{ X86::VINSERTF128rm, X86::VINSERTF128rm, X86::VINSERTI128rm },
{ X86::VINSERTF128rr, X86::VINSERTF128rr, X86::VINSERTI128rr },
{ X86::VPERM2F128rm, X86::VPERM2F128rm, X86::VPERM2I128rm },
- { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr }
+ { X86::VPERM2F128rr, X86::VPERM2F128rr, X86::VPERM2I128rr },
+ { X86::VBROADCASTSSrm, X86::VBROADCASTSSrm, X86::VPBROADCASTDrm},
+ { X86::VBROADCASTSSrr, X86::VBROADCASTSSrr, X86::VPBROADCASTDrr},
+ { X86::VBROADCASTSSYrr, X86::VBROADCASTSSYrr, X86::VPBROADCASTDYrr},
+ { X86::VBROADCASTSSYrm, X86::VBROADCASTSSYrm, X86::VPBROADCASTDYrm},
+ { X86::VBROADCASTSDYrr, X86::VBROADCASTSDYrr, X86::VPBROADCASTQYrr},
+ { X86::VBROADCASTSDYrm, X86::VBROADCASTSDYrm, X86::VPBROADCASTQYrm}
};
// FIXME: Some shuffle and unpack instructions have equivalents in different
%qf = insertelement <16 x i16> %qe, i16 %q, i32 15
ret <16 x i16> %qf
}
-; CHECK: vpbroadcastd (%
+; CHECK: vbroadcastss (%
define <4 x i32> @D32(i32* %ptr) nounwind uwtable readnone ssp {
entry:
%q = load i32* %ptr, align 4
%q3 = insertelement <4 x i32> %q2, i32 %q, i32 3
ret <4 x i32> %q3
}
-; CHECK: vpbroadcastd (%
+; CHECK: vbroadcastss (%
define <8 x i32> @DD32(i32* %ptr) nounwind uwtable readnone ssp {
entry:
%q = load i32* %ptr, align 4
%q1 = insertelement <2 x i64> %q0, i64 %q, i32 1
ret <2 x i64> %q1
}
-; CHECK: vpbroadcastq (%
+; CHECK: vbroadcastsd (%
define <4 x i64> @QQ64(i64* %ptr) nounwind uwtable readnone ssp {
entry:
%q = load i64* %ptr, align 4
;CHECK-LABEL: _inreg4xi64:
-;CHECK: vpbroadcastq
+;CHECK: vbroadcastsd
;CHECK: ret
define <4 x i64> @_inreg4xi64(<4 x i64> %a) {
%b = shufflevector <4 x i64> %a, <4 x i64> undef, <4 x i32> zeroinitializer
}
;CHECK-LABEL: _inreg8xi32:
-;CHECK: vpbroadcastd
+;CHECK: vbroadcastss
;CHECK: ret
define <8 x i32> @_inreg8xi32(<8 x i32> %a) {
%b = shufflevector <8 x i32> %a, <8 x i32> undef, <8 x i32> zeroinitializer
}
;CHECK-LABEL: _inreg4xi32:
-;CHECK: vpbroadcastd
+;CHECK: vbroadcastss
;CHECK: ret
define <4 x i32> @_inreg4xi32(<4 x i32> %a) {
%b = shufflevector <4 x i32> %a, <4 x i32> undef, <4 x i32> zeroinitializer
--- /dev/null
+; RUN: llc -O3 -mtriple=x86_64-apple-macosx -o - < %s -mattr=+avx2 -enable-unsafe-fp-math -mcpu=core2 | FileCheck %s
+; Check that the ExeDepsFix pass correctly fixes the domain for broadcast instructions.
+; <rdar://problem/16354675>
+
+; CHECK-LABEL: ExeDepsFix_broadcastss
+; CHECK: broadcastss
+; CHECK: vandps
+; CHECK: vmaxps
+; CHECK: ret
+define <4 x float> @ExeDepsFix_broadcastss(<4 x float> %arg, <4 x float> %arg2) {
+ %bitcast = bitcast <4 x float> %arg to <4 x i32>
+ %and = and <4 x i32> %bitcast, <i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647>
+ %floatcast = bitcast <4 x i32> %and to <4 x float>
+ %max_is_x = fcmp oge <4 x float> %floatcast, %arg2
+ %max = select <4 x i1> %max_is_x, <4 x float> %floatcast, <4 x float> %arg2
+ ret <4 x float> %max
+}
+
+; CHECK-LABEL: ExeDepsFix_broadcastss256
+; CHECK: broadcastss
+; CHECK: vandps
+; CHECK: vmaxps
+; CHECK: ret
+define <8 x float> @ExeDepsFix_broadcastss256(<8 x float> %arg, <8 x float> %arg2) {
+ %bitcast = bitcast <8 x float> %arg to <8 x i32>
+ %and = and <8 x i32> %bitcast, <i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647, i32 2147483647>
+ %floatcast = bitcast <8 x i32> %and to <8 x float>
+ %max_is_x = fcmp oge <8 x float> %floatcast, %arg2
+ %max = select <8 x i1> %max_is_x, <8 x float> %floatcast, <8 x float> %arg2
+ ret <8 x float> %max
+}
+
+
+; CHECK-LABEL: ExeDepsFix_broadcastss_inreg
+; CHECK: broadcastss
+; CHECK: vandps
+; CHECK: vmaxps
+; CHECK: ret
+define <4 x float> @ExeDepsFix_broadcastss_inreg(<4 x float> %arg, <4 x float> %arg2, i32 %broadcastvalue) {
+ %bitcast = bitcast <4 x float> %arg to <4 x i32>
+ %in = insertelement <4 x i32> undef, i32 %broadcastvalue, i32 0
+ %mask = shufflevector <4 x i32> %in, <4 x i32> undef, <4 x i32> zeroinitializer
+ %and = and <4 x i32> %bitcast, %mask
+ %floatcast = bitcast <4 x i32> %and to <4 x float>
+ %max_is_x = fcmp oge <4 x float> %floatcast, %arg2
+ %max = select <4 x i1> %max_is_x, <4 x float> %floatcast, <4 x float> %arg2
+ ret <4 x float> %max
+}
+
+; CHECK-LABEL: ExeDepsFix_broadcastss256_inreg
+; CHECK: broadcastss
+; CHECK: vandps
+; CHECK: vmaxps
+; CHECK: ret
+define <8 x float> @ExeDepsFix_broadcastss256_inreg(<8 x float> %arg, <8 x float> %arg2, i32 %broadcastvalue) {
+ %bitcast = bitcast <8 x float> %arg to <8 x i32>
+ %in = insertelement <8 x i32> undef, i32 %broadcastvalue, i32 0
+ %mask = shufflevector <8 x i32> %in, <8 x i32> undef, <8 x i32> zeroinitializer
+ %and = and <8 x i32> %bitcast, %mask
+ %floatcast = bitcast <8 x i32> %and to <8 x float>
+ %max_is_x = fcmp oge <8 x float> %floatcast, %arg2
+ %max = select <8 x i1> %max_is_x, <8 x float> %floatcast, <8 x float> %arg2
+ ret <8 x float> %max
+}
+
+; CHECK-LABEL: ExeDepsFix_broadcastsd
+; In that case the broadcast is directly folded into vandpd.
+; CHECK: vandpd
+; CHECK: vmaxpd
+; CHECK:ret
+define <2 x double> @ExeDepsFix_broadcastsd(<2 x double> %arg, <2 x double> %arg2) {
+ %bitcast = bitcast <2 x double> %arg to <2 x i64>
+ %and = and <2 x i64> %bitcast, <i64 2147483647, i64 2147483647>
+ %floatcast = bitcast <2 x i64> %and to <2 x double>
+ %max_is_x = fcmp oge <2 x double> %floatcast, %arg2
+ %max = select <2 x i1> %max_is_x, <2 x double> %floatcast, <2 x double> %arg2
+ ret <2 x double> %max
+}
+
+; CHECK-LABEL: ExeDepsFix_broadcastsd256
+; CHECK: broadcastsd
+; CHECK: vandpd
+; CHECK: vmaxpd
+; CHECK: ret
+define <4 x double> @ExeDepsFix_broadcastsd256(<4 x double> %arg, <4 x double> %arg2) {
+ %bitcast = bitcast <4 x double> %arg to <4 x i64>
+ %and = and <4 x i64> %bitcast, <i64 2147483647, i64 2147483647, i64 2147483647, i64 2147483647>
+ %floatcast = bitcast <4 x i64> %and to <4 x double>
+ %max_is_x = fcmp oge <4 x double> %floatcast, %arg2
+ %max = select <4 x i1> %max_is_x, <4 x double> %floatcast, <4 x double> %arg2
+ ret <4 x double> %max
+}
+
+
+; CHECK-LABEL: ExeDepsFix_broadcastsd_inreg
+; ExeDepsFix works top down, thus it coalesces vmovlhps domain with
+; vandps and there is nothing more you can do to match vmaxpd.
+; CHECK: vmovlhps
+; CHECK: vandps
+; CHECK: vmaxpd
+; CHECK: ret
+define <2 x double> @ExeDepsFix_broadcastsd_inreg(<2 x double> %arg, <2 x double> %arg2, i64 %broadcastvalue) {
+ %bitcast = bitcast <2 x double> %arg to <2 x i64>
+ %in = insertelement <2 x i64> undef, i64 %broadcastvalue, i32 0
+ %mask = shufflevector <2 x i64> %in, <2 x i64> undef, <2 x i32> zeroinitializer
+ %and = and <2 x i64> %bitcast, %mask
+ %floatcast = bitcast <2 x i64> %and to <2 x double>
+ %max_is_x = fcmp oge <2 x double> %floatcast, %arg2
+ %max = select <2 x i1> %max_is_x, <2 x double> %floatcast, <2 x double> %arg2
+ ret <2 x double> %max
+}
+
+; CHECK-LABEL: ExeDepsFix_broadcastsd256_inreg
+; CHECK: broadcastsd
+; CHECK: vandpd
+; CHECK: vmaxpd
+; CHECK: ret
+define <4 x double> @ExeDepsFix_broadcastsd256_inreg(<4 x double> %arg, <4 x double> %arg2, i64 %broadcastvalue) {
+ %bitcast = bitcast <4 x double> %arg to <4 x i64>
+ %in = insertelement <4 x i64> undef, i64 %broadcastvalue, i32 0
+ %mask = shufflevector <4 x i64> %in, <4 x i64> undef, <4 x i32> zeroinitializer
+ %and = and <4 x i64> %bitcast, %mask
+ %floatcast = bitcast <4 x i64> %and to <4 x double>
+ %max_is_x = fcmp oge <4 x double> %floatcast, %arg2
+ %max = select <4 x i1> %max_is_x, <4 x double> %floatcast, <4 x double> %arg2
+ ret <4 x double> %max
+}
+