANT i32 24
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; BIG: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; BIG: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; BIG: %high_half:_(s32) = nuw G_SHL %elt1, %cst_24(s32)
; BIG: %full_load:_(s32) = G_OR %low_half, %high_half
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_24:_(s32) = G_CONSTANT i32 24
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%high_half:_(s32) = nuw G_SHL %elt1, %cst_24(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_combine_wrong_offset_2
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; This does not correspond to a 32-bit load with/without a BSWAP:
;
; s16 *x = ...
; s32 y = x[0] | (x[1] << 8)
; LITTLE-LABEL: name: dont_combine_wrong_offset_2
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s64) = G_CONSTANT i64 1
; LITTLE: %cst_8:_(s32) = G_CONSTANT i32 8
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; LITTLE: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; LITTLE: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; LITTLE: %high_half:_(s32) = nuw G_SHL %elt1, %cst_8(s32)
; LITTLE: %full_load:_(s32) = G_OR %low_half, %high_half
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_combine_wrong_offset_2
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s64) = G_CONSTANT i64 1
; BIG: %cst_8:_(s32) = G_CONSTANT i32 8
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; BIG: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; BIG: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; BIG: %high_half:_(s32) = nuw G_SHL %elt1, %cst_8(s32)
; BIG: %full_load:_(s32) = G_OR %low_half, %high_half
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_8:_(s32) = G_CONSTANT i32 8
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%high_half:_(s32) = nuw G_SHL %elt1, %cst_8(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_combine_missing_load
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; This is missing x[2], so we shouldn't combine:
;
; s16 *x = ...
; s64 y = (x[0] | (x[1] << 16)) | (x[3] << 48)
; LITTLE-LABEL: name: dont_combine_missing_load
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s64) = G_CONSTANT i64 1
; LITTLE: %cst_3:_(s64) = G_CONSTANT i64 3
; LITTLE: %cst_16:_(s64) = G_CONSTANT i64 16
; LITTLE: %cst_48:_(s64) = G_CONSTANT i64 48
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; LITTLE: %ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s64)
; LITTLE: %byte0_byte1:_(s64) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; LITTLE: %elt1:_(s64) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; LITTLE: %elt3:_(s64) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s16))
; LITTLE: %byte2_byte3:_(s64) = nuw G_SHL %elt1, %cst_16(s64)
; LITTLE: %byte6_byte7:_(s64) = nuw G_SHL %elt3, %cst_48(s64)
; LITTLE: %or1:_(s64) = G_OR %byte0_byte1, %byte2_byte3
; LITTLE: %full_load:_(s64) = G_OR %or1, %byte6_byte7
; LITTLE: $x1 = COPY %full_load(s64)
; LITTLE: RET_ReallyLR implicit $x1
; BIG-LABEL: name: dont_combine_missing_load
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s64) = G_CONSTANT i64 1
; BIG: %cst_3:_(s64) = G_CONSTANT i64 3
; BIG: %cst_16:_(s64) = G_CONSTANT i64 16
; BIG: %cst_48:_(s64) = G_CONSTANT i64 48
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; BIG: %ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s64)
; BIG: %byte0_byte1:_(s64) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; BIG: %elt1:_(s64) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; BIG: %elt3:_(s64) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s16))
; BIG: %byte2_byte3:_(s64) = nuw G_SHL %elt1, %cst_16(s64)
; BIG: %byte6_byte7:_(s64) = nuw G_SHL %elt3, %cst_48(s64)
; BIG: %or1:_(s64) = G_OR %byte0_byte1, %byte2_byte3
; BIG: %full_load:_(s64) = G_OR %or1, %byte6_byte7
; BIG: $x1 = COPY %full_load(s64)
; BIG: RET_ReallyLR implicit $x1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_3:_(s64) = G_CONSTANT i64 3
%cst_16:_(s64) = G_CONSTANT i64 16
%cst_48:_(s64) = G_CONSTANT i64 48
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s64)
%byte0_byte1:_(s64) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%elt1:_(s64) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%elt3:_(s64) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s16))
%byte2_byte3:_(s64) = nuw G_SHL %elt1, %cst_16(s64)
%byte6_byte7:_(s64) = nuw G_SHL %elt3, %cst_48(s64)
%or1:_(s64) = G_OR %byte0_byte1, %byte2_byte3
%full_load:_(s64) = G_OR %or1, %byte6_byte7
$x1 = COPY %full_load(s64)
RET_ReallyLR implicit $x1
...
---
name: dont_combine_different_addr_spaces
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; When the loads are from different address spaces, don't combine.
; LITTLE-LABEL: name: dont_combine_different_addr_spaces
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s64) = G_CONSTANT i64 1
; LITTLE: %cst_16:_(s32) = G_CONSTANT i32 16
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; LITTLE: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; LITTLE: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16), addrspace 1)
; LITTLE: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; LITTLE: %full_load:_(s32) = G_OR %low_half, %high_half
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_combine_different_addr_spaces
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s64) = G_CONSTANT i64 1
; BIG: %cst_16:_(s32) = G_CONSTANT i32 16
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; BIG: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; BIG: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16), addrspace 1)
; BIG: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; BIG: %full_load:_(s32) = G_OR %low_half, %high_half
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_16:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16), addrspace 0)
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16), addrspace 1)
%high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_combine_duplicate_idx
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; If two of the G_PTR_ADDs have the same index, then don't combine.
;
; sN *x = ...
; sM y = (x[i] << A) | (x[i] << B) ...
; LITTLE-LABEL: name: dont_combine_duplicate_idx
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s32) = G_CONSTANT i32 1
; LITTLE: %reused_idx:_(s32) = G_CONSTANT i32 2
; LITTLE: %cst_8:_(s32) = G_CONSTANT i32 8
; LITTLE: %cst_16:_(s32) = G_CONSTANT i32 16
; LITTLE: %cst_24:_(s32) = G_CONSTANT i32 24
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
; LITTLE: %uses_idx_2:_(p0) = G_PTR_ADD %ptr, %reused_idx(s32)
; LITTLE: %also_uses_idx_2:_(p0) = G_PTR_ADD %ptr, %reused_idx(s32)
; LITTLE: %byte0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
; LITTLE: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
; LITTLE: %elt2:_(s32) = G_ZEXTLOAD %uses_idx_2(p0) :: (load (s8))
; LITTLE: %elt3:_(s32) = G_ZEXTLOAD %also_uses_idx_2(p0) :: (load (s8))
; LITTLE: %byte1:_(s32) = nuw G_SHL %elt1, %cst_8(s32)
; LITTLE: %byte2:_(s32) = nuw G_SHL %elt2, %cst_16(s32)
; LITTLE: %byte3:_(s32) = nuw G_SHL %elt3, %cst_24(s32)
; LITTLE: %or1:_(s32) = G_OR %byte0, %byte1
; LITTLE: %or2:_(s32) = G_OR %byte2, %byte3
; LITTLE: %full_load:_(s32) = G_OR %or1, %or2
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_combine_duplicate_idx
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s32) = G_CONSTANT i32 1
; BIG: %reused_idx:_(s32) = G_CONSTANT i32 2
; BIG: %cst_8:_(s32) = G_CONSTANT i32 8
; BIG: %cst_16:_(s32) = G_CONSTANT i32 16
; BIG: %cst_24:_(s32) = G_CONSTANT i32 24
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
; BIG: %uses_idx_2:_(p0) = G_PTR_ADD %ptr, %reused_idx(s32)
; BIG: %also_uses_idx_2:_(p0) = G_PTR_ADD %ptr, %reused_idx(s32)
; BIG: %byte0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
; BIG: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
; BIG: %elt2:_(s32) = G_ZEXTLOAD %uses_idx_2(p0) :: (load (s8))
; BIG: %elt3:_(s32) = G_ZEXTLOAD %also_uses_idx_2(p0) :: (load (s8))
; BIG: %byte1:_(s32) = nuw G_SHL %elt1, %cst_8(s32)
; BIG: %byte2:_(s32) = nuw G_SHL %elt2, %cst_16(s32)
; BIG: %byte3:_(s32) = nuw G_SHL %elt3, %cst_24(s32)
; BIG: %or1:_(s32) = G_OR %byte0, %byte1
; BIG: %or2:_(s32) = G_OR %byte2, %byte3
; BIG: %full_load:_(s32) = G_OR %or1, %or2
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s32) = G_CONSTANT i32 1
%reused_idx:_(s32) = G_CONSTANT i32 2
%cst_8:_(s32) = G_CONSTANT i32 8
%cst_16:_(s32) = G_CONSTANT i32 16
%cst_24:_(s32) = G_CONSTANT i32 24
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
%uses_idx_2:_(p0) = G_PTR_ADD %ptr, %reused_idx(s32)
%also_uses_idx_2:_(p0) = G_PTR_ADD %ptr, %reused_idx(s32)
%byte0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
%elt2:_(s32) = G_ZEXTLOAD %uses_idx_2(p0) :: (load (s8))
%elt3:_(s32) = G_ZEXTLOAD %also_uses_idx_2(p0) :: (load (s8))
%byte1:_(s32) = nuw G_SHL %elt1, %cst_8(s32)
%byte2:_(s32) = nuw G_SHL %elt2, %cst_16(s32)
%byte3:_(s32) = nuw G_SHL %elt3, %cst_24(s32)
%or1:_(s32) = G_OR %byte0, %byte1
%or2:_(s32) = G_OR %byte2, %byte3
%full_load:_(s32) = G_OR %or1, %or2
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_combine_duplicate_offset
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; If two of the G_SHLs have the same constant, then we should not combine.
;
; sN *x = ...
; sM y = (x[i] << A) | (x[i+1] << A) ...
; LITTLE-LABEL: name: dont_combine_duplicate_offset
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s32) = G_CONSTANT i32 1
; LITTLE: %cst_2:_(s32) = G_CONSTANT i32 2
; LITTLE: %cst_3:_(s32) = G_CONSTANT i32 3
; LITTLE: %cst_8:_(s32) = G_CONSTANT i32 8
; LITTLE: %duplicate_shl_cst:_(s32) = G_CONSTANT i32 16
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
; LITTLE: %ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s32)
; LITTLE: %ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s32)
; LITTLE: %byte0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
; LITTLE: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
; LITTLE: %elt2:_(s32) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s8))
; LITTLE: %elt3:_(s32) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s8))
; LITTLE: %byte1:_(s32) = nuw G_SHL %elt1, %cst_8(s32)
; LITTLE: %duplicate_shl_1:_(s32) = nuw G_SHL %elt2, %duplicate_shl_cst(s32)
; LITTLE: %duplicate_shl_2:_(s32) = nuw G_SHL %elt3, %duplicate_shl_cst(s32)
; LITTLE: %or1:_(s32) = G_OR %byte0, %byte1
; LITTLE: %or2:_(s32) = G_OR %duplicate_shl_1, %duplicate_shl_2
; LITTLE: %full_load:_(s32) = G_OR %or1, %or2
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_combine_duplicate_offset
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s32) = G_CONSTANT i32 1
; BIG: %cst_2:_(s32) = G_CONSTANT i32 2
; BIG: %cst_3:_(s32) = G_CONSTANT i32 3
; BIG: %cst_8:_(s32) = G_CONSTANT i32 8
; BIG: %duplicate_shl_cst:_(s32) = G_CONSTANT i32 16
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
; BIG: %ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s32)
; BIG: %ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s32)
; BIG: %byte0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
; BIG: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
; BIG: %elt2:_(s32) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s8))
; BIG: %elt3:_(s32) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s8))
; BIG: %byte1:_(s32) = nuw G_SHL %elt1, %cst_8(s32)
; BIG: %duplicate_shl_1:_(s32) = nuw G_SHL %elt2, %duplicate_shl_cst(s32)
; BIG: %duplicate_shl_2:_(s32) = nuw G_SHL %elt3, %duplicate_shl_cst(s32)
; BIG: %or1:_(s32) = G_OR %byte0, %byte1
; BIG: %or2:_(s32) = G_OR %duplicate_shl_1, %duplicate_shl_2
; BIG: %full_load:_(s32) = G_OR %or1, %or2
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s32) = G_CONSTANT i32 1
%cst_2:_(s32) = G_CONSTANT i32 2
%cst_3:_(s32) = G_CONSTANT i32 3
%cst_8:_(s32) = G_CONSTANT i32 8
%duplicate_shl_cst:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
%ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s32)
%ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s32)
%byte0:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
%elt2:_(s32) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s8))
%elt3:_(s32) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s8))
%byte1:_(s32) = nuw G_SHL %elt1, %cst_8(s32)
%duplicate_shl_1:_(s32) = nuw G_SHL %elt2, %duplicate_shl_cst(s32)
%duplicate_shl_2:_(s32) = nuw G_SHL %elt3, %duplicate_shl_cst(s32)
%or1:_(s32) = G_OR %byte0, %byte1
%or2:_(s32) = G_OR %duplicate_shl_1, %duplicate_shl_2
%full_load:_(s32) = G_OR %or1, %or2
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_combine_lowest_index_not_zero_offset
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; In this case, the lowest index load (e.g. x[0]) does not end up at byte
; offset 0. We shouldn't combine.
;
; s8 *x = ...
; s32 y = (x[0] << 8) | (x[1]) | (x[2] << 16) ...
; LITTLE-LABEL: name: dont_combine_lowest_index_not_zero_offset
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s32) = G_CONSTANT i32 1
; LITTLE: %cst_2:_(s32) = G_CONSTANT i32 2
; LITTLE: %cst_3:_(s32) = G_CONSTANT i32 3
; LITTLE: %cst_8:_(s32) = G_CONSTANT i32 8
; LITTLE: %cst_16:_(s32) = G_CONSTANT i32 16
; LITTLE: %cst_24:_(s32) = G_CONSTANT i32 24
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
; LITTLE: %ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s32)
; LITTLE: %ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s32)
; LITTLE: %lowest_idx_load:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
; LITTLE: %byte0:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
; LITTLE: %elt2:_(s32) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s8))
; LITTLE: %elt3:_(s32) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s8))
; LITTLE: %byte1:_(s32) = nuw G_SHL %lowest_idx_load, %cst_8(s32)
; LITTLE: %byte2:_(s32) = nuw G_SHL %elt2, %cst_16(s32)
; LITTLE: %byte3:_(s32) = nuw G_SHL %elt3, %cst_24(s32)
; LITTLE: %or1:_(s32) = G_OR %byte0, %byte1
; LITTLE: %or2:_(s32) = G_OR %byte2, %byte3
; LITTLE: %full_load:_(s32) = G_OR %or1, %or2
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_combine_lowest_index_not_zero_offset
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s32) = G_CONSTANT i32 1
; BIG: %cst_2:_(s32) = G_CONSTANT i32 2
; BIG: %cst_3:_(s32) = G_CONSTANT i32 3
; BIG: %cst_8:_(s32) = G_CONSTANT i32 8
; BIG: %cst_16:_(s32) = G_CONSTANT i32 16
; BIG: %cst_24:_(s32) = G_CONSTANT i32 24
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
; BIG: %ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s32)
; BIG: %ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s32)
; BIG: %lowest_idx_load:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
; BIG: %byte0:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
; BIG: %elt2:_(s32) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s8))
; BIG: %elt3:_(s32) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s8))
; BIG: %byte1:_(s32) = nuw G_SHL %lowest_idx_load, %cst_8(s32)
; BIG: %byte2:_(s32) = nuw G_SHL %elt2, %cst_16(s32)
; BIG: %byte3:_(s32) = nuw G_SHL %elt3, %cst_24(s32)
; BIG: %or1:_(s32) = G_OR %byte0, %byte1
; BIG: %or2:_(s32) = G_OR %byte2, %byte3
; BIG: %full_load:_(s32) = G_OR %or1, %or2
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s32) = G_CONSTANT i32 1
%cst_2:_(s32) = G_CONSTANT i32 2
%cst_3:_(s32) = G_CONSTANT i32 3
%cst_8:_(s32) = G_CONSTANT i32 8
%cst_16:_(s32) = G_CONSTANT i32 16
%cst_24:_(s32) = G_CONSTANT i32 24
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s32)
%ptr_elt_2:_(p0) = G_PTR_ADD %ptr, %cst_2(s32)
%ptr_elt_3:_(p0) = G_PTR_ADD %ptr, %cst_3(s32)
; This load is index 0
%lowest_idx_load:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s8))
%byte0:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s8))
%elt2:_(s32) = G_ZEXTLOAD %ptr_elt_2(p0) :: (load (s8))
%elt3:_(s32) = G_ZEXTLOAD %ptr_elt_3(p0) :: (load (s8))
; ... But it ends up being shifted, so we shouldn't combine.
%byte1:_(s32) = nuw G_SHL %lowest_idx_load, %cst_8(s32)
%byte2:_(s32) = nuw G_SHL %elt2, %cst_16(s32)
%byte3:_(s32) = nuw G_SHL %elt3, %cst_24(s32)
%or1:_(s32) = G_OR %byte0, %byte1
%or2:_(s32) = G_OR %byte2, %byte3
%full_load:_(s32) = G_OR %or1, %or2
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_combine_more_than_one_use_load
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; If any load is used more than once, don't combine. We want to remove the
; entire tree.
; LITTLE-LABEL: name: dont_combine_more_than_one_use_load
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s64) = G_CONSTANT i64 1
; LITTLE: %cst_16:_(s32) = G_CONSTANT i32 16
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; LITTLE: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; LITTLE: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; LITTLE: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; LITTLE: %full_load:_(s32) = G_OR %low_half, %high_half
; LITTLE: %extra_use:_(s32) = G_AND %full_load, %low_half
; LITTLE: $w1 = COPY %extra_use(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_combine_more_than_one_use_load
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s64) = G_CONSTANT i64 1
; BIG: %cst_16:_(s32) = G_CONSTANT i32 16
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; BIG: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; BIG: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; BIG: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; BIG: %full_load:_(s32) = G_OR %low_half, %high_half
; BIG: %extra_use:_(s32) = G_AND %full_load, %low_half
; BIG: $w1 = COPY %extra_use(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_16:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
%extra_use:_(s32) = G_AND %full_load, %low_half
$w1 = COPY %extra_use(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_combine_more_than_one_use_shl
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; If anything feeding into any of the ors is used more than once, don't
; combine.
; LITTLE-LABEL: name: dont_combine_more_than_one_use_shl
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s64) = G_CONSTANT i64 1
; LITTLE: %cst_16:_(s32) = G_CONSTANT i32 16
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; LITTLE: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; LITTLE: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; LITTLE: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; LITTLE: %full_load:_(s32) = G_OR %low_half, %high_half
; LITTLE: %extra_use:_(s32) = G_AND %full_load, %high_half
; LITTLE: $w1 = COPY %extra_use(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_combine_more_than_one_use_shl
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s64) = G_CONSTANT i64 1
; BIG: %cst_16:_(s32) = G_CONSTANT i32 16
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; BIG: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; BIG: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; BIG: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; BIG: %full_load:_(s32) = G_OR %low_half, %high_half
; BIG: %extra_use:_(s32) = G_AND %full_load, %high_half
; BIG: $w1 = COPY %extra_use(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_16:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
%extra_use:_(s32) = G_AND %full_load, %high_half
$w1 = COPY %extra_use(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_combine_store_between_same_mbb
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; If there is a store between any of the loads, then do not combine.
; LITTLE-LABEL: name: dont_combine_store_between_same_mbb
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s64) = G_CONSTANT i64 1
; LITTLE: %cst_16:_(s32) = G_CONSTANT i32 16
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; LITTLE: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; LITTLE: %other_ptr:_(p0) = COPY $x1
; LITTLE: %some_val:_(s32) = G_CONSTANT i32 12
; LITTLE: G_STORE %some_val(s32), %other_ptr(p0) :: (store (s16))
; LITTLE: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; LITTLE: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; LITTLE: %full_load:_(s32) = G_OR %low_half, %high_half
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_combine_store_between_same_mbb
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s64) = G_CONSTANT i64 1
; BIG: %cst_16:_(s32) = G_CONSTANT i32 16
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; BIG: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; BIG: %other_ptr:_(p0) = COPY $x1
; BIG: %some_val:_(s32) = G_CONSTANT i32 12
; BIG: G_STORE %some_val(s32), %other_ptr(p0) :: (store (s16))
; BIG: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; BIG: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; BIG: %full_load:_(s32) = G_OR %low_half, %high_half
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_16:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; Memory could be modified here, so don't combine!
%other_ptr:_(p0) = COPY $x1
%some_val:_(s32) = G_CONSTANT i32 12
G_STORE %some_val, %other_ptr :: (store (s16))
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: dont_combine_store_between_different_mbb
tracksRegLiveness: true
body: |
; LITTLE-LABEL: name: dont_combine_store_between_different_mbb
; LITTLE: bb.0:
; LITTLE: successors: %bb.1(0x80000000)
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s64) = G_CONSTANT i64 1
; LITTLE: %cst_16:_(s32) = G_CONSTANT i32 16
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; LITTLE: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; LITTLE: bb.1:
; LITTLE: successors: %bb.2(0x80000000)
; LITTLE: liveins: $x0, $x1
; LITTLE: %other_ptr:_(p0) = COPY $x1
; LITTLE: %some_val:_(s32) = G_CONSTANT i32 12
; LITTLE: G_STORE %some_val(s32), %other_ptr(p0) :: (store (s16))
; LITTLE: bb.2:
; LITTLE: liveins: $x0, $x1
; LITTLE: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; LITTLE: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; LITTLE: %full_load:_(s32) = G_OR %low_half, %high_half
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: dont_combine_store_between_different_mbb
; BIG: bb.0:
; BIG: successors: %bb.1(0x80000000)
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s64) = G_CONSTANT i64 1
; BIG: %cst_16:_(s32) = G_CONSTANT i32 16
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; BIG: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; BIG: bb.1:
; BIG: successors: %bb.2(0x80000000)
; BIG: liveins: $x0, $x1
; BIG: %other_ptr:_(p0) = COPY $x1
; BIG: %some_val:_(s32) = G_CONSTANT i32 12
; BIG: G_STORE %some_val(s32), %other_ptr(p0) :: (store (s16))
; BIG: bb.2:
; BIG: liveins: $x0, $x1
; BIG: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; BIG: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; BIG: %full_load:_(s32) = G_OR %low_half, %high_half
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
; There is a store between the two loads, hidden away in a different MBB.
; We should not combine here.
bb.0:
successors: %bb.1(0x80000000)
liveins: $x0, $x1
; If there is a store between any of the loads, then do not combine.
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_16:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
bb.1:
liveins: $x0, $x1
successors: %bb.2(0x80000000)
; Memory could be modified here, so don't combine!
%other_ptr:_(p0) = COPY $x1
%some_val:_(s32) = G_CONSTANT i32 12
G_STORE %some_val, %other_ptr :: (store (s16))
bb.2:
liveins: $x0, $x1
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: different_mbb
tracksRegLiveness: true
body: |
; LITTLE-LABEL: name: different_mbb
; LITTLE: bb.0:
; LITTLE: successors: %bb.1(0x80000000)
; LITTLE: liveins: $x0, $x1
; LITTLE: %cst_1:_(s64) = G_CONSTANT i64 1
; LITTLE: %cst_16:_(s32) = G_CONSTANT i32 16
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; LITTLE: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; LITTLE: bb.1:
; LITTLE: liveins: $x0, $x1
; LITTLE: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; LITTLE: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; LITTLE: %full_load:_(s32) = G_OR %low_half, %high_half
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: different_mbb
; BIG: bb.0:
; BIG: successors: %bb.1(0x80000000)
; BIG: liveins: $x0, $x1
; BIG: %cst_1:_(s64) = G_CONSTANT i64 1
; BIG: %cst_16:_(s32) = G_CONSTANT i32 16
; BIG: %ptr:_(p0) = COPY $x1
; BIG: %ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
; BIG: %low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
; BIG: bb.1:
; BIG: liveins: $x0, $x1
; BIG: %elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
; BIG: %high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
; BIG: %full_load:_(s32) = G_OR %low_half, %high_half
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
; It should be possible to combine here, but it's not supported right now.
bb.0:
successors: %bb.1(0x80000000)
liveins: $x0, $x1
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_16:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
bb.1:
liveins: $x0, $x1
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: load_first
tracksRegLiveness: true
body: |
bb.0:
liveins: $x0, $x1
; Test for a bug fix for predecessor-checking code.
; LITTLE-LABEL: name: load_first
; LITTLE: liveins: $x0, $x1
; LITTLE: %ptr:_(p0) = COPY $x1
; LITTLE: %full_load:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 2)
; LITTLE: $w1 = COPY %full_load(s32)
; LITTLE: RET_ReallyLR implicit $w1
; BIG-LABEL: name: load_first
; BIG: liveins: $x0, $x1
; BIG: %ptr:_(p0) = COPY $x1
; BIG: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD %ptr(p0) :: (load (s32), align 2)
; BIG: %full_load:_(s32) = G_BSWAP [[LOAD]]
; BIG: $w1 = COPY %full_load(s32)
; BIG: RET_ReallyLR implicit $w1
%low_half:_(s32) = G_ZEXTLOAD %ptr(p0) :: (load (s16))
%cst_1:_(s64) = G_CONSTANT i64 1
%cst_16:_(s32) = G_CONSTANT i32 16
%ptr:_(p0) = COPY $x1
%ptr_elt_1:_(p0) = G_PTR_ADD %ptr, %cst_1(s64)
%elt1:_(s32) = G_ZEXTLOAD %ptr_elt_1(p0) :: (load (s16))
%high_half:_(s32) = nuw G_SHL %elt1, %cst_16(s32)
%full_load:_(s32) = G_OR %low_half, %high_half
$w1 = COPY %full_load(s32)
RET_ReallyLR implicit $w1
...
---
name: store_between_loads_and_or
alignment: 4
tracksRegLiveness: true
liveins:
- { reg: '$x0' }
- { reg: '$x1' }
frameInfo:
maxAlignment: 1
body: |
bb.1:
liveins: $x0, $x1
; Check that we build the G_LOAD at the point of the last load, instead of place of the G_OR.
; We could have a G_STORE in between which may not be safe to move the load across.
liveins: $x0, $x1
; LITTLE-LABEL: name: store_between_loads_and_or
; LITTLE: liveins: $x0, $x1, $x0, $x1
; LITTLE: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
; LITTLE: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
; LITTLE: [[C:%[0-9]+]]:_(s8) = G_CONSTANT i8 1
; LITTLE: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p0) :: (load (s32), align 1)
; LITTLE: G_STORE [[C]](s8), [[COPY1]](p0) :: (store (s8))
; LITTLE: $w0 = COPY [[LOAD]](s32)
; LITTLE: RET_ReallyLR implicit $w0
; BIG-LABEL: name: store_between_loads_and_or
; BIG: liveins: $x0, $x1, $x0, $x1
; BIG: [[COPY:%[0-9]+]]:_(p0) = COPY $x0
; BIG: [[COPY1:%[0-9]+]]:_(p0) = COPY $x1
; BIG: [[C:%[0-9]+]]:_(s8) = G_CONSTANT i8 1
; BIG: [[LOAD:%[0-9]+]]:_(s32) = G_LOAD [[COPY]](p0) :: (load (s32), align 1)
; BIG: [[BSWAP:%[0-9]+]]:_(s32) = G_BSWAP [[LOAD]]
; BIG: G_STORE [[C]](s8), [[COPY1]](p0) :: (store (s8))
; BIG: $w0 = COPY [[BSWAP]](s32)
; BIG: RET_ReallyLR implicit $w0
%0:_(p0) = COPY $x0
%1:_(p0) = COPY $x1
%12:_(s8) = G_CONSTANT i8 1
%15:_(s32) = G_CONSTANT i32 8
%19:_(s32) = G_CONSTANT i32 16
%23:_(s32) = G_CONSTANT i32 24
%13:_(s32) = G_ZEXTLOAD %0:_(p0) :: (load (s8))
%3:_(s64) = G_CONSTANT i64 1
%4:_(p0) = G_PTR_ADD %0:_, %3:_(s64)
%14:_(s32) = G_ZEXTLOAD %4:_(p0) :: (load (s8))
%6:_(s64) = G_CONSTANT i64 2
%7:_(p0) = G_PTR_ADD %0:_, %6:_(s64)
%18:_(s32) = G_ZEXTLOAD %7:_(p0) :: (load (s8))
%9:_(s64) = G_CONSTANT i64 3
%10:_(p0) = G_PTR_ADD %0:_, %9:_(s64)
%22:_(s32) = G_ZEXTLOAD %10:_(p0) :: (load (s8))
G_STORE %12:_(s8), %1:_(p0) :: (store (s8))
%16:_(s32) = nuw nsw G_SHL %14:_, %15:_(s32)
%17:_(s32) = G_OR %16:_, %13:_
%20:_(s32) = nuw nsw G_SHL %18:_, %19:_(s32)
%21:_(s32) = G_OR %17:_, %20:_
%24:_(s32) = nuw G_SHL %22:_, %23:_(s32)
%25:_(s32) = G_OR %21:_, %24:_
$w0 = COPY %25:_(s32)
RET_ReallyLR implicit $w0
...