[X86] Add broadcast instructions to the table used by ExeDepsFix pass.

Adds the different broadcast instructions to the ReplaceableInstrsAVX2 table.
That way the ExeDepsFix pass can take better decisions when AVX2 broadcasts are
across domain (int <-> float).

In particular, prior to this patch we were generating:
  vpbroadcastd  LCPI1_0(%rip), %ymm2
  vpand %ymm2, %ymm0, %ymm0
  vmaxps  %ymm1, %ymm0, %ymm0 ## <- domain change penalty

Now, we generate the following nice sequence where everything is in the float
domain:
  vbroadcastss  LCPI1_0(%rip), %ymm2
  vandps  %ymm2, %ymm0, %ymm0
  vmaxps  %ymm1, %ymm0, %ymm0

<rdar://problem/16354675>


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@204770 91177308-0d34-0410-b5e6-96231b3b80d8

3 files changed, 137 insertions, 9 deletions

diff --git a/test/CodeGen/X86/avx2-intrinsics-x86.ll b/test/CodeGen/X86/avx2-intrinsics-x86.ll
index 7ee944b7c9..ab3d591e1d 100644
--- a/test/CodeGen/X86/avx2-intrinsics-x86.ll
+++ b/test/CodeGen/X86/avx2-intrinsics-x86.ll
@@ -753,7 +753,7 @@ declare <16 x i16> @llvm.x86.avx2.pbroadcastw.256(<8 x i16>) nounwind readonly
 
 
 define <4 x i32> @test_x86_avx2_pbroadcastd_128(<4 x i32> %a0) {
-  ; CHECK: vpbroadcastd
+  ; CHECK: vbroadcastss
   %res = call <4 x i32> @llvm.x86.avx2.pbroadcastd.128(<4 x i32> %a0) ; <<4 x i32>> [#uses=1]
   ret <4 x i32> %res
 }
@@ -761,7 +761,7 @@ declare <4 x i32> @llvm.x86.avx2.pbroadcastd.128(<4 x i32>) nounwind readonly
 
 
 define <8 x i32> @test_x86_avx2_pbroadcastd_256(<4 x i32> %a0) {
-  ; CHECK: vpbroadcastd
+  ; CHECK: vbroadcastss {{[^,]+}}, %ymm{{[0-9]+}}
   %res = call <8 x i32> @llvm.x86.avx2.pbroadcastd.256(<4 x i32> %a0) ; <<8 x i32>> [#uses=1]
   ret <8 x i32> %res
 }
@@ -777,7 +777,7 @@ declare <2 x i64> @llvm.x86.avx2.pbroadcastq.128(<2 x i64>) nounwind readonly
 
 
 define <4 x i64> @test_x86_avx2_pbroadcastq_256(<2 x i64> %a0) {
-  ; CHECK: vpbroadcastq
+  ; CHECK: vbroadcastsd {{[^,]+}}, %ymm{{[0-9]+}}
   %res = call <4 x i64> @llvm.x86.avx2.pbroadcastq.256(<2 x i64> %a0) ; <<4 x i64>> [#uses=1]
   ret <4 x i64> %res
 }
diff --git a/test/CodeGen/X86/avx2-vbroadcast.ll b/test/CodeGen/X86/avx2-vbroadcast.ll
index bac9c66c3a..66f586d23d 100644
--- a/test/CodeGen/X86/avx2-vbroadcast.ll
+++ b/test/CodeGen/X86/avx2-vbroadcast.ll
@@ -98,7 +98,7 @@ entry:
   %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
@@ -108,7 +108,7 @@ entry:
   %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
@@ -130,7 +130,7 @@ entry:
   %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
@@ -293,7 +293,7 @@ define   <8 x i16> @_inreg8xi16(<8 x i16> %a) {
 
 
 ;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
@@ -325,7 +325,7 @@ define   <2 x double> @_inreg2xdouble(<2 x double> %a) {
 }
 
 ;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
@@ -333,7 +333,7 @@ define   <8 x i32> @_inreg8xi32(<8 x i32> %a) {
 }
 
 ;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
diff --git a/test/CodeGen/X86/exedepsfix-broadcast.ll b/test/CodeGen/X86/exedepsfix-broadcast.ll
new file mode 100644
index 0000000000..a18f751956
--- /dev/null
+++ b/test/CodeGen/X86/exedepsfix-broadcast.ll
@@ -0,0 +1,128 @@
+; 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
+}
+