172 lines · plain
1; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 52; RUN: llc -fixup-byte-word-insts=1 < %s | FileCheck %s -check-prefix=BWON3; RUN: llc -fixup-byte-word-insts=0 < %s | FileCheck %s -check-prefix=BWOFF4 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 single11; basic-block loop. The upper portion should be dead, so the movb loads12; 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, %edi17; BWON-NEXT: jle LBB0_218; BWON-NEXT: .p2align 419; BWON-NEXT: LBB0_1: ## %a420; BWON-NEXT: ## =>This Inner Loop Header: Depth=121; BWON-NEXT: movzbl (%rsi), %eax22; BWON-NEXT: movb %al, (%rdx)23; BWON-NEXT: movzbl 1(%rsi), %eax24; BWON-NEXT: movb %al, 1(%rdx)25; BWON-NEXT: addq $8, %rdx26; BWON-NEXT: decl %edi27; BWON-NEXT: jne LBB0_128; BWON-NEXT: LBB0_2: ## %._crit_edge29; BWON-NEXT: retq30;31; BWOFF-LABEL: foo1:32; BWOFF: ## %bb.0:33; BWOFF-NEXT: testl %edi, %edi34; BWOFF-NEXT: jle LBB0_235; BWOFF-NEXT: .p2align 436; BWOFF-NEXT: LBB0_1: ## %a437; BWOFF-NEXT: ## =>This Inner Loop Header: Depth=138; BWOFF-NEXT: movb (%rsi), %al39; BWOFF-NEXT: movb %al, (%rdx)40; BWOFF-NEXT: movb 1(%rsi), %al41; BWOFF-NEXT: movb %al, 1(%rdx)42; BWOFF-NEXT: addq $8, %rdx43; BWOFF-NEXT: decl %edi44; BWOFF-NEXT: jne LBB0_145; BWOFF-NEXT: LBB0_2: ## %._crit_edge46; BWOFF-NEXT: retq47 %1 = icmp sgt i32 %count, 048 br i1 %1, label %.lr.ph, label %._crit_edge49 50.lr.ph: ; preds = %051 %2 = getelementptr inbounds %struct.A, ptr %q, i64 0, i32 152 br label %a453 54a4: ; preds = %3, %.lr.ph55 %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 158 store i8 %a5, ptr %.01, align 159 %a8 = getelementptr inbounds %struct.A, ptr %.01, i64 0, i32 160 %a6 = load i8, ptr %2, align 161 store i8 %a6, ptr %a8, align 162 %a9 = add nsw i32 %i.02, 163 %a10 = getelementptr inbounds %struct.A, ptr %.01, i64 164 %exitcond = icmp eq i32 %a9, %count65 br i1 %exitcond, label %._crit_edge, label %a466 67._crit_edge: ; preds = %3, %068 ret void69}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 have75; 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, %edi80; BWON-NEXT: jle LBB1_281; BWON-NEXT: .p2align 482; BWON-NEXT: LBB1_1: ## %a483; BWON-NEXT: ## =>This Inner Loop Header: Depth=184; BWON-NEXT: movzwl (%rsi), %eax85; BWON-NEXT: movw %ax, (%rdx)86; BWON-NEXT: movzwl 2(%rsi), %eax87; BWON-NEXT: movw %ax, 2(%rdx)88; BWON-NEXT: addq $16, %rdx89; BWON-NEXT: decl %edi90; BWON-NEXT: jne LBB1_191; BWON-NEXT: LBB1_2: ## %._crit_edge92; BWON-NEXT: retq93;94; BWOFF-LABEL: foo2:95; BWOFF: ## %bb.0:96; BWOFF-NEXT: testl %edi, %edi97; BWOFF-NEXT: jle LBB1_298; BWOFF-NEXT: .p2align 499; BWOFF-NEXT: LBB1_1: ## %a4100; BWOFF-NEXT: ## =>This Inner Loop Header: Depth=1101; BWOFF-NEXT: movw (%rsi), %ax102; BWOFF-NEXT: movw %ax, (%rdx)103; BWOFF-NEXT: movw 2(%rsi), %ax104; BWOFF-NEXT: movw %ax, 2(%rdx)105; BWOFF-NEXT: addq $16, %rdx106; BWOFF-NEXT: decl %edi107; BWOFF-NEXT: jne LBB1_1108; BWOFF-NEXT: LBB1_2: ## %._crit_edge109; BWOFF-NEXT: retq110 %1 = icmp sgt i32 %count, 0111 br i1 %1, label %.lr.ph, label %._crit_edge112 113.lr.ph: ; preds = %0114 %2 = getelementptr inbounds %struct.B, ptr %q, i64 0, i32 1115 br label %a4116 117a4: ; preds = %3, %.lr.ph118 %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 2121 store i16 %a5, ptr %.01, align 2122 %a8 = getelementptr inbounds %struct.B, ptr %.01, i64 0, i32 1123 %a6 = load i16, ptr %2, align 2124 store i16 %a6, ptr %a8, align 2125 %a9 = add nsw i32 %i.02, 1126 %a10 = getelementptr inbounds %struct.B, ptr %.01, i64 1127 %exitcond = icmp eq i32 %a9, %count128 br i1 %exitcond, label %._crit_edge, label %a4129 130._crit_edge: ; preds = %3, %0131 ret void132}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), %eax140; BWON-NEXT: movb %al, (%rdi)141; BWON-NEXT: retq142;143; BWOFF-LABEL: foo3:144; BWOFF: ## %bb.0:145; BWOFF-NEXT: movb (%rsi), %al146; BWOFF-NEXT: movb %al, (%rdi)147; BWOFF-NEXT: retq148 %t0 = load i8, ptr%src, align 1149 store i8 %t0, ptr%dst, align 1150 ret void151}152 153; This test contains nothing but a simple word load and store. Since154; movw and movzwl are the same size, we should always choose to use155; movzwl instead.156define void @foo4(ptr%dst, ptr%src) {157; BWON-LABEL: foo4:158; BWON: ## %bb.0:159; BWON-NEXT: movzwl (%rsi), %eax160; BWON-NEXT: movw %ax, (%rdi)161; BWON-NEXT: retq162;163; BWOFF-LABEL: foo4:164; BWOFF: ## %bb.0:165; BWOFF-NEXT: movw (%rsi), %ax166; BWOFF-NEXT: movw %ax, (%rdi)167; BWOFF-NEXT: retq168 %t0 = load i16, ptr%src, align 2169 store i16 %t0, ptr%dst, align 2170 ret void171}172