diff --git a/test/CodeGen/X86/test-nofold.ll b/test/CodeGen/X86/test-nofold.ll index 772ff6c3e76..f1063dcabf4 100644 --- a/test/CodeGen/X86/test-nofold.ll +++ b/test/CodeGen/X86/test-nofold.ll @@ -1,22 +1,35 @@ -; RUN: llc < %s -march=x86 -mcpu=yonah | grep {testl.*%e.x.*%e.x} +; RUN: llc < %s -march=x86 -mcpu=yonah | FileCheck %s ; rdar://5752025 -; We don't want to fold the and into the test, because the and clobbers its -; input forcing a copy. We want: -; movl $15, %ecx -; andl 4(%esp), %ecx -; testl %ecx, %ecx +; We want: +; CHECK: movl 4(%esp), %ecx +; CHECK-NEXT: andl $15, %ecx +; CHECK-NEXT: movl $42, %eax +; CHECK-NEXT: cmovel %ecx, %eax +; CHECK-NEXT: ret +; +; We don't want: +; movl 4(%esp), %eax +; movl %eax, %ecx # bad: extra copy +; andl $15, %ecx +; testl $15, %eax # bad: peep obstructed ; movl $42, %eax -; cmove %ecx, %eax +; cmovel %ecx, %eax ; ret ; -; Not: -; movl 4(%esp), %eax -; movl %eax, %ecx -; andl $15, %ecx -; testl $15, %eax +; We also don't want: +; movl $15, %ecx # bad: larger encoding +; andl 4(%esp), %ecx ; movl $42, %eax -; cmove %ecx, %eax +; cmovel %ecx, %eax +; ret +; +; We also don't want: +; movl 4(%esp), %ecx +; andl $15, %ecx +; testl %ecx, %ecx # bad: unnecessary test +; movl $42, %eax +; cmovel %ecx, %eax ; ret define i32 @t1(i32 %X) nounwind { diff --git a/test/CodeGen/X86/twoaddr-lea.ll b/test/CodeGen/X86/twoaddr-lea.ll new file mode 100644 index 00000000000..a245ed7caa8 --- /dev/null +++ b/test/CodeGen/X86/twoaddr-lea.ll @@ -0,0 +1,24 @@ +;; X's live range extends beyond the shift, so the register allocator +;; cannot coalesce it with Y. Because of this, a copy needs to be +;; emitted before the shift to save the register value before it is +;; clobbered. However, this copy is not needed if the register +;; allocator turns the shift into an LEA. This also occurs for ADD. + +; Check that the shift gets turned into an LEA. +; RUN: llc < %s -march=x86 -x86-asm-syntax=intel | \ +; RUN: not grep {mov E.X, E.X} + +@G = external global i32 ; [#uses=3] + +define i32 @test1(i32 %X, i32 %Y) { + %Z = add i32 %X, %Y ; [#uses=1] + volatile store i32 %Y, i32* @G + volatile store i32 %Z, i32* @G + ret i32 %X +} + +define i32 @test2(i32 %X) { + %Z = add i32 %X, 1 ; [#uses=1] + volatile store i32 %Z, i32* @G + ret i32 %X +} diff --git a/test/CodeGen/X86/use-add-flags.ll b/test/CodeGen/X86/use-add-flags.ll new file mode 100644 index 00000000000..8252030830e --- /dev/null +++ b/test/CodeGen/X86/use-add-flags.ll @@ -0,0 +1,56 @@ +; RUN: llc < %s -march=x86-64 -o - | FileCheck %s + +; Reuse the flags value from the add instructions instead of emitting separate +; testl instructions. + +; Use the flags on the add. + +; CHECK: add_zf: +; CHECK: addl (%rdi), %esi +; CHECK-NEXT: movl %edx, %eax +; CHECK-NEXT: cmovnsl %ecx, %eax +; CHECK-NEXT: ret + +define i32 @add_zf(i32* %x, i32 %y, i32 %a, i32 %b) nounwind { + %tmp2 = load i32* %x, align 4 ; [#uses=1] + %tmp4 = add i32 %tmp2, %y ; [#uses=1] + %tmp5 = icmp slt i32 %tmp4, 0 ; [#uses=1] + %tmp.0 = select i1 %tmp5, i32 %a, i32 %b ; [#uses=1] + ret i32 %tmp.0 +} + +declare void @foo(i32) + +; Don't use the flags result of the and here, since the and has no +; other use. A simple test is better. + +; CHECK: bar: +; CHECK: testb $16, %dil + +define void @bar(i32 %x) nounwind { + %y = and i32 %x, 16 + %t = icmp eq i32 %y, 0 + br i1 %t, label %true, label %false +true: + call void @foo(i32 %x) + ret void +false: + ret void +} + +; Do use the flags result of the and here, since the and has another use. + +; CHECK: qux: +; CHECK: andl $16, %edi +; CHECK-NEXT: jne .LBB3_2 + +define void @qux(i32 %x) nounwind { + %y = and i32 %x, 16 + %t = icmp eq i32 %y, 0 + br i1 %t, label %true, label %false +true: + call void @foo(i32 %y) + ret void +false: + ret void +}