[X86][Btver2] Split i8/i16/i32/i64 div/idiv costs

We were assuming a mixture of 32/64 division costs.

llvm-svn: 327407

2 files changed, 59 insertions, 20 deletions

diff --git a/llvm/lib/Target/X86/X86ScheduleBtVer2.td b/llvm/lib/Target/X86/X86ScheduleBtVer2.td
index ada5f3f9b5f..ac101102ca9 100644
--- a/llvm/lib/Target/X86/X86ScheduleBtVer2.td
+++ b/llvm/lib/Target/X86/X86ScheduleBtVer2.td
@@ -121,20 +121,59 @@ def  : WriteRes<WriteIMulH, [JALU1]> {
   let ResourceCycles = [4];
 }
 
-// FIXME 8/16 bit divisions
+// Worst case (i64 division)
 def : WriteRes<WriteIDiv, [JALU1, JDiv]> {
-  let Latency = 25;
-  let ResourceCycles = [1, 25];
-}
-def : WriteRes<WriteIDivLd, [JALU1, JLAGU, JDiv]> {
   let Latency = 41;
-  let ResourceCycles = [1, 1, 25];
+  let ResourceCycles = [1, 41];
+  let NumMicroOps = 2;
+}
+def : WriteRes<WriteIDivLd, [JLAGU, JALU1, JDiv]> {
+  let Latency = 44;
+  let ResourceCycles = [1, 1, 41];
+  let NumMicroOps = 2;
 }
 
 // This is for simple LEAs with one or two input operands.
 // FIXME: SAGU 3-operand LEA
 def : WriteRes<WriteLEA, [JALU01]>;
 
+def JWriteIDiv8 : SchedWriteRes<[JALU1, JDiv]> {
+  let Latency = 12;
+  let ResourceCycles = [1, 12];
+}
+def JWriteIDiv8Ld : SchedWriteRes<[JLAGU, JALU1, JDiv]> {
+  let Latency = 15;
+  let ResourceCycles = [1, 1, 12];
+}
+def : InstRW<[JWriteIDiv8], (instrs DIV8r, IDIV8r)>;
+def : InstRW<[JWriteIDiv8Ld], (instrs DIV8m, IDIV8m)>;
+
+def JWriteIDiv16 : SchedWriteRes<[JALU1, JDiv]> {
+  let Latency = 17;
+  let ResourceCycles = [1, 17];
+  let NumMicroOps = 2;
+}
+def JWriteIDiv16Ld : SchedWriteRes<[JLAGU, JALU1, JDiv]> {
+  let Latency = 20;
+  let ResourceCycles = [1, 1, 17];
+  let NumMicroOps = 2;
+}
+def : InstRW<[JWriteIDiv16], (instrs DIV16r, IDIV16r)>;
+def : InstRW<[JWriteIDiv16Ld], (instrs DIV16m, IDIV16m)>;
+
+def JWriteIDiv32 : SchedWriteRes<[JALU1, JDiv]> {
+  let Latency = 25;
+  let ResourceCycles = [1, 25];
+  let NumMicroOps = 2;
+}
+def JWriteIDiv32Ld : SchedWriteRes<[JLAGU, JALU1, JDiv]> {
+  let Latency = 28;
+  let ResourceCycles = [1, 1, 25];
+  let NumMicroOps = 2;
+}
+def : InstRW<[JWriteIDiv32], (instrs DIV32r, IDIV32r)>;
+def : InstRW<[JWriteIDiv32Ld], (instrs DIV32m, IDIV32m)>;
+
 ////////////////////////////////////////////////////////////////////////////////
 // Integer shifts and rotates.
 ////////////////////////////////////////////////////////////////////////////////
diff --git a/llvm/test/CodeGen/X86/schedule-x86_64.ll b/llvm/test/CodeGen/X86/schedule-x86_64.ll
index 999229ec118..3b13fb69fa7 100644
--- a/llvm/test/CodeGen/X86/schedule-x86_64.ll
+++ b/llvm/test/CodeGen/X86/schedule-x86_64.ll
@@ -5259,14 +5259,14 @@ define void @test_div(i8 %a0, i16 %a1, i32 %a2, i64 %a3, i8 *%p0, i16 *%p1, i32
 ; BTVER2-NEXT:    movq {{[0-9]+}}(%rsp), %r10 # sched: [5:1.00]
 ; BTVER2-NEXT:    movq {{[0-9]+}}(%rsp), %rax # sched: [5:1.00]
 ; BTVER2-NEXT:    #APP
-; BTVER2-NEXT:    divb %dil # sched: [25:25.00]
-; BTVER2-NEXT:    divb (%r8) # sched: [41:25.00]
-; BTVER2-NEXT:    divw %si # sched: [25:25.00]
-; BTVER2-NEXT:    divw (%r9) # sched: [41:25.00]
+; BTVER2-NEXT:    divb %dil # sched: [12:12.00]
+; BTVER2-NEXT:    divb (%r8) # sched: [15:12.00]
+; BTVER2-NEXT:    divw %si # sched: [17:17.00]
+; BTVER2-NEXT:    divw (%r9) # sched: [20:17.00]
 ; BTVER2-NEXT:    divl %edx # sched: [25:25.00]
-; BTVER2-NEXT:    divl (%rax) # sched: [41:25.00]
-; BTVER2-NEXT:    divq %rcx # sched: [25:25.00]
-; BTVER2-NEXT:    divq (%r10) # sched: [41:25.00]
+; BTVER2-NEXT:    divl (%rax) # sched: [28:25.00]
+; BTVER2-NEXT:    divq %rcx # sched: [41:41.00]
+; BTVER2-NEXT:    divq (%r10) # sched: [44:41.00]
 ; BTVER2-NEXT:    #NO_APP
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;
@@ -5507,14 +5507,14 @@ define void @test_idiv(i8 %a0, i16 %a1, i32 %a2, i64 %a3, i8 *%p0, i16 *%p1, i32
 ; BTVER2-NEXT:    movq {{[0-9]+}}(%rsp), %r10 # sched: [5:1.00]
 ; BTVER2-NEXT:    movq {{[0-9]+}}(%rsp), %rax # sched: [5:1.00]
 ; BTVER2-NEXT:    #APP
-; BTVER2-NEXT:    idivb %dil # sched: [25:25.00]
-; BTVER2-NEXT:    idivb (%r8) # sched: [41:25.00]
-; BTVER2-NEXT:    idivw %si # sched: [25:25.00]
-; BTVER2-NEXT:    idivw (%r9) # sched: [41:25.00]
+; BTVER2-NEXT:    idivb %dil # sched: [12:12.00]
+; BTVER2-NEXT:    idivb (%r8) # sched: [15:12.00]
+; BTVER2-NEXT:    idivw %si # sched: [17:17.00]
+; BTVER2-NEXT:    idivw (%r9) # sched: [20:17.00]
 ; BTVER2-NEXT:    idivl %edx # sched: [25:25.00]
-; BTVER2-NEXT:    idivl (%rax) # sched: [41:25.00]
-; BTVER2-NEXT:    idivq %rcx # sched: [25:25.00]
-; BTVER2-NEXT:    idivq (%r10) # sched: [41:25.00]
+; BTVER2-NEXT:    idivl (%rax) # sched: [28:25.00]
+; BTVER2-NEXT:    idivq %rcx # sched: [41:41.00]
+; BTVER2-NEXT:    idivq (%r10) # sched: [44:41.00]
 ; BTVER2-NEXT:    #NO_APP
 ; BTVER2-NEXT:    retq # sched: [4:1.00]
 ;