diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp index 58bdc77ccdb..4b1c411824f 100644 --- a/lib/Target/X86/X86ISelLowering.cpp +++ b/lib/Target/X86/X86ISelLowering.cpp @@ -15098,7 +15098,8 @@ TargetLowering::ConstraintWeight break; case 'x': case 'Y': - if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM()) + if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM()) || + ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasAVX())) weight = CW_Register; break; case 'I': @@ -15378,8 +15379,8 @@ X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, case 'Y': // SSE_REGS if SSE2 allowed if (!Subtarget->hasXMMInt()) break; // FALL THROUGH. - case 'x': // SSE_REGS if SSE1 allowed - if (!Subtarget->hasXMM()) break; + case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed + if (!Subtarget->hasXMM() && !Subtarget->hasAVX()) break; switch (VT.getSimpleVT().SimpleTy) { default: break; @@ -15398,6 +15399,15 @@ X86TargetLowering::getRegForInlineAsmConstraint(const std::string &Constraint, case MVT::v4f32: case MVT::v2f64: return std::make_pair(0U, X86::VR128RegisterClass); + // AVX types. + case MVT::v32i8: + case MVT::v16i16: + case MVT::v8i32: + case MVT::v4i64: + case MVT::v8f32: + case MVT::v4f64: + return std::make_pair(0U, X86::VR256RegisterClass); + } break; } diff --git a/test/CodeGen/X86/inline-asm-q-regs.ll b/test/CodeGen/X86/inline-asm-q-regs.ll index 617bd39f096..fca68baac6e 100644 --- a/test/CodeGen/X86/inline-asm-q-regs.ll +++ b/test/CodeGen/X86/inline-asm-q-regs.ll @@ -1,4 +1,4 @@ -; RUN: llc < %s -march=x86-64 +; RUN: llc < %s -march=x86-64 -mattr=+avx ; rdar://7066579 %0 = type { i64, i64, i64, i64, i64 } ; type %0 @@ -27,3 +27,11 @@ entry: %0 = tail call { i8, i8, i8, i8, i8 } asm "foo $1, $2, $3, $4, $1\0Axchgb ${0:b}, ${0:h}", "=q,={ax},={bx},={cx},={dx},0,1,2,3,4,~{dirflag},~{fpsr},~{flags}"(i8 %val, i8 %a, i8 %b, i8 %c, i8 %d) nounwind ret void } + +; rdar://10614894 +define <8 x float> @test5(<8 x float> %a, <8 x float> %b) nounwind { +entry: + %0 = tail call <8 x float> asm "vperm2f128 $3, $2, $1, $0", "=x,x,x,i,~{dirflag},~{fpsr},~{flags}"(<8 x float> %a, <8 x float> %b, i32 16) nounwind + ret <8 x float> %0 +} +