diff options
Diffstat (limited to 'llvm/test')
| -rw-r--r-- | llvm/test/CodeGen/SPARC/atomics.ll | 132 | ||||
| -rw-r--r-- | llvm/test/Transforms/AtomicExpand/SPARC/partword.ll | 166 |
2 files changed, 296 insertions, 2 deletions
diff --git a/llvm/test/CodeGen/SPARC/atomics.ll b/llvm/test/CodeGen/SPARC/atomics.ll index 6fe333a4be8..5e608e728c3 100644 --- a/llvm/test/CodeGen/SPARC/atomics.ll +++ b/llvm/test/CodeGen/SPARC/atomics.ll @@ -64,6 +64,90 @@ entry: ret i64 %2 } +;; TODO: the "move %icc" and related instructions are totally +;; redundant here. There's something weird happening in optimization +;; of the success value of cmpxchg. + +; CHECK-LABEL: test_cmpxchg_i8 +; CHECK: and %o1, -4, %o2 +; CHECK: mov 3, %o3 +; CHECK: andn %o3, %o1, %o1 +; CHECK: sll %o1, 3, %o1 +; CHECK: mov 255, %o3 +; CHECK: sll %o3, %o1, %o5 +; CHECK: xor %o5, -1, %o3 +; CHECK: mov 123, %o4 +; CHECK: ld [%o2], %g2 +; CHECK: sll %o4, %o1, %o4 +; CHECK: and %o0, 255, %o0 +; CHECK: sll %o0, %o1, %o0 +; CHECK: andn %g2, %o5, %g2 +; CHECK: sethi 0, %o5 +; CHECK: [[LABEL1:\.L.*]]: +; CHECK: or %g2, %o4, %g3 +; CHECK: or %g2, %o0, %g4 +; CHECK: cas [%o2], %g4, %g3 +; CHECK: cmp %g3, %g4 +; CHECK: mov %o5, %g4 +; CHECK: move %icc, 1, %g4 +; CHECK: cmp %g4, 0 +; CHECK: bne [[LABEL2:\.L.*]] +; CHECK: nop +; CHECK: and %g3, %o3, %g4 +; CHECK: cmp %g2, %g4 +; CHECK: bne [[LABEL1]] +; CHECK: mov %g4, %g2 +; CHECK: [[LABEL2]]: +; CHECK: retl +; CHECK: srl %g3, %o1, %o0 +define i8 @test_cmpxchg_i8(i8 %a, i8* %ptr) { +entry: + %pair = cmpxchg i8* %ptr, i8 %a, i8 123 monotonic monotonic + %b = extractvalue { i8, i1 } %pair, 0 + ret i8 %b +} + +; CHECK-LABEL: test_cmpxchg_i16 + +; CHECK: and %o1, -4, %o2 +; CHECK: and %o1, 3, %o1 +; CHECK: xor %o1, 2, %o1 +; CHECK: sll %o1, 3, %o1 +; CHECK: sethi 63, %o3 +; CHECK: or %o3, 1023, %o4 +; CHECK: sll %o4, %o1, %o5 +; CHECK: xor %o5, -1, %o3 +; CHECK: and %o0, %o4, %o4 +; CHECK: ld [%o2], %g2 +; CHECK: mov 123, %o0 +; CHECK: sll %o0, %o1, %o0 +; CHECK: sll %o4, %o1, %o4 +; CHECK: andn %g2, %o5, %g2 +; CHECK: sethi 0, %o5 +; CHECK: [[LABEL1:\.L.*]]: +; CHECK: or %g2, %o0, %g3 +; CHECK: or %g2, %o4, %g4 +; CHECK: cas [%o2], %g4, %g3 +; CHECK: cmp %g3, %g4 +; CHECK: mov %o5, %g4 +; CHECK: move %icc, 1, %g4 +; CHECK: cmp %g4, 0 +; CHECK: bne [[LABEL2:\.L.*]] +; CHECK: nop +; CHECK: and %g3, %o3, %g4 +; CHECK: cmp %g2, %g4 +; CHECK: bne [[LABEL1]] +; CHECK: mov %g4, %g2 +; CHECK: [[LABEL2]]: +; CHECK: retl +; CHECK: srl %g3, %o1, %o0 +define i16 @test_cmpxchg_i16(i16 %a, i16* %ptr) { +entry: + %pair = cmpxchg i16* %ptr, i16 %a, i16 123 monotonic monotonic + %b = extractvalue { i16, i1 } %pair, 0 + ret i16 %b +} + ; CHECK-LABEL: test_cmpxchg_i32 ; CHECK: mov 123, [[R:%[gilo][0-7]]] ; CHECK: cas [%o1], %o0, [[R]] @@ -86,6 +170,26 @@ entry: ret i64 %b } +; CHECK-LABEL: test_swap_i8 +; CHECK: mov 42, [[R:%[gilo][0-7]]] +; CHECK: cas + +define i8 @test_swap_i8(i8 %a, i8* %ptr) { +entry: + %b = atomicrmw xchg i8* %ptr, i8 42 monotonic + ret i8 %b +} + +; CHECK-LABEL: test_swap_i16 +; CHECK: mov 42, [[R:%[gilo][0-7]]] +; CHECK: cas + +define i16 @test_swap_i16(i16 %a, i16* %ptr) { +entry: + %b = atomicrmw xchg i16* %ptr, i16 42 monotonic + ret i16 %b +} + ; CHECK-LABEL: test_swap_i32 ; CHECK: mov 42, [[R:%[gilo][0-7]]] ; CHECK: swap [%o1], [[R]] @@ -105,12 +209,36 @@ entry: ret i64 %b } -; CHECK-LABEL: test_load_add_32 +; CHECK-LABEL: test_load_sub_i8 +; CHECK: membar +; CHECK: .L{{.*}}: +; CHECK: sub +; CHECK: cas [{{%[gilo][0-7]}}] +; CHECK: membar +define zeroext i8 @test_load_sub_i8(i8* %p, i8 zeroext %v) { +entry: + %0 = atomicrmw sub i8* %p, i8 %v seq_cst + ret i8 %0 +} + +; CHECK-LABEL: test_load_sub_i16 +; CHECK: membar +; CHECK: .L{{.*}}: +; CHECK: sub +; CHECK: cas [{{%[gilo][0-7]}}] +; CHECK: membar +define zeroext i16 @test_load_sub_i16(i16* %p, i16 zeroext %v) { +entry: + %0 = atomicrmw sub i16* %p, i16 %v seq_cst + ret i16 %0 +} + +; CHECK-LABEL: test_load_add_i32 ; CHECK: membar ; CHECK: add [[V:%[gilo][0-7]]], %o1, [[U:%[gilo][0-7]]] ; CHECK: cas [%o0], [[V]], [[U]] ; CHECK: membar -define zeroext i32 @test_load_add_32(i32* %p, i32 zeroext %v) { +define zeroext i32 @test_load_add_i32(i32* %p, i32 zeroext %v) { entry: %0 = atomicrmw add i32* %p, i32 %v seq_cst ret i32 %0 diff --git a/llvm/test/Transforms/AtomicExpand/SPARC/partword.ll b/llvm/test/Transforms/AtomicExpand/SPARC/partword.ll new file mode 100644 index 00000000000..9963d17c242 --- /dev/null +++ b/llvm/test/Transforms/AtomicExpand/SPARC/partword.ll @@ -0,0 +1,166 @@ +; RUN: opt -S %s -atomic-expand | FileCheck %s + +;; Verify the cmpxchg and atomicrmw expansions where sub-word-size +;; instructions are not available. + +;;; NOTE: this test is mostly target-independent -- any target which +;;; doesn't support cmpxchg of sub-word sizes would do. +target datalayout = "E-m:e-i64:64-n32:64-S128" +target triple = "sparcv9-unknown-unknown" + +; CHECK-LABEL: @test_cmpxchg_i8( +; CHECK: fence seq_cst +; CHECK: %0 = ptrtoint i8* %arg to i64 +; CHECK: %1 = and i64 %0, -4 +; CHECK: %AlignedAddr = inttoptr i64 %1 to i32* +; CHECK: %PtrLSB = and i64 %0, 3 +; CHECK: %2 = xor i64 %PtrLSB, 3 +; CHECK: %3 = shl i64 %2, 3 +; CHECK: %ShiftAmt = trunc i64 %3 to i32 +; CHECK: %Mask = shl i32 255, %ShiftAmt +; CHECK: %Inv_Mask = xor i32 %Mask, -1 +; CHECK: %4 = zext i8 %new to i32 +; CHECK: %5 = shl i32 %4, %ShiftAmt +; CHECK: %6 = zext i8 %old to i32 +; CHECK: %7 = shl i32 %6, %ShiftAmt +; CHECK: %8 = load i32, i32* %AlignedAddr +; CHECK: %9 = and i32 %8, %Inv_Mask +; CHECK: br label %partword.cmpxchg.loop +; CHECK:partword.cmpxchg.loop: +; CHECK: %10 = phi i32 [ %9, %entry ], [ %16, %partword.cmpxchg.failure ] +; CHECK: %11 = or i32 %10, %5 +; CHECK: %12 = or i32 %10, %7 +; CHECK: %13 = cmpxchg i32* %AlignedAddr, i32 %12, i32 %11 monotonic monotonic +; CHECK: %14 = extractvalue { i32, i1 } %13, 0 +; CHECK: %15 = extractvalue { i32, i1 } %13, 1 +; CHECK: br i1 %15, label %partword.cmpxchg.end, label %partword.cmpxchg.failure +; CHECK:partword.cmpxchg.failure: +; CHECK: %16 = and i32 %14, %Inv_Mask +; CHECK: %17 = icmp ne i32 %10, %16 +; CHECK: br i1 %17, label %partword.cmpxchg.loop, label %partword.cmpxchg.end +; CHECK:partword.cmpxchg.end: +; CHECK: %18 = lshr i32 %14, %ShiftAmt +; CHECK: %19 = trunc i32 %18 to i8 +; CHECK: %20 = insertvalue { i8, i1 } undef, i8 %19, 0 +; CHECK: %21 = insertvalue { i8, i1 } %20, i1 %15, 1 +; CHECK: fence seq_cst +; CHECK: %ret = extractvalue { i8, i1 } %21, 0 +; CHECK: ret i8 %ret +define i8 @test_cmpxchg_i8(i8* %arg, i8 %old, i8 %new) { +entry: + %ret_succ = cmpxchg i8* %arg, i8 %old, i8 %new seq_cst monotonic + %ret = extractvalue { i8, i1 } %ret_succ, 0 + ret i8 %ret +} + +; CHECK-LABEL: @test_cmpxchg_i16( +; CHECK: fence seq_cst +; CHECK: %0 = ptrtoint i16* %arg to i64 +; CHECK: %1 = and i64 %0, -4 +; CHECK: %AlignedAddr = inttoptr i64 %1 to i32* +; CHECK: %PtrLSB = and i64 %0, 3 +; CHECK: %2 = xor i64 %PtrLSB, 2 +; CHECK: %3 = shl i64 %2, 3 +; CHECK: %ShiftAmt = trunc i64 %3 to i32 +; CHECK: %Mask = shl i32 65535, %ShiftAmt +; CHECK: %Inv_Mask = xor i32 %Mask, -1 +; CHECK: %4 = zext i16 %new to i32 +; CHECK: %5 = shl i32 %4, %ShiftAmt +; CHECK: %6 = zext i16 %old to i32 +; CHECK: %7 = shl i32 %6, %ShiftAmt +; CHECK: %8 = load i32, i32* %AlignedAddr +; CHECK: %9 = and i32 %8, %Inv_Mask +; CHECK: br label %partword.cmpxchg.loop +; CHECK:partword.cmpxchg.loop: +; CHECK: %10 = phi i32 [ %9, %entry ], [ %16, %partword.cmpxchg.failure ] +; CHECK: %11 = or i32 %10, %5 +; CHECK: %12 = or i32 %10, %7 +; CHECK: %13 = cmpxchg i32* %AlignedAddr, i32 %12, i32 %11 monotonic monotonic +; CHECK: %14 = extractvalue { i32, i1 } %13, 0 +; CHECK: %15 = extractvalue { i32, i1 } %13, 1 +; CHECK: br i1 %15, label %partword.cmpxchg.end, label %partword.cmpxchg.failure +; CHECK:partword.cmpxchg.failure: +; CHECK: %16 = and i32 %14, %Inv_Mask +; CHECK: %17 = icmp ne i32 %10, %16 +; CHECK: br i1 %17, label %partword.cmpxchg.loop, label %partword.cmpxchg.end +; CHECK:partword.cmpxchg.end: +; CHECK: %18 = lshr i32 %14, %ShiftAmt +; CHECK: %19 = trunc i32 %18 to i16 +; CHECK: %20 = insertvalue { i16, i1 } undef, i16 %19, 0 +; CHECK: %21 = insertvalue { i16, i1 } %20, i1 %15, 1 +; CHECK: fence seq_cst +; CHECK: %ret = extractvalue { i16, i1 } %21, 0 +; CHECK: ret i16 %ret +define i16 @test_cmpxchg_i16(i16* %arg, i16 %old, i16 %new) { +entry: + %ret_succ = cmpxchg i16* %arg, i16 %old, i16 %new seq_cst monotonic + %ret = extractvalue { i16, i1 } %ret_succ, 0 + ret i16 %ret +} + + +; CHECK-LABEL: @test_add_i16( +; CHECK: fence seq_cst +; CHECK: %0 = ptrtoint i16* %arg to i64 +; CHECK: %1 = and i64 %0, -4 +; CHECK: %AlignedAddr = inttoptr i64 %1 to i32* +; CHECK: %PtrLSB = and i64 %0, 3 +; CHECK: %2 = xor i64 %PtrLSB, 2 +; CHECK: %3 = shl i64 %2, 3 +; CHECK: %ShiftAmt = trunc i64 %3 to i32 +; CHECK: %Mask = shl i32 65535, %ShiftAmt +; CHECK: %Inv_Mask = xor i32 %Mask, -1 +; CHECK: %4 = zext i16 %val to i32 +; CHECK: %ValOperand_Shifted = shl i32 %4, %ShiftAmt +; CHECK: %5 = load i32, i32* %AlignedAddr, align 4 +; CHECK: br label %atomicrmw.start +; CHECK:atomicrmw.start: +; CHECK: %loaded = phi i32 [ %5, %entry ], [ %newloaded, %atomicrmw.start ] +; CHECK: %new = add i32 %loaded, %ValOperand_Shifted +; CHECK: %6 = and i32 %new, %Mask +; CHECK: %7 = and i32 %loaded, %Inv_Mask +; CHECK: %8 = or i32 %7, %6 +; CHECK: %9 = cmpxchg i32* %AlignedAddr, i32 %loaded, i32 %8 monotonic monotonic +; CHECK: %success = extractvalue { i32, i1 } %9, 1 +; CHECK: %newloaded = extractvalue { i32, i1 } %9, 0 +; CHECK: br i1 %success, label %atomicrmw.end, label %atomicrmw.start +; CHECK:atomicrmw.end: +; CHECK: %10 = lshr i32 %newloaded, %ShiftAmt +; CHECK: %11 = trunc i32 %10 to i16 +; CHECK: fence seq_cst +; CHECK: ret i16 %11 +define i16 @test_add_i16(i16* %arg, i16 %val) { +entry: + %ret = atomicrmw add i16* %arg, i16 %val seq_cst + ret i16 %ret +} + +; CHECK-LABEL: @test_xor_i16( +; (I'm going to just assert on the bits that differ from add, above.) +; CHECK:atomicrmw.start: +; CHECK: %new = xor i32 %loaded, %ValOperand_Shifted +; CHECK: %6 = cmpxchg i32* %AlignedAddr, i32 %loaded, i32 %new monotonic monotonic +; CHECK:atomicrmw.end: +define i16 @test_xor_i16(i16* %arg, i16 %val) { +entry: + %ret = atomicrmw xor i16* %arg, i16 %val seq_cst + ret i16 %ret +} + +; CHECK-LABEL: @test_min_i16( +; CHECK:atomicrmw.start: +; CHECK: %6 = lshr i32 %loaded, %ShiftAmt +; CHECK: %7 = trunc i32 %6 to i16 +; CHECK: %8 = icmp sle i16 %7, %val +; CHECK: %new = select i1 %8, i16 %7, i16 %val +; CHECK: %9 = zext i16 %new to i32 +; CHECK: %10 = shl i32 %9, %ShiftAmt +; CHECK: %11 = and i32 %loaded, %Inv_Mask +; CHECK: %12 = or i32 %11, %10 +; CHECK: %13 = cmpxchg i32* %AlignedAddr, i32 %loaded, i32 %12 monotonic monotonic +; CHECK:atomicrmw.end: +define i16 @test_min_i16(i16* %arg, i16 %val) { +entry: + %ret = atomicrmw min i16* %arg, i16 %val seq_cst + ret i16 %ret +} |

