X86 cost model: Differentiate cost for vector shifts of constants

SSE2 has efficient support for shifts by a scalar. My previous change of making
shifts expensive did not take this into account marking all shifts as expensive.
This would prevent vectorization from happening where it is actually beneficial.

With this change we differentiate between shifts of constants and other shifts.

radar://13576547

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@178808 91177308-0d34-0410-b5e6-96231b3b80d8

1 files changed, 287 insertions, 0 deletions

diff --git a/test/Analysis/CostModel/X86/testshiftlshr.ll b/test/Analysis/CostModel/X86/testshiftlshr.ll
index 7d665fc..8d6ef38 100644
--- a/test/Analysis/CostModel/X86/testshiftlshr.ll
+++ b/test/Analysis/CostModel/X86/testshiftlshr.ll
@@ -241,3 +241,290 @@ entry:
   ret %shifttype32i8 %0
 }
 
+; Test shift by a constant vector.
+
+%shifttypec = type <2 x i16>
+define %shifttypec @shift2i16const(%shifttypec %a, %shifttypec %b) {
+entry:
+  ; SSE2: shift2i16const
+  ; SSE2: cost of 1 {{.*}} lshr
+  ; SSE2-CODEGEN: shift2i16const
+  ; SSE2-CODEGEN: psrlq $3
+
+  %0 = lshr %shifttypec %a , <i16 3, i16 3>
+  ret %shifttypec %0
+}
+
+%shifttypec4i16 = type <4 x i16>
+define %shifttypec4i16 @shift4i16const(%shifttypec4i16 %a, %shifttypec4i16 %b) {
+entry:
+  ; SSE2: shift4i16const
+  ; SSE2: cost of 1 {{.*}} lshr
+  ; SSE2-CODEGEN: shift4i16const
+  ; SSE2-CODEGEN: psrld $3
+
+  %0 = lshr %shifttypec4i16 %a , <i16 3, i16 3, i16 3, i16 3>
+  ret %shifttypec4i16 %0
+}
+
+%shifttypec8i16 = type <8 x i16>
+define %shifttypec8i16 @shift8i16const(%shifttypec8i16 %a, %shifttypec8i16 %b) {
+entry:
+  ; SSE2: shift8i16const
+  ; SSE2: cost of 1 {{.*}} lshr
+  ; SSE2-CODEGEN: shift8i16const
+  ; SSE2-CODEGEN: psrlw $3
+
+  %0 = lshr %shifttypec8i16 %a , <i16 3, i16 3, i16 3, i16 3,
+                                  i16 3, i16 3, i16 3, i16 3>
+  ret %shifttypec8i16 %0
+}
+
+%shifttypec16i16 = type <16 x i16>
+define %shifttypec16i16 @shift16i16const(%shifttypec16i16 %a,
+                                         %shifttypec16i16 %b) {
+entry:
+  ; SSE2: shift16i16const
+  ; SSE2: cost of 2 {{.*}} lshr
+  ; SSE2-CODEGEN: shift16i16const
+  ; SSE2-CODEGEN: psrlw $3
+
+  %0 = lshr %shifttypec16i16 %a , <i16 3, i16 3, i16 3, i16 3,
+                                   i16 3, i16 3, i16 3, i16 3,
+                                   i16 3, i16 3, i16 3, i16 3,
+                                   i16 3, i16 3, i16 3, i16 3>
+  ret %shifttypec16i16 %0
+}
+
+%shifttypec32i16 = type <32 x i16>
+define %shifttypec32i16 @shift32i16const(%shifttypec32i16 %a,
+                                        %shifttypec32i16 %b) {
+entry:
+  ; SSE2: shift32i16const
+  ; SSE2: cost of 4 {{.*}} lshr
+  ; SSE2-CODEGEN: shift32i16const
+  ; SSE2-CODEGEN: psrlw $3
+
+  %0 = lshr %shifttypec32i16 %a , <i16 3, i16 3, i16 3, i16 3,
+                                   i16 3, i16 3, i16 3, i16 3,
+                                   i16 3, i16 3, i16 3, i16 3,
+                                   i16 3, i16 3, i16 3, i16 3,
+                                   i16 3, i16 3, i16 3, i16 3,
+                                   i16 3, i16 3, i16 3, i16 3,
+                                   i16 3, i16 3, i16 3, i16 3,
+                                   i16 3, i16 3, i16 3, i16 3>
+  ret %shifttypec32i16 %0
+}
+
+%shifttypec2i32 = type <2 x i32>
+define %shifttypec2i32 @shift2i32c(%shifttypec2i32 %a, %shifttypec2i32 %b) {
+entry:
+  ; SSE2: shift2i32c
+  ; SSE2: cost of 1 {{.*}} lshr
+  ; SSE2-CODEGEN: shift2i32c
+  ; SSE2-CODEGEN: psrlq $3
+
+  %0 = lshr %shifttypec2i32 %a , <i32 3, i32 3>
+  ret %shifttypec2i32 %0
+}
+
+%shifttypec4i32 = type <4 x i32>
+define %shifttypec4i32 @shift4i32c(%shifttypec4i32 %a, %shifttypec4i32 %b) {
+entry:
+  ; SSE2: shift4i32c
+  ; SSE2: cost of 1 {{.*}} lshr
+  ; SSE2-CODEGEN: shift4i32c
+  ; SSE2-CODEGEN: psrld $3
+
+  %0 = lshr %shifttypec4i32 %a , <i32 3, i32 3, i32 3, i32 3>
+  ret %shifttypec4i32 %0
+}
+
+%shifttypec8i32 = type <8 x i32>
+define %shifttypec8i32 @shift8i32c(%shifttypec8i32 %a, %shifttypec8i32 %b) {
+entry:
+  ; SSE2: shift8i32c
+  ; SSE2: cost of 2 {{.*}} lshr
+  ; SSE2-CODEGEN: shift8i32c
+  ; SSE2-CODEGEN: psrld $3
+
+  %0 = lshr %shifttypec8i32 %a , <i32 3, i32 3, i32 3, i32 3,
+                                  i32 3, i32 3, i32 3, i32 3>
+  ret %shifttypec8i32 %0
+}
+
+%shifttypec16i32 = type <16 x i32>
+define %shifttypec16i32 @shift16i32c(%shifttypec16i32 %a, %shifttypec16i32 %b) {
+entry:
+  ; SSE2: shift16i32c
+  ; SSE2: cost of 4 {{.*}} lshr
+  ; SSE2-CODEGEN: shift16i32c
+  ; SSE2-CODEGEN: psrld $3
+
+  %0 = lshr %shifttypec16i32 %a , <i32 3, i32 3, i32 3, i32 3,
+                                   i32 3, i32 3, i32 3, i32 3,
+                                   i32 3, i32 3, i32 3, i32 3,
+                                   i32 3, i32 3, i32 3, i32 3>
+  ret %shifttypec16i32 %0
+}
+
+%shifttypec32i32 = type <32 x i32>
+define %shifttypec32i32 @shift32i32c(%shifttypec32i32 %a, %shifttypec32i32 %b) {
+entry:
+  ; SSE2: shift32i32c
+  ; getTypeConversion fails here and promotes this to a i64.
+  ; SSE2: cost of 256 {{.*}} lshr
+  ; SSE2-CODEGEN: shift32i32c
+  ; SSE2-CODEGEN: psrld $3
+  %0 = lshr %shifttypec32i32 %a , <i32 3, i32 3, i32 3, i32 3,
+                                   i32 3, i32 3, i32 3, i32 3,
+                                   i32 3, i32 3, i32 3, i32 3,
+                                   i32 3, i32 3, i32 3, i32 3,
+                                   i32 3, i32 3, i32 3, i32 3,
+                                   i32 3, i32 3, i32 3, i32 3,
+                                   i32 3, i32 3, i32 3, i32 3,
+                                   i32 3, i32 3, i32 3, i32 3>
+  ret %shifttypec32i32 %0
+}
+
+%shifttypec2i64 = type <2 x i64>
+define %shifttypec2i64 @shift2i64c(%shifttypec2i64 %a, %shifttypec2i64 %b) {
+entry:
+  ; SSE2: shift2i64c
+  ; SSE2: cost of 1 {{.*}} lshr
+  ; SSE2-CODEGEN: shift2i64c
+  ; SSE2-CODEGEN: psrlq $3
+
+  %0 = lshr %shifttypec2i64 %a , <i64 3, i64 3>
+  ret %shifttypec2i64 %0
+}
+
+%shifttypec4i64 = type <4 x i64>
+define %shifttypec4i64 @shift4i64c(%shifttypec4i64 %a, %shifttypec4i64 %b) {
+entry:
+  ; SSE2: shift4i64c
+  ; SSE2: cost of 2 {{.*}} lshr
+  ; SSE2-CODEGEN: shift4i64c
+  ; SSE2-CODEGEN: psrlq $3
+
+  %0 = lshr %shifttypec4i64 %a , <i64 3, i64 3, i64 3, i64 3>
+  ret %shifttypec4i64 %0
+}
+
+%shifttypec8i64 = type <8 x i64>
+define %shifttypec8i64 @shift8i64c(%shifttypec8i64 %a, %shifttypec8i64 %b) {
+entry:
+  ; SSE2: shift8i64c
+  ; SSE2: cost of 4 {{.*}} lshr
+  ; SSE2-CODEGEN: shift8i64c
+  ; SSE2-CODEGEN: psrlq $3
+
+ %0 = lshr %shifttypec8i64 %a , <i64 3, i64 3, i64 3, i64 3,
+                                 i64 3, i64 3, i64 3, i64 3>
+  ret %shifttypec8i64 %0
+}
+
+%shifttypec16i64 = type <16 x i64>
+define %shifttypec16i64 @shift16i64c(%shifttypec16i64 %a, %shifttypec16i64 %b) {
+entry:
+  ; SSE2: shift16i64c
+  ; SSE2: cost of 8 {{.*}} lshr
+  ; SSE2-CODEGEN: shift16i64c
+  ; SSE2-CODEGEN: psrlq $3
+
+  %0 = lshr %shifttypec16i64 %a , <i64 3, i64 3, i64 3, i64 3,
+                                   i64 3, i64 3, i64 3, i64 3,
+                                   i64 3, i64 3, i64 3, i64 3,
+                                   i64 3, i64 3, i64 3, i64 3>
+  ret %shifttypec16i64 %0
+}
+
+%shifttypec32i64 = type <32 x i64>
+define %shifttypec32i64 @shift32i64c(%shifttypec32i64 %a, %shifttypec32i64 %b) {
+entry:
+  ; SSE2: shift32i64c
+  ; SSE2: cost of 256 {{.*}} lshr
+  ; SSE2-CODEGEN: shift32i64c
+  ; SSE2-CODEGEN: psrlq $3
+
+  %0 = lshr %shifttypec32i64 %a ,<i64 3, i64 3, i64 3, i64 3,
+                                  i64 3, i64 3, i64 3, i64 3,
+                                  i64 3, i64 3, i64 3, i64 3,
+                                  i64 3, i64 3, i64 3, i64 3,
+                                  i64 3, i64 3, i64 3, i64 3,
+                                  i64 3, i64 3, i64 3, i64 3,
+                                  i64 3, i64 3, i64 3, i64 3,
+                                  i64 3, i64 3, i64 3, i64 3>
+  ret %shifttypec32i64 %0
+}
+
+%shifttypec2i8 = type <2 x i8>
+define %shifttypec2i8 @shift2i8c(%shifttypec2i8 %a, %shifttypec2i8 %b) {
+entry:
+  ; SSE2: shift2i8c
+  ; SSE2: cost of 1 {{.*}} lshr
+  ; SSE2-CODEGEN: shift2i8c
+  ; SSE2-CODEGEN: psrlq $3
+
+  %0 = lshr %shifttypec2i8 %a , <i8 3, i8 3>
+  ret %shifttypec2i8 %0
+}
+
+%shifttypec4i8 = type <4 x i8>
+define %shifttypec4i8 @shift4i8c(%shifttypec4i8 %a, %shifttypec4i8 %b) {
+entry:
+  ; SSE2: shift4i8c
+  ; SSE2: cost of 1 {{.*}} lshr
+  ; SSE2-CODEGEN: shift4i8c
+  ; SSE2-CODEGEN: psrld $3
+
+  %0 = lshr %shifttypec4i8 %a , <i8 3, i8 3, i8 3, i8 3>
+  ret %shifttypec4i8 %0
+}
+
+%shifttypec8i8 = type <8 x i8>
+define %shifttypec8i8 @shift8i8c(%shifttypec8i8 %a, %shifttypec8i8 %b) {
+entry:
+  ; SSE2: shift8i8c
+  ; SSE2: cost of 1 {{.*}} lshr
+  ; SSE2-CODEGEN: shift8i8c
+  ; SSE2-CODEGEN: psrlw $3
+
+  %0 = lshr %shifttypec8i8 %a , <i8 3, i8 3, i8 3, i8 3,
+                                 i8 3, i8 3, i8 3, i8 3>
+  ret %shifttypec8i8 %0
+}
+
+%shifttypec16i8 = type <16 x i8>
+define %shifttypec16i8 @shift16i8c(%shifttypec16i8 %a, %shifttypec16i8 %b) {
+entry:
+  ; SSE2: shift16i8c
+  ; SSE2: cost of 1 {{.*}} lshr
+  ; SSE2-CODEGEN: shift16i8c
+  ; SSE2-CODEGEN: psrlw $3
+
+  %0 = lshr %shifttypec16i8 %a , <i8 3, i8 3, i8 3, i8 3,
+                                  i8 3, i8 3, i8 3, i8 3,
+                                  i8 3, i8 3, i8 3, i8 3,
+                                  i8 3, i8 3, i8 3, i8 3>
+  ret %shifttypec16i8 %0
+}
+
+%shifttypec32i8 = type <32 x i8>
+define %shifttypec32i8 @shift32i8c(%shifttypec32i8 %a, %shifttypec32i8 %b) {
+entry:
+  ; SSE2: shift32i8c
+  ; SSE2: cost of 2 {{.*}} lshr
+  ; SSE2-CODEGEN: shift32i8c
+  ; SSE2-CODEGEN: psrlw $3
+
+  %0 = lshr %shifttypec32i8 %a , <i8 3, i8 3, i8 3, i8 3,
+                                  i8 3, i8 3, i8 3, i8 3,
+                                  i8 3, i8 3, i8 3, i8 3,
+                                  i8 3, i8 3, i8 3, i8 3,
+                                  i8 3, i8 3, i8 3, i8 3,
+                                  i8 3, i8 3, i8 3, i8 3,
+                                  i8 3, i8 3, i8 3, i8 3,
+                                  i8 3, i8 3, i8 3, i8 3>
+  ret %shifttypec32i8 %0
+}