[SystemZ] Fix register ordering for BinaryRRF instructions

[• zhanjunl]

Summary

The ordering of registers in BinaryRRF instructions are wrong, and affects the copysign instruction (CPSDR). This results in the wrong magnitude and sign being set.

[• zhanjunl]

A simple testcase is:

#include <stdio.h>
#include <cmath>

double cpsdr(double a, double b) {
  return copysign(a,b);
}

double a = 1.0;
double b = -2.0;

int main () {
  printf("ret is %lf\n", cpsdr(a, b));
  return 0;
}

The expected result is -1.0, but the return of clang is 2.0.

[uweigand]

LGTM, thanks!

llvm/trunk/lib/Target/SystemZ/SystemZInstrFormats.td

class BinaryRRE<string mnemonic, bits<16> opcode, SDPatternOperator operator,
   let OpKey = mnemonic ## cls1;
   let OpType = "reg";
   let Constraints = "$R1 = $R1src";
   let DisableEncoding = "$R1src";
 }
 
 class BinaryRRF<string mnemonic, bits<16> opcode, SDPatternOperator operator,
                 RegisterOperand cls1, RegisterOperand cls2>
-  : InstRRF<opcode, (outs cls1:$R1), (ins cls1:$R3, cls2:$R2),
+  : InstRRF<opcode, (outs cls1:$R1), (ins cls1:$R2, cls2:$R3),
             mnemonic#"r\t$R1, $R3, $R2",
-            [(set cls1:$R1, (operator cls1:$R3, cls2:$R2))]> {
+            [(set cls1:$R1, (operator cls1:$R2, cls2:$R3))]> {
   let OpKey = mnemonic ## cls1;
   let OpType = "reg";
   let R4 = 0;
 }
 
 class BinaryRRFK<string mnemonic, bits<16> opcode, SDPatternOperator operator,
                  RegisterOperand cls1, RegisterOperand cls2>
   : InstRRF<opcode, (outs cls1:$R1), (ins cls1:$R2, cls2:$R3),

llvm/trunk/test/CodeGen/SystemZ/fp-copysign-01.ll

 ; Test copysign operations.
 ;
 ; RUN: llc < %s -mtriple=s390x-linux-gnu | FileCheck %s
 
 declare float @copysignf(float, float) readnone
 declare double @copysign(double, double) readnone
 ; FIXME: not really the correct prototype for SystemZ.
 declare fp128 @copysignl(fp128, fp128) readnone
 
 ; Test f32 copies in which the sign comes from an f32.
 define float @f1(float %a, float %b) {
 ; CHECK-LABEL: f1:
 ; CHECK-NOT: %f2
-; CHECK: cpsdr %f0, %f0, %f2
+; CHECK: cpsdr %f0, %f2, %f0
 ; CHECK: br %r14
   %res = call float @copysignf(float %a, float %b) readnone
   ret float %res
 }
 
 ; Test f32 copies in which the sign comes from an f64.
 define float @f2(float %a, double %bd) {
 ; CHECK-LABEL: f2:
 ; CHECK-NOT: %f2
-; CHECK: cpsdr %f0, %f0, %f2
+; CHECK: cpsdr %f0, %f2, %f0
 ; CHECK: br %r14
   %b = fptrunc double %bd to float
   %res = call float @copysignf(float %a, float %b) readnone
   ret float %res
 }
 
 ; Test f32 copies in which the sign comes from an f128.
 define float @f3(float %a, fp128 *%bptr) {
 ; CHECK-LABEL: f3:
 ; CHECK: ld [[BHIGH:%f[0-7]]], 0(%r2)
 ; CHECK: ld [[BLOW:%f[0-7]]], 8(%r2)
-; CHECK: cpsdr %f0, %f0, [[BHIGH]]
+; CHECK: cpsdr %f0, [[BHIGH]], %f0
 ; CHECK: br %r14
   %bl = load volatile fp128 , fp128 *%bptr
   %b = fptrunc fp128 %bl to float
   %res = call float @copysignf(float %a, float %b) readnone
   ret float %res
 }
 
 ; Test f64 copies in which the sign comes from an f32.
 define double @f4(double %a, float %bf) {
 ; CHECK-LABEL: f4:
 ; CHECK-NOT: %f2
-; CHECK: cpsdr %f0, %f0, %f2
+; CHECK: cpsdr %f0, %f2, %f0
 ; CHECK: br %r14
   %b = fpext float %bf to double
   %res = call double @copysign(double %a, double %b) readnone
   ret double %res
 }
 
 ; Test f64 copies in which the sign comes from an f64.
 define double @f5(double %a, double %b) {
 ; CHECK-LABEL: f5:
 ; CHECK-NOT: %f2
-; CHECK: cpsdr %f0, %f0, %f2
+; CHECK: cpsdr %f0, %f2, %f0
 ; CHECK: br %r14
   %res = call double @copysign(double %a, double %b) readnone
   ret double %res
 }
 
 ; Test f64 copies in which the sign comes from an f128.
 define double @f6(double %a, fp128 *%bptr) {
 ; CHECK-LABEL: f6:
 ; CHECK: ld [[BHIGH:%f[0-7]]], 0(%r2)
 ; CHECK: ld [[BLOW:%f[0-7]]], 8(%r2)
-; CHECK: cpsdr %f0, %f0, [[BHIGH]]
+; CHECK: cpsdr %f0, [[BHIGH]], %f0
 ; CHECK: br %r14
   %bl = load volatile fp128 , fp128 *%bptr
   %b = fptrunc fp128 %bl to double
   %res = call double @copysign(double %a, double %b) readnone
   ret double %res
 }
 
 ; Test f128 copies in which the sign comes from an f32.  We shouldn't
 ; need any register shuffling here; %a should be tied to %c, with CPSDR
 ; just changing the high register.
 define void @f7(fp128 *%cptr, fp128 *%aptr, float %bf) {
 ; CHECK-LABEL: f7:
 ; CHECK: ld [[AHIGH:%f[0-7]]], 0(%r3)
 ; CHECK: ld [[ALOW:%f[0-7]]], 8(%r3)
-; CHECK: cpsdr [[AHIGH]], [[AHIGH]], %f0
+; CHECK: cpsdr [[AHIGH]], %f0, [[AHIGH]]
 ; CHECK: std [[AHIGH]], 0(%r2)
 ; CHECK: std [[ALOW]], 8(%r2)
 ; CHECK: br %r14
   %a = load volatile fp128 , fp128 *%aptr
   %b = fpext float %bf to fp128
   %c = call fp128 @copysignl(fp128 %a, fp128 %b) readnone
   store fp128 %c, fp128 *%cptr
   ret void
 }
 
 ; As above, but the sign comes from an f64.
 define void @f8(fp128 *%cptr, fp128 *%aptr, double %bd) {
 ; CHECK-LABEL: f8:
 ; CHECK: ld [[AHIGH:%f[0-7]]], 0(%r3)
 ; CHECK: ld [[ALOW:%f[0-7]]], 8(%r3)
-; CHECK: cpsdr [[AHIGH]], [[AHIGH]], %f0
+; CHECK: cpsdr [[AHIGH]], %f0, [[AHIGH]]
 ; CHECK: std [[AHIGH]], 0(%r2)
 ; CHECK: std [[ALOW]], 8(%r2)
 ; CHECK: br %r14
   %a = load volatile fp128 , fp128 *%aptr
   %b = fpext double %bd to fp128
   %c = call fp128 @copysignl(fp128 %a, fp128 %b) readnone
   store fp128 %c, fp128 *%cptr
   ret void
 }
 
 ; As above, but the sign comes from an f128.  Don't require the low part
 ; of %b to be loaded, since it isn't used.
 define void @f9(fp128 *%cptr, fp128 *%aptr, fp128 *%bptr) {
 ; CHECK-LABEL: f9:
 ; CHECK: ld [[AHIGH:%f[0-7]]], 0(%r3)
 ; CHECK: ld [[ALOW:%f[0-7]]], 8(%r3)
 ; CHECK: ld [[BHIGH:%f[0-7]]], 0(%r4)
-; CHECK: cpsdr [[AHIGH]], [[AHIGH]], [[BHIGH]]
+; CHECK: cpsdr [[AHIGH]], [[BHIGH]], [[AHIGH]]
 ; CHECK: std [[AHIGH]], 0(%r2)
 ; CHECK: std [[ALOW]], 8(%r2)
 ; CHECK: br %r14
   %a = load volatile fp128 , fp128 *%aptr
   %b = load volatile fp128 , fp128 *%bptr
   %c = call fp128 @copysignl(fp128 %a, fp128 %b) readnone
   store fp128 %c, fp128 *%cptr
   ret void
 }