Subzero: Fix lowering of the fabs() intrinsic.

tests_lit/llvm2ice_tests/nacl-other-intrinsics.ll

 ; This tests the NaCl intrinsics not related to atomic operations.
 
 ; RUN: %p2i -i %s --filetype=obj --disassemble --args -O2 -sandbox \
 ; RUN:   | FileCheck %s
 ; RUN: %p2i -i %s --filetype=obj --disassemble --args -Om1 -sandbox \
 ; RUN:   | FileCheck %s
 
 ; Do another run w/ O2 and a different check-prefix (otherwise O2 and Om1
 ; share the same "CHECK" prefix). This separate run helps check that
 ; some code is optimized out.
@@ skipping 261 matching lines @@
 ; CHECKO2REM-NOT: sqrtsd
 
 define float @test_fabs_float(float %x) {
 entry:
   %r = call float @llvm.fabs.f32(float %x)
   %r2 = call float @llvm.fabs.f32(float %r)
   %r3 = call float @llvm.fabs.f32(float -0.0)
   %r4 = fadd float %r2, %r3
   ret float %r4
 }
+;;; Specially check that the pand instruction doesn't try to operate on a 32-bit
+;;; (f32) memory operand, and instead uses two xmm registers.
 ; CHECK-LABEL: test_fabs_float
 ; CHECK: pcmpeqd
 ; CHECK: psrld
-; CHECK: pand
+; CHECK: pand {{.*}}xmm{{.*}}xmm
 ; CHECK: pcmpeqd
 ; CHECK: psrld
-; CHECK: pand
+; CHECK: pand {{.*}}xmm{{.*}}xmm
 ; CHECK: pcmpeqd
 ; CHECK: psrld
-; CHECK: pand
+; CHECK: pand {{.*}}xmm{{.*}}xmm
 
 define double @test_fabs_double(double %x) {
 entry:
   %r = call double @llvm.fabs.f64(double %x)
   %r2 = call double @llvm.fabs.f64(double %r)
   %r3 = call double @llvm.fabs.f64(double -0.0)
   %r4 = fadd double %r2, %r3
   ret double %r4
 }
+;;; Specially check that the pand instruction doesn't try to operate on a 64-bit
+;;; (f64) memory operand, and instead uses two xmm registers.
 ; CHECK-LABEL: test_fabs_double
 ; CHECK: pcmpeqd
 ; CHECK: psrlq
-; CHECK: pand
+; CHECK: pand {{.*}}xmm{{.*}}xmm
 ; CHECK: pcmpeqd
 ; CHECK: psrlq
-; CHECK: pand
+; CHECK: pand {{.*}}xmm{{.*}}xmm
 ; CHECK: pcmpeqd
 ; CHECK: psrlq
-; CHECK: pand
+; CHECK: pand {{.*}}xmm{{.*}}xmm
 
 define <4 x float> @test_fabs_v4f32(<4 x float> %x) {
 entry:
   %r = call <4 x float> @llvm.fabs.v4f32(<4 x float> %x)
   %r2 = call <4 x float> @llvm.fabs.v4f32(<4 x float> %r)
   %r3 = call <4 x float> @llvm.fabs.v4f32(<4 x float> undef)
   %r4 = fadd <4 x float> %r2, %r3
   ret <4 x float> %r4
 }
 ; CHECK-LABEL: test_fabs_v4f32
@@ skipping 212 matching lines @@
 
   call void @llvm.stackrestore(i8* %sp1)
   ret void
 }
 ; CHECK-LABEL: test_stacksave_multiple
 ; At least 3 copies of esp, but probably more from having to do the allocas.
 ; CHECK: mov {{.*}},esp
 ; CHECK: mov {{.*}},esp
 ; CHECK: mov {{.*}},esp
 ; CHECK: mov esp,{{.*}}