[PowerPC] Replace float load/store pair with integer load/store pair when it's only used in load/store

Replace float load/store pair with integer load/store pair when it's only used in load/store,
because float load/store instructions cost more cycles then integer load/store.

A typical scenario is when there is a call with more than 13 float arguments passing, we need pass them by stack.
So we need a load/store pair to do such memory operation if the variable is global variable.

Differential Revision: https://reviews.llvm.org/D64195

llvm-svn: 366775

Author: wuzish

llvm/lib/Target/PowerPC/PPCISelLowering.h

@@ -834,6 +834,18 @@ namespace llvm {
       return true;
     }
 
+    bool isDesirableToTransformToIntegerOp(unsigned Opc,
+                                           EVT VT) const override {
+      // Only handle float load/store pair because float(fpr) load/store
+      // instruction has more cycles than integer(gpr) load/store in PPC.
+      if (Opc != ISD::LOAD && Opc != ISD::STORE)
+        return false;
+      if (VT != MVT::f32 && VT != MVT::f64)
+        return false;
+
+      return true; 
+    }
+
     // Returns true if the address of the global is stored in TOC entry.
     bool isAccessedAsGotIndirect(SDValue N) const;
 

llvm/test/CodeGen/PowerPC/float-load-store-pair.ll

@@ -54,27 +54,27 @@ define signext i32 @test() {
 ; CHECK-NEXT:    addis 3, 2, a10@toc@ha
 ; CHECK-NEXT:    lfd 10, a10@toc@l(3)
 ; CHECK-NEXT:    addis 3, 2, a11@toc@ha
+; CHECK-NEXT:    addis 6, 2, a17@toc@ha
+; CHECK-NEXT:    addis 5, 2, a16@toc@ha
+; CHECK-NEXT:    addi 6, 6, a17@toc@l
+; CHECK-NEXT:    addi 5, 5, a16@toc@l
+; CHECK-NEXT:    lxvx 34, 0, 6
+; CHECK-NEXT:    addis 4, 2, a15@toc@ha
+; CHECK-NEXT:    lxvx 0, 0, 5
+; CHECK-NEXT:    ld 4, a15@toc@l(4)
+; CHECK-NEXT:    li 5, 168
 ; CHECK-NEXT:    lfd 11, a11@toc@l(3)
 ; CHECK-NEXT:    addis 3, 2, a12@toc@ha
 ; CHECK-NEXT:    lfd 12, a12@toc@l(3)
 ; CHECK-NEXT:    addis 3, 2, a13@toc@ha
 ; CHECK-NEXT:    lfd 13, a13@toc@l(3)
 ; CHECK-NEXT:    addis 3, 2, a14@toc@ha
-; CHECK-NEXT:    lfd 0, a14@toc@l(3)
-; CHECK-NEXT:    addis 3, 2, a15@toc@ha
-; CHECK-NEXT:    addis 4, 2, a17@toc@ha
-; CHECK-NEXT:    addi 4, 4, a17@toc@l
-; CHECK-NEXT:    lxsd 2, a15@toc@l(3)
-; CHECK-NEXT:    addis 3, 2, a16@toc@ha
-; CHECK-NEXT:    addi 3, 3, a16@toc@l
-; CHECK-NEXT:    lxvx 36, 0, 4
-; CHECK-NEXT:    lxvx 35, 0, 3
-; CHECK-NEXT:    li 3, 168
-; CHECK-NEXT:    stxvx 36, 1, 3
-; CHECK-NEXT:    li 3, 152
-; CHECK-NEXT:    stxvx 35, 1, 3
-; CHECK-NEXT:    stxsd 2, 144(1)
-; CHECK-NEXT:    stfd 0, 136(1)
+; CHECK-NEXT:    ld 3, a14@toc@l(3)
+; CHECK-NEXT:    stxvx 34, 1, 5
+; CHECK-NEXT:    li 5, 152
+; CHECK-NEXT:    stxvx 0, 1, 5
+; CHECK-NEXT:    std 4, 144(1)
+; CHECK-NEXT:    std 3, 136(1)
 ; CHECK-NEXT:    bl _Z3fooddddddddddddddd
 ; CHECK-NEXT:    nop
 ; CHECK-NEXT:    li 3, 0

llvm/test/CodeGen/PowerPC/ppc64-smallarg.ll

@@ -52,7 +52,7 @@ entry:
   ret void
 }
 ; CHECK: @caller2
-; CHECK: stfs {{[0-9]+}}, 156(1)
+; CHECK: std {{[0-9]+}}, 16(1)
 ; CHECK: bl test2
 
 declare float @test2(float, float, float, float, float, float, float, float, float, float, float, float, float, float)

llvm/test/CodeGen/PowerPC/ppc64le-smallarg.ll

@@ -52,7 +52,7 @@ entry:
   ret void
 }
 ; CHECK: @caller2
-; CHECK: stfs {{[0-9]+}}, 136({{[0-9]+}})
+; CHECK: std {{[0-9]+}}, 16({{[0-9]+}})
 ; CHECK: bl test2
 
 declare float @test2(float, float, float, float, float, float, float, float, float, float, float, float, float, float)

llvm/test/CodeGen/PowerPC/pwr7-gt-nop.ll

@@ -16,12 +16,12 @@ entry:
   store float %2, float* %d, align 4
   ret void
 
-; CHECK: lfs [[REG1:[0-9]+]], 0(4)
-; CHECK: stfs [[REG1]], 0(3)
-; CHECK: lfs [[REG2:[0-9]+]], 0(5)
-; CHECK: stfs [[REG2]], 0(4)
-; CHECK: lfs [[REG3:[0-9]+]], 0(3)
-; CHECK: stfs [[REG3]], 0(6)
+; CHECK: lwz [[REG1:[0-9]+]], 0(4)
+; CHECK: stw [[REG1]], 0(3)
+; CHECK: lwz [[REG2:[0-9]+]], 0(5)
+; CHECK: stw [[REG2]], 0(4)
+; CHECK: lwz [[REG3:[0-9]+]], 0(3)
+; CHECK: stw [[REG3]], 0(6)
 ; CHECK: blr
 }