[ARM] Hoist fabs/fneg above a conversion to float.

[jmolloy]

Summary

This is especially visible in softfp mode, for example in the implementation of libm fabs/fneg functions. If we have:

%1 = vmovdrr r0, r1
%2 = fabs %1

then move the fabs before the vmovdrr:

%1 = and r1, #0x7FFFFFFF
%2 = vmovdrr r0, r1

This is never a lose, and could be a serious win because the vmovdrr may be followed by a vmovrrd, which would enable us to remove the conversion into FPRs completely.

We already do this for f32, but not for f64. Tests are added for both.

[jmolloy]

Hi Quentin,

I've added you as reviewer because of your work eliding cross class copies. Do you have any comments on this?

Cheers,
James

[qcolombet]

Hi James,

Sounds like a good idea.

Cheers,
-Quentin

[jmolloy]

Thanks Quentin,

Committed in r246360.

lib/Target/ARM/ARMInstrVFP.td

 def arm_cmpfp0 : SDNode<"ARMISD::CMPFPw0", SDT_CMPFP0, [SDNPOutGlue]>;
 def arm_fmdrr  : SDNode<"ARMISD::VMOVDRR", SDT_VMOVDRR>;
 
-
 //===----------------------------------------------------------------------===//
 // Operand Definitions.
 //

   let isRegSequence = 1;
 }
 
+// Hoist an fabs or a fneg of a value coming from integer registers
+// and do the fabs/fneg on the integer value. This is never a lose
+// and could enable the conversion to float to be removed completely.
+def : Pat<(fabs (arm_fmdrr GPR:$Rl, GPR:$Rh)),
+          (VMOVDRR GPR:$Rl, (BFC GPR:$Rh, (i32 0x7FFFFFFF)))>,
+      Requires<[IsARM]>;
+def : Pat<(fabs (arm_fmdrr GPR:$Rl, GPR:$Rh)),
+          (VMOVDRR GPR:$Rl, (t2BFC GPR:$Rh, (i32 0x7FFFFFFF)))>,
+      Requires<[IsThumb2]>;
+def : Pat<(fneg (arm_fmdrr GPR:$Rl, GPR:$Rh)),
+          (VMOVDRR GPR:$Rl, (EORri GPR:$Rh, (i32 0x80000000)))>,
+      Requires<[IsARM]>;
+def : Pat<(fneg (arm_fmdrr GPR:$Rl, GPR:$Rh)),
+          (VMOVDRR GPR:$Rl, (t2EORri GPR:$Rh, (i32 0x80000000)))>,
+      Requires<[IsThumb2]>;
+
 let hasSideEffects = 0 in
 def VMOVSRR : AVConv5I<0b11000100, 0b1010,
                      (outs SPR:$dst1, SPR:$dst2), (ins GPR:$src1, GPR:$src2),

test/CodeGen/ARM/softfp-fabs-fneg.ll

+; RUN: llc -mtriple=armv7 < %s | FileCheck %s --check-prefix=CHECK-ARM --check-prefix=CHECK
+; RUN: llc -mtriple=thumbv7 < %s | FileCheck %s --check-prefix=CHECK-THUMB --check-prefix=CHECK
+
+target datalayout = "e-m:e-p:32:32-i64:64-v128:64:128-a:0:32-n32-S64"
+target triple = "armv7--"
+
+define double @f(double %a) {
+  ; CHECK-LABEL: f:
+  ; CHECK: bfc r1, #31, #1
+  ; CHECK-NEXT: bx lr
+  %x = call double @llvm.fabs.f64(double %a) readnone
+  ret double %x
+}
+
+define float @g(float %a) {
+  ; CHECK-LABEL: g:
+  ; CHECK-THUMB: bic r0, r0, #-2147483648
+  ; CHECK-ARM: bfc r0, #31, #1
+  ; CHECK-NEXT: bx lr
+  %x = call float @llvm.fabs.f32(float %a) readnone
+  ret float %x
+}
+
+define double @h(double %a) {
+  ; CHECK-LABEL: h:
+  ; CHECK: eor r1, r1, #-2147483648
+  ; CHECK-NEXT: bx lr
+  %x = fsub nsz double -0.0, %a
+  ret double %x
+}
+
+define float @i(float %a) {
+  ; CHECK-LABEL: i:
+  ; CHECK: eor r0, r0, #-2147483648
+  ; CHECK-NEXT: bx lr
+  %x = fsub nsz float -0.0, %a
+  ret float %x
+}
+
+declare double @llvm.fabs.f64(double) readnone
+declare float @llvm.fabs.f32(float) readnone