[AArch32] Added patterns for VCVT{A,N, P,M}.

[mcrosier]

Summary

All,
This patch adds patterns for VCVT{A,N, P,M}. Patterns for lowering libm calls to VCVT{A,N,P,M}, e.g. (int)floorf(float) => VCVTM.s32.f32, are also included. Please have a look.

Regards,

Chad

[compnerd]

First, thanks for adding these.

The changes to the table data looks fine.

I think that you should add a bit more test coverage though. The hard float shouldn't matter towards use of these instructions. So, I think you should also test gnueabi. Furthermore, I believe that these instructions are also available on non-v8 hardware if you have a NEON FPU. So, it may be nice to also add cases for armv7 with neon on hard float and soft float. The tests would be identical in those cases, simply a matter of adding the appropriate RUN lines.

[rengolin]

In D5033#4, @compnerd wrote:

I think that you should add a bit more test coverage though. The hard float shouldn't matter towards use of these instructions. So, I think you should also test gnueabi.

Shouldn't matter, but adding won't hurt either. Is VCVTN tested elsewhere?

Furthermore, I believe that these instructions are also available on non-v8 hardware if you have a NEON FPU. So, it may be nice to also add cases for armv7 with neon on hard float and soft float.

AFAIK, VCVT with user rounding in ARMv7 can only be done via the FPSCR, which is not the same thing here (and will need more code if there isn't already). These tests are clearly for v8, so I'm fine with it the way it is.

cheers,
--renato

[mcrosier]

I went ahead and committed this in r216388, per Renato's feedback. This was only designed to target ARMv8 and I'll defer any additional improvements to a later patch.

Chad

[rengolin]

Thanks!

I think making this work in v7 will involve a lot more thinking, since this changes the FPSCR register, even if for a brief moment, while exception can still occur. It's a whole new problem.

[mcrosier]

Committed r216388.

lib/Target/ARM/ARMInstrVFP.td

 def : Pat<(f64 (f16_to_fp GPR:$a)),
           (VCVTBHD (COPY_TO_REGCLASS GPR:$a, SPR))>;
 
-
-multiclass vcvt_inst<string opc, bits<2> rm> {
+multiclass vcvt_inst<string opc, bits<2> rm,
+                     SDPatternOperator node = null_frag> {
   let PostEncoderMethod = "", DecoderNamespace = "VFPV8" in {
     def SS : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
                     (outs SPR:$Sd), (ins SPR:$Sm),
                     NoItinerary, !strconcat("vcvt", opc, ".s32.f32\t$Sd, $Sm"),
-                    []>, Requires<[HasFPARMv8]> {
+                    [(set SPR:$Sd, (arm_ftosi (node SPR:$Sm)))]>,
+                    Requires<[HasFPARMv8]> {
       let Inst{17-16} = rm;
     }
 
     def US : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
                     (outs SPR:$Sd), (ins SPR:$Sm),
                     NoItinerary, !strconcat("vcvt", opc, ".u32.f32\t$Sd, $Sm"),
-                    []>, Requires<[HasFPARMv8]> {
+                    [(set SPR:$Sd, (arm_ftoui (node SPR:$Sm)))]>,
+                    Requires<[HasFPARMv8]> {
       let Inst{17-16} = rm;
     }
 
     def SD : ASuInp<0b11101, 0b11, 0b1100, 0b11, 0,
                     (outs SPR:$Sd), (ins DPR:$Dm),
                     NoItinerary, !strconcat("vcvt", opc, ".s32.f64\t$Sd, $Dm"),
-                    []>, Requires<[HasFPARMv8, HasDPVFP]> {
+                    [(set SPR:$Sd, (arm_ftosi (f64 (node (f64 DPR:$Dm)))))]>,
+                    Requires<[HasFPARMv8, HasDPVFP]> {
       bits<5> Dm;
 
       let Inst{17-16} = rm;

     def UD : ASuInp<0b11101, 0b11, 0b1100, 0b01, 0,
                     (outs SPR:$Sd), (ins DPR:$Dm),
                     NoItinerary, !strconcat("vcvt", opc, ".u32.f64\t$Sd, $Dm"),
-                    []>, Requires<[HasFPARMv8, HasDPVFP]> {
+                    [(set SPR:$Sd, (arm_ftoui (f64 (node (f64 DPR:$Dm)))))]>,
+                    Requires<[HasFPARMv8, HasDPVFP]> {
       bits<5> Dm;
 
       let Inst{17-16} = rm;

   }
 }
 
-defm VCVTA : vcvt_inst<"a", 0b00>;
+defm VCVTA : vcvt_inst<"a", 0b00, frnd>;
 defm VCVTN : vcvt_inst<"n", 0b01>;
-defm VCVTP : vcvt_inst<"p", 0b10>;
-defm VCVTM : vcvt_inst<"m", 0b11>;
+defm VCVTP : vcvt_inst<"p", 0b10, fceil>;
+defm VCVTM : vcvt_inst<"m", 0b11, ffloor>;
 
 def VNEGD  : ADuI<0b11101, 0b11, 0b0001, 0b01, 0,
                   (outs DPR:$Dd), (ins DPR:$Dm),

test/CodeGen/ARM/arm32-round-conv.ll

+; RUN: llc < %s -mtriple=armv8-linux-gnueabihf -mattr=+fp-armv8 | FileCheck %s
+
+; CHECK-LABEL: test1
+; CHECK: vcvtm.s32.f32
+define i32 @test1(float %a) {
+entry:
+  %call = call float @floorf(float %a) nounwind readnone
+  %conv = fptosi float %call to i32
+  ret i32 %conv
+}
+
+; CHECK-LABEL: test2
+; CHECK: vcvtm.u32.f32
+define i32 @test2(float %a) {
+entry:
+  %call = call float @floorf(float %a) nounwind readnone
+  %conv = fptoui float %call to i32
+  ret i32 %conv
+}
+
+; CHECK-LABEL: test3
+; CHECK: vcvtm.s32.f64
+define i32 @test3(double %a) {
+entry:
+  %call = call double @floor(double %a) nounwind readnone
+  %conv = fptosi double %call to i32
+  ret i32 %conv
+}
+
+; CHECK-LABEL: test4
+; CHECK: vcvtm.u32.f64
+define i32 @test4(double %a) {
+entry:
+  %call = call double @floor(double %a) nounwind readnone
+  %conv = fptoui double %call to i32
+  ret i32 %conv
+}
+
+; CHECK-LABEL: test5
+; CHECK: vcvtp.s32.f32
+define i32 @test5(float %a) {
+entry:
+  %call = call float @ceilf(float %a) nounwind readnone
+  %conv = fptosi float %call to i32
+  ret i32 %conv
+}
+
+; CHECK-LABEL: test6
+; CHECK: vcvtp.u32.f32
+define i32 @test6(float %a) {
+entry:
+  %call = call float @ceilf(float %a) nounwind readnone
+  %conv = fptoui float %call to i32
+  ret i32 %conv
+}
+
+; CHECK-LABEL: test7
+; CHECK: vcvtp.s32.f64
+define i32 @test7(double %a) {
+entry:
+  %call = call double @ceil(double %a) nounwind readnone
+  %conv = fptosi double %call to i32
+  ret i32 %conv
+}
+
+; CHECK-LABEL: test8
+; CHECK: vcvtp.u32.f64
+define i32 @test8(double %a) {
+entry:
+  %call = call double @ceil(double %a) nounwind readnone
+  %conv = fptoui double %call to i32
+  ret i32 %conv
+}
+
+; CHECK-LABEL: test9
+; CHECK: vcvta.s32.f32
+define i32 @test9(float %a) {
+entry:
+  %call = call float @roundf(float %a) nounwind readnone
+  %conv = fptosi float %call to i32
+  ret i32 %conv
+}
+
+; CHECK-LABEL: test10
+; CHECK: vcvta.u32.f32
+define i32 @test10(float %a) {
+entry:
+  %call = call float @roundf(float %a) nounwind readnone
+  %conv = fptoui float %call to i32
+  ret i32 %conv
+}
+
+; CHECK-LABEL: test11
+; CHECK: vcvta.s32.f64
+define i32 @test11(double %a) {
+entry:
+  %call = call double @round(double %a) nounwind readnone
+  %conv = fptosi double %call to i32
+  ret i32 %conv
+}
+
+; CHECK-LABEL: test12
+; CHECK: vcvta.u32.f64
+define i32 @test12(double %a) {
+entry:
+  %call = call double @round(double %a) nounwind readnone
+  %conv = fptoui double %call to i32
+  ret i32 %conv
+}
+
+declare float @floorf(float) nounwind readnone
+declare double @floor(double) nounwind readnone
+declare float @ceilf(float) nounwind readnone
+declare double @ceil(double) nounwind readnone
+declare float @roundf(float) nounwind readnone
+declare double @round(double) nounwind readnone