1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5 2; RUN: llc -fixup-byte-word-insts=1 < %s | FileCheck %s -check-prefix=BWON 3; RUN: llc -fixup-byte-word-insts=0 < %s | FileCheck %s -check-prefix=BWOFF 4 5target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128" 6target triple = "x86_64-apple-macosx10.8.0" 7 8%struct.A = type { i8, i8, i8, i8, i8, i8, i8, i8 } 9 10; This has byte loads interspersed with byte stores, in a single 11; basic-block loop. The upper portion should be dead, so the movb loads 12; should have been changed into movzbl instead. 13define void @foo1(i32 %count, ptr noalias nocapture %q, ptr noalias nocapture %p) nounwind uwtable noinline ssp { 14; BWON-LABEL: foo1: 15; BWON: ## %bb.0: 16; BWON-NEXT: testl %edi, %edi 17; BWON-NEXT: jle LBB0_2 18; BWON-NEXT: .p2align 4 19; BWON-NEXT: LBB0_1: ## %a4 20; BWON-NEXT: ## =>This Inner Loop Header: Depth=1 21; BWON-NEXT: movzbl (%rsi), %eax 22; BWON-NEXT: movb %al, (%rdx) 23; BWON-NEXT: movzbl 1(%rsi), %eax 24; BWON-NEXT: movb %al, 1(%rdx) 25; BWON-NEXT: addq $8, %rdx 26; BWON-NEXT: decl %edi 27; BWON-NEXT: jne LBB0_1 28; BWON-NEXT: LBB0_2: ## %._crit_edge 29; BWON-NEXT: retq 30; 31; BWOFF-LABEL: foo1: 32; BWOFF: ## %bb.0: 33; BWOFF-NEXT: testl %edi, %edi 34; BWOFF-NEXT: jle LBB0_2 35; BWOFF-NEXT: .p2align 4 36; BWOFF-NEXT: LBB0_1: ## %a4 37; BWOFF-NEXT: ## =>This Inner Loop Header: Depth=1 38; BWOFF-NEXT: movb (%rsi), %al 39; BWOFF-NEXT: movb %al, (%rdx) 40; BWOFF-NEXT: movb 1(%rsi), %al 41; BWOFF-NEXT: movb %al, 1(%rdx) 42; BWOFF-NEXT: addq $8, %rdx 43; BWOFF-NEXT: decl %edi 44; BWOFF-NEXT: jne LBB0_1 45; BWOFF-NEXT: LBB0_2: ## %._crit_edge 46; BWOFF-NEXT: retq 47 %1 = icmp sgt i32 %count, 0 48 br i1 %1, label %.lr.ph, label %._crit_edge 49 50.lr.ph: ; preds = %0 51 %2 = getelementptr inbounds %struct.A, ptr %q, i64 0, i32 1 52 br label %a4 53 54a4: ; preds = %3, %.lr.ph 55 %i.02 = phi i32 [ 0, %.lr.ph ], [ %a9, %a4 ] 56 %.01 = phi ptr [ %p, %.lr.ph ], [ %a10, %a4 ] 57 %a5 = load i8, ptr %q, align 1 58 store i8 %a5, ptr %.01, align 1 59 %a8 = getelementptr inbounds %struct.A, ptr %.01, i64 0, i32 1 60 %a6 = load i8, ptr %2, align 1 61 store i8 %a6, ptr %a8, align 1 62 %a9 = add nsw i32 %i.02, 1 63 %a10 = getelementptr inbounds %struct.A, ptr %.01, i64 1 64 %exitcond = icmp eq i32 %a9, %count 65 br i1 %exitcond, label %._crit_edge, label %a4 66 67._crit_edge: ; preds = %3, %0 68 ret void 69} 70 71%struct.B = type { i16, i16, i16, i16, i16, i16, i16, i16 } 72 73; This has word loads interspersed with word stores. 74; The upper portion should be dead, so the movw loads should have 75; been changed into movzwl instead. 76define void @foo2(i32 %count, ptr noalias nocapture %q, ptr noalias nocapture %p) nounwind uwtable noinline ssp { 77; BWON-LABEL: foo2: 78; BWON: ## %bb.0: 79; BWON-NEXT: testl %edi, %edi 80; BWON-NEXT: jle LBB1_2 81; BWON-NEXT: .p2align 4 82; BWON-NEXT: LBB1_1: ## %a4 83; BWON-NEXT: ## =>This Inner Loop Header: Depth=1 84; BWON-NEXT: movzwl (%rsi), %eax 85; BWON-NEXT: movw %ax, (%rdx) 86; BWON-NEXT: movzwl 2(%rsi), %eax 87; BWON-NEXT: movw %ax, 2(%rdx) 88; BWON-NEXT: addq $16, %rdx 89; BWON-NEXT: decl %edi 90; BWON-NEXT: jne LBB1_1 91; BWON-NEXT: LBB1_2: ## %._crit_edge 92; BWON-NEXT: retq 93; 94; BWOFF-LABEL: foo2: 95; BWOFF: ## %bb.0: 96; BWOFF-NEXT: testl %edi, %edi 97; BWOFF-NEXT: jle LBB1_2 98; BWOFF-NEXT: .p2align 4 99; BWOFF-NEXT: LBB1_1: ## %a4 100; BWOFF-NEXT: ## =>This Inner Loop Header: Depth=1 101; BWOFF-NEXT: movw (%rsi), %ax 102; BWOFF-NEXT: movw %ax, (%rdx) 103; BWOFF-NEXT: movw 2(%rsi), %ax 104; BWOFF-NEXT: movw %ax, 2(%rdx) 105; BWOFF-NEXT: addq $16, %rdx 106; BWOFF-NEXT: decl %edi 107; BWOFF-NEXT: jne LBB1_1 108; BWOFF-NEXT: LBB1_2: ## %._crit_edge 109; BWOFF-NEXT: retq 110 %1 = icmp sgt i32 %count, 0 111 br i1 %1, label %.lr.ph, label %._crit_edge 112 113.lr.ph: ; preds = %0 114 %2 = getelementptr inbounds %struct.B, ptr %q, i64 0, i32 1 115 br label %a4 116 117a4: ; preds = %3, %.lr.ph 118 %i.02 = phi i32 [ 0, %.lr.ph ], [ %a9, %a4 ] 119 %.01 = phi ptr [ %p, %.lr.ph ], [ %a10, %a4 ] 120 %a5 = load i16, ptr %q, align 2 121 store i16 %a5, ptr %.01, align 2 122 %a8 = getelementptr inbounds %struct.B, ptr %.01, i64 0, i32 1 123 %a6 = load i16, ptr %2, align 2 124 store i16 %a6, ptr %a8, align 2 125 %a9 = add nsw i32 %i.02, 1 126 %a10 = getelementptr inbounds %struct.B, ptr %.01, i64 1 127 %exitcond = icmp eq i32 %a9, %count 128 br i1 %exitcond, label %._crit_edge, label %a4 129 130._crit_edge: ; preds = %3, %0 131 ret void 132} 133 134; This test contains nothing but a simple byte load and store. 135; movb encodes smaller, but we use movzbl for the load for better perf. 136define void @foo3(ptr%dst, ptr%src) { 137; BWON-LABEL: foo3: 138; BWON: ## %bb.0: 139; BWON-NEXT: movzbl (%rsi), %eax 140; BWON-NEXT: movb %al, (%rdi) 141; BWON-NEXT: retq 142; 143; BWOFF-LABEL: foo3: 144; BWOFF: ## %bb.0: 145; BWOFF-NEXT: movb (%rsi), %al 146; BWOFF-NEXT: movb %al, (%rdi) 147; BWOFF-NEXT: retq 148 %t0 = load i8, ptr%src, align 1 149 store i8 %t0, ptr%dst, align 1 150 ret void 151} 152 153; This test contains nothing but a simple word load and store. Since 154; movw and movzwl are the same size, we should always choose to use 155; movzwl instead. 156define void @foo4(ptr%dst, ptr%src) { 157; BWON-LABEL: foo4: 158; BWON: ## %bb.0: 159; BWON-NEXT: movzwl (%rsi), %eax 160; BWON-NEXT: movw %ax, (%rdi) 161; BWON-NEXT: retq 162; 163; BWOFF-LABEL: foo4: 164; BWOFF: ## %bb.0: 165; BWOFF-NEXT: movw (%rsi), %ax 166; BWOFF-NEXT: movw %ax, (%rdi) 167; BWOFF-NEXT: retq 168 %t0 = load i16, ptr%src, align 2 169 store i16 %t0, ptr%dst, align 2 170 ret void 171} 172