From 7f6eb639bd9563c53a6eb035d198f8d525ce3c0e Mon Sep 17 00:00:00 2001 From: Dale Johannesen Date: Sat, 7 Aug 2010 00:33:42 +0000 Subject: Use sdmem and sse_load_f64 (etc.) for the vector form of CMPSD (etc.) Matching a 128-bit memory operand is wrong, the instruction uses only 64 bits (same as ADDSD etc.) 8193553. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@110491 91177308-0d34-0410-b5e6-96231b3b80d8 --- test/CodeGen/X86/2010-08-06-cmpsd.ll | 27 +++++++++++++++++++++++++++ 1 file changed, 27 insertions(+) create mode 100644 test/CodeGen/X86/2010-08-06-cmpsd.ll (limited to 'test/CodeGen') diff --git a/test/CodeGen/X86/2010-08-06-cmpsd.ll b/test/CodeGen/X86/2010-08-06-cmpsd.ll new file mode 100644 index 0000000000..99b963f453 --- /dev/null +++ b/test/CodeGen/X86/2010-08-06-cmpsd.ll @@ -0,0 +1,27 @@ +; RUN: llc < %s -mtriple=x86_64-applecl-darwin11 | FileCheck %s +; 8193553 + +define void @__math_kernel_Vectorized_wrapper(<4 x double> addrspace(1)* %a, <4 x double> addrspace(1)* %b, i64 addrspace(1)* %c, i64 addrspace(1)* %d) nounwind { +entry.i: ; preds = %entry.i, %loop +; CHECK: math_kernel_Vectorized_wrapper +; CHECK-NOT: cmpordsd (%rsi), + %0 = alloca i8 + %1 = alloca i8 + %2 = alloca i8 + %tmp213.i = load <4 x double> addrspace(1)* %a ; <<4 x double>> [#uses=4] + %extract25.i = extractelement <4 x double> %tmp213.i, i32 1 ; [#uses=1] + %tmp723.i = load <4 x double> addrspace(1)* %b ; <<4 x double>> [#uses=4] + %extract29.i = extractelement <4 x double> %tmp723.i, i32 1 ; [#uses=1] + %tmp2.i26 = insertelement <2 x double> undef, double %extract25.i, i32 0 ; <<2 x double>> [#uses=1] + %tmp5.i27 = insertelement <2 x double> undef, double %extract29.i, i32 1 ; <<2 x double>> [#uses=1] + %cmpsd.i.i28 = call <2 x double> @llvm.x86.sse2.cmp.sd(<2 x double> %tmp2.i26, <2 x double> %tmp5.i27, i8 7) nounwind ; <<2 x double>> [#uses=1] + %3 = bitcast <2 x double> %cmpsd.i.i28 to <4 x i32> ; <<4 x i32>> [#uses=1] + %tmp12.i29 = extractelement <4 x i32> %3, i32 0 ; [#uses=1] + %and.i30 = and i32 %tmp12.i29, 1 ; [#uses=1] + %conv937.i36 = zext i32 %and.i30 to i64 ; [#uses=1] + store i64 %conv937.i36, i64 addrspace(1)* %d + ret void +; CHECK: ret +} + +declare <2 x double> @llvm.x86.sse2.cmp.sd(<2 x double>, <2 x double>, i8) nounwind readnone -- cgit v1.2.3