104 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py2; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefix=CHECK3 4; fold (sra (load i32), 16)) -> (sextload i16)5define i32 @sra_half(ptr %p) {6; CHECK-LABEL: sra_half:7; CHECK: # %bb.0:8; CHECK-NEXT: movswl 2(%rdi), %eax9; CHECK-NEXT: retq10 %load = load i32, ptr %p11 %shift = ashr i32 %load, 1612 ret i32 %shift13}14 15; Vector version not folded.16define <4 x i32> @sra_half_vec(ptr %p) {17; CHECK-LABEL: sra_half_vec:18; CHECK: # %bb.0:19; CHECK-NEXT: movdqa (%rdi), %xmm020; CHECK-NEXT: psrad $16, %xmm021; CHECK-NEXT: retq22 %load = load <4 x i32>, ptr %p23 %shift = ashr <4 x i32> %load, <i32 16, i32 16, i32 16, i32 16>24 ret <4 x i32> %shift25}26 27; fold (sra (load i64), 48)) -> (sextload i16)28define i64 @sra_large_shift(ptr %r) {29; CHECK-LABEL: sra_large_shift:30; CHECK: # %bb.0:31; CHECK-NEXT: movswq 6(%rdi), %rax32; CHECK-NEXT: retq33 %t0 = load i64, ptr %r34 %conv = ashr i64 %t0, 4835 ret i64 %conv36}37 38; Negative test, no fold expected.39define i32 @sra_small_shift(ptr %p) {40; CHECK-LABEL: sra_small_shift:41; CHECK: # %bb.0:42; CHECK-NEXT: movl (%rdi), %eax43; CHECK-NEXT: sarl $8, %eax44; CHECK-NEXT: retq45 %load = load i32, ptr %p46 %shift = ashr i32 %load, 847 ret i32 %shift48}49 50; This should be folded to a zextload.51define i32 @sra_of_zextload(ptr %p) {52; CHECK-LABEL: sra_of_zextload:53; CHECK: # %bb.0:54; CHECK-NEXT: movzbl 1(%rdi), %eax55; CHECK-NEXT: retq56 %load = load i16, ptr %p57 %zext = zext i16 %load to i3258 %shift = ashr i32 %zext, 859 ret i32 %shift60}61 62; fold (sra (sextload i16 to i32), 8) -> (sextload i8)63define i32 @sra_of_sextload(ptr %p) {64; CHECK-LABEL: sra_of_sextload:65; CHECK: # %bb.0:66; CHECK-NEXT: movsbl 1(%rdi), %eax67; CHECK-NEXT: retq68 %load = load i16, ptr %p69 %sext = sext i16 %load to i3270 %shift = ashr i32 %sext, 871 ret i32 %shift72}73 74; Negative test. If the shift amount is larger than the memory type then75; we're not accessing any of the loaded bytes (only the extended bits). So the76; shift is expected to remain.77define i32 @sra_of_sextload_no_fold(ptr %p) {78; CHECK-LABEL: sra_of_sextload_no_fold:79; CHECK: # %bb.0:80; CHECK-NEXT: movswl (%rdi), %eax81; CHECK-NEXT: sarl $16, %eax82; CHECK-NEXT: retq83 %load = load i16, ptr %p84 %sext = sext i16 %load to i3285 %shift = ashr i32 %sext, 1686 ret i32 %shift87}88 89; Fold even if SRA has multiple uses.90define i32 @sra_to_sextload_multiple_sra_uses(ptr %p) {91; CHECK-LABEL: sra_to_sextload_multiple_sra_uses:92; CHECK: # %bb.0:93; CHECK-NEXT: movswl 2(%rdi), %ecx94; CHECK-NEXT: movl %ecx, %eax95; CHECK-NEXT: xorl $6, %eax96; CHECK-NEXT: imull %ecx, %eax97; CHECK-NEXT: retq98 %load = load i32, ptr %p99 %shift = ashr i32 %load, 16100 %use1 = xor i32 %shift, 6101 %use2 = mul i32 %shift, %use1102 ret i32 %use2103}104