[X86][BtVer2] Update latency of mmx horizontal operations

D56777 added +1cy local forwarding penalty for horizontal operations, but this penalty only affects sse2/xmm variants, the mmx variants don't suffer the penalty.

Confirmed with @andreadb

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@351755 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ScheduleBtVer2.td b/lib/Target/X86/X86ScheduleBtVer2.td
index f54e52ffa46dc9ecfe580ae4f1f1d9ed0395d815..e81f8445b3ec238e5c8e8b01b8055ce85e3f26eb 100644 (file)
--- a/lib/Target/X86/X86ScheduleBtVer2.td
+++ b/lib/Target/X86/X86ScheduleBtVer2.td
@@ -577,7 +577,7 @@ defm : JWriteResFpuPair<WriteAESDecEnc,   [JFPU01, JVALU, JFPU0, JVIMUL], 3, [1,
 
 defm : JWriteResFpuPair<WriteFHAdd,         [JFPU0, JFPA], 4>;            // +1cy latency.
 defm : JWriteResYMMPair<WriteFHAddY,        [JFPU0, JFPA], 4, [2,2], 2>;  // +1cy latency.
-defm : JWriteResFpuPair<WritePHAdd,         [JFPU01, JVALU], 2>;          // +1cy latency.
+defm : JWriteResFpuPair<WritePHAdd,         [JFPU01, JVALU], 1>;
 defm : JWriteResFpuPair<WritePHAddX,        [JFPU01, JVALU], 2>;          // +1cy latency.
 defm : X86WriteResPairUnsupported<WritePHAddY>;
 

diff --git a/test/CodeGen/X86/mmx-schedule.ll b/test/CodeGen/X86/mmx-schedule.ll
index 513332f61f11373833d42fb1c94bd1f44c8c2765..51dc5e102ff1578aa59144b24d37e4623ba23cf3 100644 (file)
--- a/test/CodeGen/X86/mmx-schedule.ll
+++ b/test/CodeGen/X86/mmx-schedule.ll
@@ -3368,8 +3368,8 @@ define i64 @test_phaddd(x86_mmx %a0, x86_mmx %a1, x86_mmx* %a2) optsize {
 ;
 ; BTVER2-LABEL: test_phaddd:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    phaddd %mm1, %mm0 # sched: [2:0.50]
-; BTVER2-NEXT:    phaddd (%rdi), %mm0 # sched: [7:1.00]
+; BTVER2-NEXT:    phaddd %mm1, %mm0 # sched: [1:0.50]
+; BTVER2-NEXT:    phaddd (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    movq %mm0, %rax # sched: [4:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
@@ -3453,8 +3453,8 @@ define i64 @test_phaddsw(x86_mmx %a0, x86_mmx %a1, x86_mmx* %a2) optsize {
 ;
 ; BTVER2-LABEL: test_phaddsw:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    phaddsw %mm1, %mm0 # sched: [2:0.50]
-; BTVER2-NEXT:    phaddsw (%rdi), %mm0 # sched: [7:1.00]
+; BTVER2-NEXT:    phaddsw %mm1, %mm0 # sched: [1:0.50]
+; BTVER2-NEXT:    phaddsw (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    movq %mm0, %rax # sched: [4:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
@@ -3538,8 +3538,8 @@ define i64 @test_phaddw(x86_mmx %a0, x86_mmx %a1, x86_mmx* %a2) optsize {
 ;
 ; BTVER2-LABEL: test_phaddw:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    phaddw %mm1, %mm0 # sched: [2:0.50]
-; BTVER2-NEXT:    phaddw (%rdi), %mm0 # sched: [7:1.00]
+; BTVER2-NEXT:    phaddw %mm1, %mm0 # sched: [1:0.50]
+; BTVER2-NEXT:    phaddw (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    movq %mm0, %rax # sched: [4:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
@@ -3623,8 +3623,8 @@ define i64 @test_phsubd(x86_mmx %a0, x86_mmx %a1, x86_mmx* %a2) optsize {
 ;
 ; BTVER2-LABEL: test_phsubd:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    phsubd %mm1, %mm0 # sched: [2:0.50]
-; BTVER2-NEXT:    phsubd (%rdi), %mm0 # sched: [7:1.00]
+; BTVER2-NEXT:    phsubd %mm1, %mm0 # sched: [1:0.50]
+; BTVER2-NEXT:    phsubd (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    movq %mm0, %rax # sched: [4:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
@@ -3708,8 +3708,8 @@ define i64 @test_phsubsw(x86_mmx %a0, x86_mmx %a1, x86_mmx* %a2) optsize {
 ;
 ; BTVER2-LABEL: test_phsubsw:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    phsubsw %mm1, %mm0 # sched: [2:0.50]
-; BTVER2-NEXT:    phsubsw (%rdi), %mm0 # sched: [7:1.00]
+; BTVER2-NEXT:    phsubsw %mm1, %mm0 # sched: [1:0.50]
+; BTVER2-NEXT:    phsubsw (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    movq %mm0, %rax # sched: [4:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
@@ -3793,8 +3793,8 @@ define i64 @test_phsubw(x86_mmx %a0, x86_mmx %a1, x86_mmx* %a2) optsize {
 ;
 ; BTVER2-LABEL: test_phsubw:
 ; BTVER2:       # %bb.0:
-; BTVER2-NEXT:    phsubw %mm1, %mm0 # sched: [2:0.50]
-; BTVER2-NEXT:    phsubw (%rdi), %mm0 # sched: [7:1.00]
+; BTVER2-NEXT:    phsubw %mm1, %mm0 # sched: [1:0.50]
+; BTVER2-NEXT:    phsubw (%rdi), %mm0 # sched: [6:1.00]
 ; BTVER2-NEXT:    movq %mm0, %rax # sched: [4:1.00]
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;

diff --git a/test/tools/llvm-mca/X86/BtVer2/resources-ssse3.s b/test/tools/llvm-mca/X86/BtVer2/resources-ssse3.s
index 0b64d1da48b62143e37fd89f530afdf572470b09..e35a474514966521462131f6a95a2fb5d6cc21b2 100644 (file)
--- a/test/tools/llvm-mca/X86/BtVer2/resources-ssse3.s
+++ b/test/tools/llvm-mca/X86/BtVer2/resources-ssse3.s
@@ -122,28 +122,28 @@ psignw      (%rax), %xmm2
 # CHECK-NEXT:  1      6     1.00    *                   palignr        $1, (%rax), %mm2
 # CHECK-NEXT:  1      1     0.50                        palignr        $1, %xmm0, %xmm2
 # CHECK-NEXT:  1      6     1.00    *                   palignr        $1, (%rax), %xmm2
-# CHECK-NEXT:  1      2     0.50                        phaddd %mm0, %mm2
-# CHECK-NEXT:  1      7     1.00    *                   phaddd (%rax), %mm2
+# CHECK-NEXT:  1      1     0.50                        phaddd %mm0, %mm2
+# CHECK-NEXT:  1      6     1.00    *                   phaddd (%rax), %mm2
 # CHECK-NEXT:  1      2     0.50                        phaddd %xmm0, %xmm2
 # CHECK-NEXT:  1      7     1.00    *                   phaddd (%rax), %xmm2
-# CHECK-NEXT:  1      2     0.50                        phaddsw        %mm0, %mm2
-# CHECK-NEXT:  1      7     1.00    *                   phaddsw        (%rax), %mm2
+# CHECK-NEXT:  1      1     0.50                        phaddsw        %mm0, %mm2
+# CHECK-NEXT:  1      6     1.00    *                   phaddsw        (%rax), %mm2
 # CHECK-NEXT:  1      2     0.50                        phaddsw        %xmm0, %xmm2
 # CHECK-NEXT:  1      7     1.00    *                   phaddsw        (%rax), %xmm2
-# CHECK-NEXT:  1      2     0.50                        phaddw %mm0, %mm2
-# CHECK-NEXT:  1      7     1.00    *                   phaddw (%rax), %mm2
+# CHECK-NEXT:  1      1     0.50                        phaddw %mm0, %mm2
+# CHECK-NEXT:  1      6     1.00    *                   phaddw (%rax), %mm2
 # CHECK-NEXT:  1      2     0.50                        phaddw %xmm0, %xmm2
 # CHECK-NEXT:  1      7     1.00    *                   phaddw (%rax), %xmm2
-# CHECK-NEXT:  1      2     0.50                        phsubd %mm0, %mm2
-# CHECK-NEXT:  1      7     1.00    *                   phsubd (%rax), %mm2
+# CHECK-NEXT:  1      1     0.50                        phsubd %mm0, %mm2
+# CHECK-NEXT:  1      6     1.00    *                   phsubd (%rax), %mm2
 # CHECK-NEXT:  1      2     0.50                        phsubd %xmm0, %xmm2
 # CHECK-NEXT:  1      7     1.00    *                   phsubd (%rax), %xmm2
-# CHECK-NEXT:  1      2     0.50                        phsubsw        %mm0, %mm2
-# CHECK-NEXT:  1      7     1.00    *                   phsubsw        (%rax), %mm2
+# CHECK-NEXT:  1      1     0.50                        phsubsw        %mm0, %mm2
+# CHECK-NEXT:  1      6     1.00    *                   phsubsw        (%rax), %mm2
 # CHECK-NEXT:  1      2     0.50                        phsubsw        %xmm0, %xmm2
 # CHECK-NEXT:  1      7     1.00    *                   phsubsw        (%rax), %xmm2
-# CHECK-NEXT:  1      2     0.50                        phsubw %mm0, %mm2
-# CHECK-NEXT:  1      7     1.00    *                   phsubw (%rax), %mm2
+# CHECK-NEXT:  1      1     0.50                        phsubw %mm0, %mm2
+# CHECK-NEXT:  1      6     1.00    *                   phsubw (%rax), %mm2
 # CHECK-NEXT:  1      2     0.50                        phsubw %xmm0, %xmm2
 # CHECK-NEXT:  1      7     1.00    *                   phsubw (%rax), %xmm2
 # CHECK-NEXT:  1      2     1.00                        pmaddubsw      %mm0, %mm2